Interrogations of DNS of Solid-Liquid Flows Daniel D

Interrogations of Direct Numerical Simulation of
Solid-Liquid Flows
Daniel D. Joseph
University of Minnesota
107 Akerman Hall
110 Union Street SE
Minneapolis, MN 55455
Table of Contents
Abstract ....................................................................................................................................................... 1
I Introduction.................................................................................................................................... 1
What is DNS? .................................................................................................................................. 1
Approximate Methods ..................................................................................................................... 2
Data Structures for Applications of DNS ........................................................................................ 3
II Equations of Motion ...................................................................................................................... 4
Weak Solutions for Total Momentum ............................................................................................. 4
III Numerical Packages for Moving Particles in Direct Numerical Simulation ............................ 6
ALE Particle Mover ........................................................................................................................ 6
A Projected Particle Mover.............................................................................................................. 7
DLM Particle Mover ....................................................................................................................... 7
Parallel Implementation .................................................................................................................. 9
Viscoelastic DLM Particle Mover ................................................................................................. 10
Computation domains for pipe flows of slurries............................................................................ 10
Collision Strategies ........................................................................................................................ 11
Addendum to Chapter III (Addendum A). Experiments on on Particle Collisions in Viscous
Liquids; Stokes number......................................................................................................... A-1
IV Weak and Strong Formulations of the DLM Method .............................................................. 13
Body Force Formulation of DLM Method: Weak Solution .......................................................... 13
Body Force Formulation of DLM: Strong Solutions ..................................................................... 13
Stress Formulation of DLM: Weak Solutions ............................................................................... 15
Stress Formulation of DLM: Strong Solutions .............................................................................. 16
Addendum to Chapter IV (Addendum B). Weak form of body force formulation when the
Lagrange Multiplier is eliminated ......................................................................................... B-1
Addendum to Chapter IV (Addendum C). Recent Developments in Direct Numerical
Methods for Solid-Liquid Flows ........................................................................................... C-1
V Applications of DNS .................................................................................................................... 17
Studies of Microstructure............................................................................................................... 17
Sedimentation and Fluidization ..................................................................................................... 20
Mechanisms of Cross-Stream Migration ...................................................................................... 21
Slot Problems for Particle Transport in Fractured Reservoirs ....................................................... 23
Lift-off, Resuspension, Equilibrium Height, Slip Velocities, Lift-Force Ratios ........................... 24
VI Modeling and DNS ...................................................................................................................... 25
Averaging...................................................................................................................................... 25
Fluidization by Drag and Fluidization by Lift ............................................................................... 30
Fluidization and Sedimentation ..................................................................................................... 30
Two Differences Between Fluidization and Sedimentation .......................................................... 31
Uniform Fluidization ..................................................................................................................... 31
© 2002 Daniel D. Joseph i

Incipient Fluidization ..................................................................................................................... 32
Hindered Settling ........................................................................................................................... 33
Drag and Hindered Settling ........................................................................................................... 36
Dynamic and One Dimensional Models ....................................................................................... 38
Particle Phase Pressure and the Stability of Uniform Fluidization ............................................... 39
Addendum to Chapter VI (Addendum D). Lift Forces on a Cylindrical Particle in Plane
Poiseuille Flow of Newtonian and Shear Thining Fluids ..................................................... D-1
Addendum to Chapter VI (Addendum E). Lift Force on a Spherical in tube flow .............. E-1
VII Direct Numerical Simulation of Fluidization of 1204 Spheres................................................. 41
Experiments ` ................................................................................................................................. 41
Numerical Method ........................................................................................................................ 41
Experiments .................................................................................................................................. 44
Numerical Simulation ................................................................................................................... 46
Qualitative comparison of experiment and simulation ................................................................. 49
Numerical computation of averaged quantities ............................................................................ 52
Dynamic and wall friction pressure in a fluidized bed .................................................................. 55
Sedimentation and fluidization velocity of single spheres ............................................................ 57
Richardson-Zaki correlations from DNS ...................................................................................... 60
Discussion ..................................................................................................................................... 63
VIII Modeling Rayleigh-Taylor instability of a sedimenting suspension of circular particles...... 65
Simulation Data ............................................................................................................................ 65
Comparison of Two-fluid Model and Simulation ......................................................................... 71
Discussion and Conclusions ......................................................................................................... 74
IX Fluidization by lift: single particle studies ................................................................................ 85
Equations of motion and dimensionless parameters ..................................................................... 86
Lift-off of a single particle in plane Poiseuille flows of a Newtonian fluid ................................. 89
Data Structure for DNS and experiments ..................................................................................... 95
X Analytical models of lift ............................................................................................................ 100
Lift in an inviscid fluid ............................................................................................................... 100
Low Reynolds numbers .............................................................................................................. 102
Lift in shear flows at low R ......................................................................................................... 104
Slip velocity and lift .................................................................................................................... 105
Non-uniqueness .......................................................................................................................... 106
Validation of lift formulas by DNS ............................................................................................ 107
Wall effects in shear flows .......................................................................................................... 109
Curvature .................................................................................................................................... 110
Regular perturbation in the wall region ...................................................................................... 111
Reciprocal theorem ..................................................................................................................... 113
Finite size sphere near a wall ...................................................................................................... 114
XI Slip velocity and lift at finite Reynolds numbers ................................................................... 118
Equilibrium positions of neutrally buoyant and heavy particles ................................................. 118
Mechanism for lift ....................................................................................................................... 119
Numerical simulation of migration and lift ................................................................................. 120
Long particle model .................................................................................................................... 121
Slip velocities, angular slip velocities and lift for neutrally buoyant circular particles ............... 127
Slip velocities, angular slip velocities and lift for non-neutrally buoyant circular particles ....... 130
Summary ..................................................................................................................................... 134
XII Stability and turning point bifurcations of a single particle in Poiseuille flow ................... 136
Pressure lift and shear lift ........................................................................................................... 143
XIII Lift off of a single particle in plane-Poiseuille flow of an Oldroyd-B fluid .......................... 149
© 2002 Daniel D. Joseph ii

Neutrally buoyant particle ........................................................................................................... 150
Pressure shear and viscoelastic lift forces ................................................................................... 154
Power law correlations for lift-off in an Oldroyd B fluid ........................................................... 158
XIV Fluidization by lift of 300 circular particles in plane Poiseuille flow by DNS ..................... 163
Case 1: η = 1 poise, RG = 9.81 .................................................................................................... 164
Case 2: η = 0.2 poise, RG = 245 .................................................................................................. 171
Case 3: η = 0.1 poise, RG = 9.81/η2 = 9.8×103 ............................................................................ 177
Case 4: η = 0.05 poise, RG = 9.81/η2 = 3.92×104 ........................................................................ 186
Case 5: η = 0.01 poise, RG = 9.81/η2 = 9.8×104 ........................................................................... 194
An engineering correlation for the lift-off of circular particle in a plane Poiseuille flow of a
Newtonian fluid .......................................................................................................................... 199
Inertial mechanism of fluidization .............................................................................................. 202
XV Bi-power law correlations for sediment transport in pressure driven channel
flows ............................................................................................................................................ 207
Analogy between fluidization by drag and lift ............................................................................ 207
Direct numerical simulation (DNS) of solid-liquid flows ........................................................... 209
Experimental setup ..................................................................................................................... 211
Experimental correlations for sediment transport ....................................................................... 213
Dimensionless parameters .................................................................................................... 213
Power law correlations for the erosion case (Patankar et al. 2002) ...................................... 213
Bi-power law correlations for the bed load transport case (Wang et al. 2002) .................... 215
Summary ..................................................................................................................................... 227
Appendix A (contribution of B. Baree) ...................................................................................... 229
Fitting power-law data with transition regions by a continuous function:
General framework and application to the Richardson-Zaki correlation ............................. 229
Appendix B ................................................................................................................................. 233
XVI Fluid Dynamics of Floating Particles ...................................................................................XVI-1
Self-aggregation, clustering, and bonding of particles in a liquid film rimming the inside of a
partially filled rotating particle ................................................................................................XVI-1
Fluid Dynamics of Floating Particles ....................................................................................XVI-37
XVII Epilogue ..................................................................................................................................... 234
Acknowledgment ........................................................................................................................ 237
References ................................................................................................................................... 241
Nomenclature list ........................................................................................................................ 252
Index ........................................................................................................................................... 255
© 2002 Daniel D. Joseph iii

Interrogations of DNS of Solid-Liquid Flows Introduction
Abstract
In direct simulation the fluid motion is resolved numerically and the forces which move
the particles are computed rather than modeled. This procedure opens new windows for
understanding and modeling. Numerical methods are discussed based on body fitted
moving unstructured grids and another on a fixed grid in which the portions of the fluid
occupied by solids are forced to move as a rigid body by a distribution of Lagrange
multipliers. Animation of the fluidization of 1204 spheres in 3D will be compared with
experiments and the concept of fluidization of slurries in conduits by lift rather than
drag will be framed in animation by direct simulation. Correlation for lift-off of single
particles and the bed height of slurries fluidized by lift are obtained by processing data
from numerical experiments.
I. Introduction
The current popularity of computational fluid dynamics is rooted in the perception that
information implicit in the equations of motion can be extracted without approximation using
direct numerical simulation (DNS).
What is DNS?
Direct numerical simulation of solid-liquid flows is a way of solving the initial value problem
for the motion of particles in fluids exactly, without approximation. The particles are moved by
Newton’s laws under the action of hydrodynamic forces computed from the numerical solution
of the fluid equations.
To perform a direct simulation in the above sense, therefore, one must simultaneously
integrate the Navier-Stokes equations (governing the motion of the fluid) and the equations of
rigid-body motion (governing the motion of the particles). These equations are coupled through
the no-slip condition on the particle boundaries, and through the hydrodynamic forces and
torques that appear in the equations of rigid-body motion.
These hydrodynamic forces and torques must of course be those arising from the computed
motion of the fluid, and so are not known in advance, but only as the integration proceeds. It is
crucial that no approximation of these forces and torques be made--other than that due to the
numerical discretization itself--so that the overall simulation will yield a solution of the exact
coupled initial value problem--up to the numerical truncation error.
Our goal is to do direct numerical solutions with many thousands of particles in three
dimensions, with large volume fractions, for the various kinds of suspensions and slurries that
model the practical particulate flows arising in applications like fluidized beds, slurry transport,
transport of drill cuttings for oil production, and proppant sands in reservoir stimulations. In this
article, we present the results of a simulation of the fluidization of 1204 spheres and other
simulations for targeted applications.
© 2002 Daniel D. Joseph 1 Updated 02/21/03 • Interog-1.doc

The hope is that the direct simulation of the motions of thousands of particles will, in many
cases, allow the large numbers of experiments used in deriving engineering correlations to be
replaced by cheaper, safer numerical experiments in which flow, material, and process-control
parameters can be altered with a computer command. With DNS you can turn physical factors on
or off to isolate effects, which is something that cannot be done in experiments. There are also
opportunities for the application of direct simulation to the diagnosis of industrial problems
involving flowing particulates, to the establishment of benchmark standards for two-phase flow
models (see Chapter VI), and to approximate numerical methods that track particles.
Approximate Methods
The signal feature of approximate numerical methods which track particles is that the forces
that the fluid exerts on the particle are modeled rather than computed as in DNS. Many excellent
numerical studies of particulate flows of many particles which are not direct simulations in the
above sense have appeared in recent years. These approximate methods include simulations
based on potential flow, Stokes flow, and point-particle approximations. They all simplify the
computation by ignoring some possibly important effects like viscosity and wakes in the case of
potential flow; inertial forces which produce lateral migration and across-the-stream orientations
in the case of Stokes flow; and the effects of stagnation and separation points in the case of
point-particle approximations.
Particle tracking methods take into account the particles and the fluid motion to understand
particulate flow. Particle tracking methods move the particles by Newton’s equations for rigid
bodies using forces that are modeled from single particle analysis or from empirical correlation
rather than from forces which are obtained by direct computation from the fluid motion. This
kind of simulation is called Lagrangian because the particles are tracked. The fluid motion in
particle tracking methods are computed from field equations, like the Navier-Stokes defined at
every point of the field including points occupied by the particles. This kind of computation from
continuum partial differential equations on a field is called Eulerian; hence, particle tracking
methods are Eulerian-Lagrangian. The particle tracking can have a one way coupling in which
the fluid motion is computed without particles, and a two-way coupling in which some effects of
the particle motion on the fluid motion are recognized. In most cases the two-way coupling is
introduced by momentum exchange terms representing the force of the particles on the fluid. A
more general coupling was introduced by Andrews and O’Rourke 1996 and Snider, O’Rourke
and Andrews 1998. They presented a scheme that considers the particle phase both as a
continuum and a discrete phase. Inter-particle stresses are calculated by treating the particle as a
continuum phase, as in the fluid phase equations of a two-fluid model. This way, they can track
the motion of the particles and at the same time model the inter-particle stress. In their scheme
the viscosity of the fluid in the fluid phase is neglected. The method can handle particulate
loading ranging from dense to dilute, and particles of different size and materials. The particles
are grouped into parcels. The motion of the parcels is tracked by the Lagrangian approach. The
number of parcels in a computational cell are used to calculate the solid phase volume fraction
on the Eulerian grid. Patankar and Joseph 1999, 2001 have extended this kind of particle-in-cell
method to account for fluid phase viscosity and other effects.
The problem of particles in turbulent flows has been treated using turbulent flow models for
the fluid with one-way coupling and two-way coupling with momentum exchange applicable to

dilute mixtures. These kind of turbulent flow models have been discussed by McLaughlin 1994
in a comprehensive review paper. Some of these turbulent flow models are conventionally called
DNS; they are not DNS for particulate flow in that the forces on the particles are modeled and
not computed from the fluid motion.
Another method used to compute solid-liquid is the Lattice Boltzmann method (LBM). This
is an unconventional computational approach that constructs simplified kinetic models for the
motion of discrete fluid particles. The LBM is inspired by Boltzmann’s equation which give rise
to the Navier-Stokes equations at 2nd order in Chapman-Enskog expansion and to Burnett
equations, which do not agree with experiments at 3rd order. The LBM gives rise to equations for
a compressible fluid, which are Navier-Stokes-like but not Navier-Stokes equations (see Qian,
d’Humieres and Lallemand 1992, Ladd 1994, How, Zou, Chen, Doolen and Cogley 1995, Chen
and Doolen 1998). The LBM may give rise to good approximations of isochoric flow when the
pressure gradients are not too large. For large pressure gradients the fluid will compress. The
LBM gives rise to particulate flows that are close to those computed by DNS (see Ladd 1997;
Aidun, Lu and Ding 1995, and Qi 1999). The good results achieved by LBM probably arise from
the fact that the forces on particles for this method are computed rather than modeled.
Data Structure for Applications of DNS

DNS generates huge amounts of data. It is necessary to post-process the data to get useful
results. The way the data is generated and processed is determined by the application; to
interrogate the data efficiently it is necessary to create a data structure for post processing. The
problem of setting up a data structure for the interrogation of results of simulations is a way of
determining the way that numerical simulations ought to be applied; it connects the
computational world to the applications.
The literature of DNS is by now fairly extensive. This literature is listed chronologically in
chapter IV. This same literature, together with the paper abstracts and animations of computer
simulations, can be found on our web page http://www.aem.umn.edu/Solid-Liquid_Flows.

Interrogations of DNS of Solid-Liquid Flows Equations of Motion
II. Equations of Motion

The equations to be solved in the ease of a Newtonian fluid are as follows. In the fluid we
have
u � 0 (II.1)
and
� �u �
Hf � � u � � u � � H f g � �p � D� 2 u . (II.2)
� �t �
For simplicity of presentation we consider spherical particles. The mass center of the spheres are
at X(t). The particles equations of motion are
dU d ω T [u]
M � Mg � F[u], I �
dt dt
(II.3)
dX dθ
� U, � M�
dt dt
where F[u] is the force and T [u] the torque on the particle. The fluid velocity is the same as the
particle velocity at the surface of the particle
u � U�ω r. (II.4)
The fluid force F[u] acting on the boundary of the particle is the integral over the body surface of
the traction
σ n � � pn � 2DD[u] n (II.5)
where
1
D[u] � (u � u T )
2
is the rate of strain and n is the outward normal and T [u] is the integral of the moment of the
traction vector.
Weak Solution for the Total Momentum

Numerical solutions of the Navier-Stokes are expressed first in terms of a weak solution,
which is in integral form and must be satisfied for all test functions in a certain space. The test
functions are chosen to convert weak solutions to a matrix in which solutions are obtained at
nodal points. In our work we use a special weak solution in which the fluid and particle
equations of motion are combined into a single weak equation of motion, which governs the
evolution of the total momentum of the system—fluid plus particles; the force F[u] and torque
T [u] cancel and do not need to be computed.
Find u, p, U, ω and satisfying

Interrogations of DNS of Solid-Liquid Flows Equations of Motion
� �u �
fluid
Hf �
� �t
� u � �u � g � � v dx
�
� p� � v dx � 2DD[u] :D[v]dx (II.6)
fluid fluid
� dU � dω
� M� � g� � V � I �ξ � 0 for all v, V , ξ,
� dt � dt
fluid
q u dx � 0 for all q , (II.7)
where u and v are both required to satisfy the no-slip condition on the particle boundary:
u � U�ω r v � V �ξ r. (II.8)
The function spaces for the functions and test functions are described in papers where they
are used.
The combined equation of motion (II.6), (II.7) and (II.8) in which hydrodynamic forces and
torques are completely eliminated was introduced by Hesla 1991 and implemented first by Hu
1996.
The same type of methods can be used for viscoelastic liquids; of course the equations for the
fluid motion are then different. For the Oldroyd-B fluids we have
du
Hf � �p � div τ (II.9)
dt

�1 τ� τ � 2D (D[u] � �2 D[u]) (II.10)
where
� (o )
(o ) � � u � � (o) � �u � (o) � (o) � �u T (II.11)
�t
is called the “upper convected derivative.” Here �1, is a relaxation time and �2 is a retardation
time. Newtonian fluids are recovered in the asymptotic limit �2 ��1, 0 � �2 � �1. It is convenient
for simulation to write
τ � τ E � 2DD[u] (II.12)
where

τ E � �1 τ E � 2D E D[u] . (II.13)
This leads to a formulation (III.1) of the problem in terms of the configuration tensor
�D �
A � τ E � �� E ��1 .
� �1 �
Interrogations of DNS of Solid-Liquid Flows Numerical Packages for Moving Particles in Direct Numerical Simulation
III. Numerical Packages for Moving Particles in Direct Numerical

Simulation
We developed two different kinds of numerical packages, called “particle movers,” one
based on body fitted moving unstructured grids and another based on fixed structured grids over
which bodies move by a method involving a system of Lagrange multipliers. Several different
versions of each kind of code have been developed and tried on a variety of applications.
ALE Particle Mover

Direct simulation of the motion of solid particles in fluids can be said to have started in the
paper of Hu, Joseph and Crochet 1992a. The first method Hu et al 1992a uses an implicit update
of the particle translational and angular velocities on unstructured grids (see figure III.1) to
prevent numerical instability. This is achieved by alternately computing the hydrodynamic force
and torque, then updating the particle translational and angular velocities using the equations of
rigid-body motion, and iterating until the translational and angular velocities converge. The
combined equations of motion (II.6, 7, 8) were used in Hu’s improved ALE scheme Hu 1996a,
2000 and Hu, Patankar and Zhu 2000.
The ALE particle mover uses a generalization of the standard Galerkin finite-element method
on an unstructured body-fitted mesh, together with an Arbitrary Lagrangian-Eulerian (ALE)
moving mesh technique to deal with the movement of particles (see, for example Hansbo 1992,
Huerta and Liu 1988, Nomura and Hughes 1992). In our implementation, the nodes on a particle
surface are assumed to move with the particle. The movement of the nodes in the interior of the
fluid is computed using a modified Laplace’s equation, to ensure that they are smoothly
distributed. At each time step, the grid is updated according to the motion of the particles. A new
grid is generated whenever the elements of the mesh get too distorted, and the flow fields are
projected onto the new grid.
The weak formulation of the ALE particle mover is based on (II.6, 7 and 8). These equations
are reduced to algebraic equations by finite element methods on the unstructured grid like that in
figure III.1. The strong form of these equations are the original equations (II.1 through 5) from
which the weak form was derived.
Two versions of the ALE method are the integrated method introduced by H. Hu 1996a, and
a splitting method implemented by H. Choi 2000. In the integrated method one solves for the
velocity and pressure at the same time; in the splitting method one solves sequentially separate
equations for the velocity and pressure.
The splitting method leads to a smaller system of equations than the integrated method;
however, the divergence free condition is not enforced at every sequential step so that the
velocity field which emerges from a divergence free step does not satisfy the momentum
equation exactly; hence small time steps are required to solve the momentum equations with
negligible error.

Figure III.1 Unstructured grid for the ALE method on a periodic domain.
The splitting method generates a symmetric saddle point matrix and leads to a symmetric
positive definite pressure equation. The pressure gradient is solved by a conjugate gradient
method without preconditioning because the matrixes involved are a composition of diagonally
dominant matrices and other matrices conveniently formed with diagonal preconditioners. The
methods run matrix-free of assembly of a global matrix.
A Projected Particle Mover

Matt Knepley developed a variation of the ALE particle mover, Knepley, Sarin, Sameh 1998,
in which the entire simulation is performed matrix-free in the space constrained to be discretely
incompressible. Apart from the elegance of this approach, it simplifies the model by treating the
particles and fluid in a decoupled fashion, and by eliminating pressure. The parallel multilevel
preconditioner due to Sarin and Sameh 1998a, is used to obtain an explicit basis, Pv, for the
discrete constrained divergence-free space. After elimination of pressure unknowns, a Krylov
~ ~
subspace method such as GMRES is used to solve the reduced system PvT APv x � b , where A is
the constrained Jacobian for velocity unknowns. In contrast to the ALE particle mover discussed
earlier, the linear systems in this method are positive-definite, and exhibit favorable convergence
properties on account of the well-conditioned basis Pv. The algorithm has demonstrated very
good scalability and efficiency for particle benchmarks on the SGI Origin 2000.
DLM Particle Mover

The DLM particle mover uses a new Distributed-Lagrange-Multiplier-based fictitious-
domain method. The basic idea is to imagine that fluid fills the space inside as well as outside the
particle boundaries. The fluid-flow problem is then posed on a larger domain (the “fictitious
domain”). This larger domain is simpler, allowing a simple regular mesh to be used. This in turn
allows specialized fast solution techniques. The larger domain is also time-independent, so the
same mesh can be used for the entire simulation, eliminating the need for repeated remeshing
and projection (see figure III.2). This is a great advantage, since for three-dimensional particulate
flow the automatic generation of unstructured body-fitted meshes in the region outside a large
number of closely spaced particles is a difficult problem. In addition, the entire computation is
performed matrix-free, resulting in significant savings.
The velocity on each particle boundary must be constrained to match the rigid-body motion
of the particle. In fact, in order to obtain a combined weak formulation with the hydrodynamic
forces and torques eliminated, the velocity inside the particle boundary must also be a rigid-body
motion. This constraint is enforced using a distributed Lagrange multiplier, which represents the
additional body force per unit volume needed to maintain the rigid-body motion inside the
particle boundary, much like the pressure in incompressible fluid flow whose gradient is the
force required to maintain the constraint of incompressibility.
The scheme uses an operator-splitting technique for discretization in time. (Operator-splitting

schemes have been used for solving the Navier-Stokes equations by many authors, starting, to
our knowledge, with Chorin, 1967, 1968, and 1973, and Teman, 1997.) The linearly constrained
quadratic minimization problems which arise from this splitting are solved using conjugate-
gradient algorithms, yielding a method that is robust, stable, and easy to implement. For further
details, see Glowinski, Pan, Hesla, Joseph and Périaux 1999. The immersed boundary method of
Peskin and his collaborators, Peskin 1997, 1981, Peskin and McQueen 1980, on the simulation of
incompressible viscous flow in regions with elastic moving boundaries also uses a fictitious-
domain method, but without Lagrange-multipliers.
The rigid motion constraint has been implemented in two ways leading to two codes, DLM1
and DLM2. The first implementation DLM1 Glowinski, Pan, Hesla and Joseph 1999 requires
that the fluid at the places P(t) occupied by the solid take on a rigid body velocity
DLM1 ux, t � U � ω r, x � P (t )
where U is the velocity of the mass center and ω r is the rotation around the mass center. The
second implementation DLM2 Patankar, Singh, Joseph, Glowinski and Pan 2000 is stress-like,
rigid motion on P(t) is enforced by requiring that the rate of strain vanish there
DLM2 D�u� � 0, x � P(t )
where D[u] is the symmetric part of �u.

Figure III.2 Fixed triangular grid used in DLM computation. The same grid covers the fluid and
solid. The fluid in the circles is bordered by Lagrange multipliers to move as a rigid body.
Parallel Implementation
The DLM particle mover uses an operator-splitting technique consisting of three steps. In the
first step, a saddle-point problem is solved using an Uzawa/conjugate-gradient algorithm,
preconditioned by the discrete analogue of the Laplacian operator with homogeneous Neumann
boundary conditions on the pressure mesh; such an algorithm is described in Turek 1996. The
second step requires the solution of a non-linear discrete advection-diffusion problem that is
solved by the algorithm discussed in Glowinski 1984. The third step solves another saddle-point
problem using an Uzawa/conjugate-gradient algorithm.
The DLM approach uses uniform grids for two and three-dimensional domains, and relies on
matrix-free operations on the velocity and pressure unknowns in the domain. This simplifies the
distribution of data on parallel architectures and ensures excellent load balance (see Pan, Sarin,
Glowinski, Sameh and Périaux 1999). The basic computational kernels, vector operations such
as additions and dot products and matrix-free matrix-vector products, yield excellent scalability
on distributed shared memory computers such as the SGI Origin 2000.
The main challenge in parallelization is posed by the solution of the Laplacian for the
pressure mesh that functions as a preconditioner for the Uzawa algorithm. Fast solvers based on
cyclic reduction for elliptic problems on uniform grids are overkill since the solution is required
only to modest accuracy. A multilevel parallel elliptic solver, Sarin and Sameh 1998b, has been
incorporated into the DLM algorithm. This has yielded speedup of over 10 for the
preconditioning step on a 16 processor Origin.
The parallel DLM particle mover has been used to simulate the expansion of a fluidized bed
discussed in chapter VII. Even though there is a serial component of the code, we have observed
an overall speedup of 10 on the SGI Origin 2000 at NCSA, using 16 processors. In addition, this
represents an impressive eight-fold increase in speed over the best serial implementation.

Viscoelastic DLM Particle Mover

The ALE particle mover has been implemented for the popular Oldroyd-B constitutive
model, which can be written in the form
� �u �
Hf � � u � � u � � ��p � � � (2DD � A),
� �t �
� � u � 0, (III.1)
� �A � D
�1 � � u � � A � A � �u � �u T � A � � A � E 1
� �t � �1
where D is the rate-of-strain tensor, A � J E � (D E / �1 )1 is the configuration tensor, JE is the

elastic stress, DE is the elastic viscosity, and �1 is the relaxation time. These equations are to be
solved subject to appropriate boundary conditions.
The system (III.1) is classified mathematically as being of composite type, Joseph 1990a:
The solution can have large gradients normal to the characteristic surfaces, which for this system
are tangent to the streamlines. A numerical error in resolving these sharp gradients can cause one
or more of the principal values of A, which are always positive in the continuous problem, to
become negative. This can cause a catastrophic amplification of short waves—a Hadamard
instability, Singh and Leal 1993.
This Hadamard instability can be prevented by ensuring that A remains positive-definite
using a method introduced by Singh in Singh and Leal 1993, Singh and Leal 1994. The method
has two key elements: a third-order upwinding scheme for discretizing the convection term in
(II.10) and a time-dependent solution algorithm which explicitly forces the principal values of A
to be positive. The combination of these two elements ensures that the scheme will remain stable
even at relatively large Deborah numbers. The equations are discretized in time using a second-
order operator-splitting technique that decouples the constitutive equation from the
incompressibility constraint. Singh, Joseph, Hesla, Glowinski and Pan 2000 have combined
Singh’s with the DLM method.
Computation domains for pipe flows of slurries

In studying slurries and other pipe flows it is necessary to set the computational problem on a
finite domain. It would be desirable to pose the problem as it is in nature using the end
conditions. The problem posed this way requires knowledge of end conditions and the details of
the motion may not determine the motion in the pipe, especially in long pipes. This same
problem occurs in analysis of pipe flows in which a constant pressure gradient is prescribed as
the ratio of the pressure drop �p over the length L of pipe. The prescription of that constant
pressure gradient is the only way in which the pressure drop is acknowledged. In analytical
studies, the remaining nonconstant part of the pressure gradient can be posed in a suitable class
of functions, say periodic in x. We could write
p( x, y, z , t ) � Px � 2( x, y, z , t ) (III.2)

where P is the constant pressure gradient and F ( x, y , z, t ) is periodic with an assigned period in
x. There are possibly different ways to choose this period and they are not equivalent. This is
exactly where we are in simulation. We decompose the pressure as in (III.2), then we require that
� and the velocity and other quantities are periodic in x with an assigned period. The assignment
of the period is made decisively by the construction of a periodic mesh like the one shown in
figure III.1. Clearly anything we compute on this mesh will have the assigned periodicity. The
choice of periodicity and the selection of the period does cloud the relevance of computed to
observed results and needs further study.
Simulations of single particle lift to equilibrium were also performed in a computational

domain, which moves in the x-direction and is such that the particle is always at its center. The
inflow and outflow boundaries are located at a specified distance from the center of the particle.
A fully developed parabolic velocity profile u(y) =�py(W - y)/2D corresponding to the applied
pressure gradient is imposed at the inflow and outflow boundaries.
Collision Strategies
It is not possible to simulate the motion of even a moderately dense suspension of particles
without a strategy to handle cases in which particles touch. In unstructured-grid methods,
frequent near-collisions force large numbers of mesh points into the narrow gap between close
particles and the mesh distorts rapidly, requiring an expensive high frequency of remeshing and
projection. A “collision strategy” is a method for preventing near collisions while still conserving
mass and momentum.
Four collision strategies are presently being used. They all define a security zone around the
particle such that when the gap between particles is smaller than the security zone a repelling
force is activated. A repelling force can be thought to represent surface roughness, for example.
The repelling force pushes the particle out of the security zone into the region in which fluid
forces computed numerically govern. The strategies differ in the nature of the repelling force and
how it is computed. It is necessary to compare these different strategies.
A collision strategy for the ALE particle mover used by Hu 2000 chooses the repelling force
so that the particles are forced just to the edge of the security zone. Another strategy for the ALE
particle mover, due to Maury and Glowinski 1997, uses the lubrication force of Kim and Karrila
1991, to separate touching particles and it is also the only strategy that requires touching particles
to transfer tangential as well as normal momentum. Maury and Glowinski developed his theory
for smooth bodies of arbitrary shape.
A yet different collision strategy has been implemented for the DLM particle mover,
Glowinski, Pan, Hesla and Joseph 1999. As in the other methods a security zone is defined. An
explicit formula, linear in the distance of penetration into the security zone, is used to keep the
particles apart. This repelling force may be tuned with a “stiffness” parameter.
Johnson and Tezduyar 1996, implemented a collision strategy based on the physics of
inelastic collisions in which a security zone is defined by structured layers of elements around
the particles. They model their strategy as an inelastic collision of elastic bodies with no fluid
present and use the coefficient of restitution as a fitting parameter. The collision strategy is
activated when the structured layers touch.

All of these strategies keep the particles farther apart than they ought to be, resulting in too
high void fractions. To have real collision of smooth rigid particles it is necessary for the film
between particles to rupture and film rupture requires physics and mathematics beyond the
Navier-Stokes equations. Without film rupture physics, the best that can be expected from
simulation is a strategy that allows for the action of lubrication forces to within the tolerance of
the mesh.
This kind of "security zone free" scheme has been devised and implemented by Singh, Hesla
and Joseph 2002. They modify the DLM method to allow the particles to come arbitrarily close
to each other and even slightly overlap. When conflicting rigid body constraints from two
different particles are applicable on a velocity node, the constraint from the particle that is closer
to that node is used and the other constraint is dropped. An elastic repulsive force is applied
when the particles overlap. The particles are allowed to overlap as much as one hundredth of the
velocity element size.
The modified DLM method was applied for both Newtonian and viscoelastic liquids.
Excellent results for particles in close approach and validation against analytical results were
achieved.

A. Addendum to Chapter III (Addendum A). Experiments on particle collisions in
viscous liquids; Stokes number
The physics of collisions is thought to correlate with the Stokes number
mv0 1 ρp
Stk = = Re
6π a µ 9 ρ f
2
here defined for a sphere of radius a, where v0 is the velocity of the sphere and
4 2av 0
m = π a 3 ρ p is its mass and Re = is the Reynolds number. The effects of particle
3 vf
momentum mv0 , collisions and rebound after collision are more important when Stk is
larger. In simulations, collisions which trigger an elastic repulsive force require this
selection of a coefficient of restitution, which is the ratio of the velocity before and after
impact. G. Joseph et al 2001, 2004, in the laboratory of Melanie Hunt, have done
experiments which clarify this point:
• G. G. JOSEPH, R. ZENIT, M. L. HUNT AND A. M. ROSENWINKEL. PARTICLE-WALL COLLISIONS IN A

VISCOUS FLUID, J. FLUID MECH. 433, 329-346 (2001).
• G. G. JOSEPH AND M. L. HUNT. OBLIQUE PARTICLE-WALL COLLISIONS IN A LIQUID. J. FLUID MECH.
510, 71-93 (2004).
(Prior literature on this topic can be found in the reference list of these two papers). They
find that
• For a particle colliding with a wall in a viscous fluid, the coefficient of restitution
(the ratio of the velocity just prior to and after impact) depends on the impact
Stokes number (defined from the Reynolds number and the density ratio) and
weakly on the elastic properties of the material.
• There exists a critical value for the Stokes number below which rebound does not
occur. This value is higher than the value predicted from hydrodynamics for zero
impact velocity.
• The slowdown of a particle prior to impact as it approaches a wall can be
computed from hydrodynamic theory to a good approximation
• Oblique collisions in a fluid obey the same laws as oblique collisions in a dry
system. However, the effective coefficient of friction at the point of contact is
drastically reduced due to lubrication effects.
Stated in another way, these experiments have shown that the elastic properties of the
particles and walls do not have a significant effect on the measured coefficients of
restitution. For a particle colliding with a wall in the normal direction, a deceleration was
observed due to the presence of the wall at Stokes numbers lower than approximately 70,
with rebound ceasing at approximately 10. The distance from the wall at which the
particle commences to decelerate increases with decreasing Stokes number. For Stokes
A-1 1/12/2005
numbers above 70, there is no apparent deceleration and above about 2000, the fluid
effects can be neglected. For numerical studies of solids in liquids, the forces arising
from lubrication are vastly more important than those arising from elasticity.
A-2 1/12/2005
Interrogations of DNS of Solid-Liquid Flows Addendum to Chapter IV (Addendum A)
IV. Weak and Strong Formulations of the DLM Method

The weak and strong forms of the equations for the DLM method are more interesting
because the Lagrange multiplier field must be introduced. There are two ways to represent the
rigid motion of a fluid in portions of the space occupied by solids; one way is to impose the
condition that the velocity is given by the velocity of the mass center plus a rigid rotation, and
the other is that on those portions of space, the rate of strain tensor vanishes. This gives rise to
two versions of the DLM method: a body force formulation and a stress formulation.
Body Force Formulation of DLM Method; Weak Solution

Find u, p, λ, U and ω satisfying
� �u �
H f � � u � � u � g � � v d x
9 � �t �
- p� � v d x � 2DD�u � : D�v �d x
9 9
� H f �� d U � d(Iω � ξ) �
+ ��1� �� M � � g�� V � �
� H s �� d t � d t ��
, H1(9), p
= λ � v � V � ξ � r d x , for all u, v in H1(�), p and � in L2(�), U, M, V, ξ in R3 (IV.1)
9s
q� � u d x � 0 ,
9s
q � H1(�) (IV.2)
µ � u � U � ω � r d x ,
9s
� � L2(�) (IV.3)
Here λ (x , t) is the Lagrange multiplier field, u, p, λ, U and ω are the unknown velocity,
pressure, U(t ) and ω(t ) are the velocity and angular velocity; v, V , ξ and µ are the test
functions, 9 s is the domain occupied by solids, and � � � f � � s is the domain occupied by
both fluid and solid, the entire domain shown as triangles in figure III.2.
Body Force Formulation of DLM: Strong Solutions

Using standard methods of the calculus of variations, we obtain the strong form of the DLM
equations of motion and constraint in fluids and solids. We get the Navier-Stokes over the whole

domain 9 but the Lagrange multiplier field appears only on the solid. We find that div u � 0 on
fluid and solid and
du
Hf � H f g � � � σ � λ@ � s (IV.4)
dt
where @ 9 s is one when x � 9 s and zero otherwise and σ � � p1 � 2DD�u� is the stress. Here
λ � a 2 2 λ is a body force. The solid is in rigid motion
u � U�ω r (IV.5)
λ satisfies a natural boundary condition
n � �λ � n � σ s � σ f (IV.6)
evaluating (IV.4) in 9 f , we find
div u � 0 ,
du
Hf � H f g � p �D 2 u (IV.7)
dt
in 9 f and from (IV.5), using no-slip, we get
u � U�ω r on �� f . (IV.8)
Equation (IV.7) and (IV.8) are a Dirichlet problem when U and ω are given and the
boundary of the particles defining �� f are known; if all these “given” were known as a function
of time we could solve any initial value problem in 9 f without reference to λ (x, t ) .
Using (IV.5) to evaluate (IV.4) on the solids 9 s we get
� dU dω �
Hf � � � r � ω � ω � r � g � � λ . (IV.9)
� dt dt �
Since there is no divergence constraint for λ we may put the pressure p in the solid to zero
and, of course, D�u� � 0 on rigid motions, and
n � � λ � pn � 2DD�u�� n (IV.10)
where p and u are the values of these fields in the fluid on the �� f of the solids. Once the fluid
flow associated with (IV.7,8) is solved, the right side of (IV.10) is known and the linear equation
(IV.9) may be solved for λ . The λ field is selected to make the fluid on the domain 9 s occupied
by solids move as a rigid body.

In the weak solution, the position of 9 s (t ) are updated using Newton’s equations (II.3) rigid
bodies. The form of the equations (II.3) is specialized for spheres; for arbitrary shape Newton’s
equations of motion are more complicated.
Stress Formulation of DLM: Weak Solutions

Patankar, Singh, Joseph, Glowinski and Pan 2000 have introduced a new formulation of the
distributed Lagrange multiplier/fictitious domain method for particulate flows. In this
formulation the deformation rate tensor in the fluid occupied by solids is constrained to be zero.
The Lagrange multiplier field λ turns out to be a velocity and the mathematics gives rise to a
stress like equation for λ in which the particle phase pressure may be put to zero.
The new formulation starts with the observation that on a rigid solid
D�u� � 0 (IV.11)
and a partial differential equation equivalent to (IV.11) is
� � D�u� � 0 in 9 s and D�u� n � 0 on �� s . (IV.12)
The combined form of the total momentum equation is formulated as follows:
Find u, p, λ, U and M satisfying
� �u �
H f � � u � � u � g � � v d x � 2DD�u� : D�v �d x � p� � v d x � q� � u d x
9 � �t � 9 9 9
� �u
H � H f �
�
+ s � u � � u � g � � v d x + D�λ � : D�v �d x � D�u � : D�µ �d x � 0 (IV.13)
9s � �t � 9s 9s
and
MU � H udx
9s
s
Iω � r x H s u d x . (IV.14)
9s

Stress Formulation of DLM: Strong Solutions

The strong form for (IV.13) may be found by the usual method of the calculus of variations.
In 9 f , we have div u=0,
du
Hf � �p � D 2 u � H f g . (IV.15)
dt
In 9 s , div u=0, D[u]=0 and
du
Hs � ��p �D� 2 u � H s g � � � D�λ � . (IV.16)
dt
The velocity and stress traction vectors are continuous across �� s � �� f
n � p f 1 � p s 1 � 2DD�u f � u s � � n D�λ �, �
� (IV.17)
u f � us . �
Noting next that (IV.5) holds in 9 s and on �� s , we arrive again at the Dirichlet problem
(IV.7) and (IV.8) for the Navier-Stokes equation in 9 f . After putting u � U � ω r into
(IV.16) and evaluating (IV.17), we get
� d U dω �
� � (D�λ �) � H s � � � r � ω � ω � r � g � (IV.18)
� dt dt �
n D�λ �s � n � p1 � 2DD�u � f . (IV.19)
This is a linear Dirichlet problem for λ when U(t), ω(t ) and the position of X and G of the
particles. This data determines u and p in 9 f through (IV.15) (or (VI.7)) and (IV.8).
D�λ � is the rate of strain for the Lagrange multiplier velocity field λ in 3D and three
equations (IV.18). It is not required that div λ � 0 ; in general div λ 0 . Hence a “pressure”
field is not associated with λ field. We might frame this result for two-fluid models of solid-
liquid flow which states the particle phase pressure vanishes.

B. Addendum to Chapter IV (Addendum B). Weak form of the body force
formulation when the Lagrange multiplier is eliminated
Using equation (IV.9) and (IV.10), we may eliminate λ in the weak solution.
∫ λ • ( v − (V + ξ ∧ r))dx
Ωs
• •
=ρf ∫ (U+ ω ∧ r + ω ∧ (ω ∧ r) − g) • ( v − (V + ξ ∧ r))dx
Ωs
(IV.20)
where
• dU • dω
U= , ω= .
dt dt
Noting now that Ω = Ω / Ω s ∪ Ω s , where Ω / Ω s = Ω f is the part of Ω occupied by the

fluid alone, we may write (IV.1), using (IV.20) as
∂u
∫ {ρ
Ωf
f [
∂t
+ (u • ∇)u − g ] • v + 2ηD[u] : D[ v] − p∇ • v}dx
∂u • •
+ ∫ρf[ + (u • ∇)u − U − ω ∧ r − ω ∧ (ω ∧ r )] • vdx
Ωs
∂t
ρf • d
+ (1 − )[ M (U − g) • V + (Iω) • ξ ]
ρs dt
• •
+ ρ f ∫ [U + ω ∧ r + ω ∧ (ω ∧ r ) − g ] • (V + ξ ∧ r )dx = 0 , (IV.21)
Ωs
where
I = ρ s ∫ [r 2 1 − r ⊗ r ]dx
Ωs
is the moment of inertia tensor.
The position vector r is measured from the mass center of the particle. Since ρ s is
uniform, the volume and mass centers are identical and
I’m indebted to Todd Hesla for his help with the reduction of the weak to strong solution and some of the calculations
leading to (IV.25).
© 2002 Daniel D. Joseph B-1 Updated 01/25/05 • AddendumA B.doc

∫ rdx = 0
Ωs
(IV.22)
where r = x − X(t ) and X is the center of mass.
Using (IV.22), we may simplify and rewrite the last term of (IV.21) as
• •
ρ f (U − g) • V ∫ dx + ρ f ∫ [ω ∧ r + ω ∧ (ω ∧ r )](ξ ∧ r )dx .
Ωs Ω
Noting next that M = ρ s ∫ dx , the terms proportional to V in (IV.21) are

Ωs
•
M [U − g] • V
and the terms proportional to ξ become
ρf d •
(1 − ) (Iω) • ξ + ρ f ∫ [ω ∧ r + ω ∧ (ω ∧ r )](ξ ∧ r )dx
ρ s dt Ω
which, after applying well known vector identities, becomes
ρf d • •
(1 − ) (Iω) • ξ + ρ f ∫ [r 2 ω − (r • ω)r + ω • r (r ∧ ω)] • ξdx
ρ s dt Ω s
or
ρf d ρf • ρf
(1 − ) (Iω) • ξ + (I ω ) • ξ + ω ∧ (Iω) • ξ . (IV.23)
ρ s dt ρs ρs
We note next that
d • dr dr dr
(Iω) = I ω + ρ s ∫ [ω(r • ) − (ω • )r − (ω • r ) ]dx .
dt Ωs
dt dt dt
Since
dr ∂ •
= [ + (U + ω ∧ r ) • ∇][x − X(t )] = − X+ U + ω ∧ r = ω ∧ r ,
dt ∂t
this reduces to

d •
(Iω) = I ω + ω ∧ (Iω) . (IV.24)
dt
After inserting (IV.24) into (IV.23), all of the terms except ξ • d (Iω) / dt vanish. The
terms proportional to V and ξ in (IV.21) therefore simplify drastically to
• d
M [U − g] • V + (Iω) • ξ .
dt
The reformulated statement of the weak form of our solid liquid system is obtained
after eliminating λ from (IV.1) by the method just given. The combined momentum
formulation, fluid plus solid, is
∂u
∫ρ
Ωf
f [
∂t
+ (u • ∇)u − g] • vdx + ∫ {2ηD[u] : D[ v] − p∇ • v}dx
Ωf
∂u • •
+ ∫ ρ[ + (u • ∇)u − U − ω ∧ r − ω ∧ (ω ∧ r )] • vdx
Ωs
∂t
• d
+ M (U − g ) • V + (Iω) • ξ = 0 , (IV.25)
dt
where the test functions v ∈ H 1 (Ω) 3 and V ∈ R 3 and ξ ∈ R 3 are vectors. The integral on
Ω s guarantees that the momentum on of the patches of fluid Ω s on places occupied by
the solids is for a rigid body motion. The fluid is weakly solenoidal
∫ q∇ • udx = 0 , q ∈ L (Ω)
2
and the condition
∫ µ • [u − (U + (ω ∧ r))]dx = 0 , µ ∈ H (Ω)
1 3
Ωs
guarantees that the fluid moves as a rigid body.

C. Addendum to Chapter IV (Addendum C). Recent Developments in Direct Numerical
Methods for Solid-Liquid flows
The method of direct simulation of solid-liquid flow has become very popular since
the first edition of this book in 2002. Here, I am going to give a short synopsis of some
of the main developments since 2002. These may be characterized as
(1) Extensions of DLM to different element bases

(2) Fast DLM methods based on quick generation of rigid body motions by “averaging”
fluid motions
(3) Explicit methods avoiding Poisson equations for the pressure using an equation of
state for weak compressibility
My book is not a source for the computational methods used to implement direct
simulation. These methods are described in papers cited throughout the book. For good
thorough reviews of the technical CFD issues confronting researchers, I recommend the
following papers. The ALE method and applications is treated by
• H. HU, N. A. PATANKAR AND M. Y. ZHU. DIRECT NUMERICAL SIMULATIONS OF FLUID-SOLID SYSTEMS

USING THE ARBITRARY LAGRANGIAN-EULERIAN TECHNIQUE. J. COMP. PHYSICS, 169, 427-462 (2001)
The DLM method and applications is treated by
• R. GLOWINSKI. FINITE ELEMENT METHODS FOR INCOMPRESSIBLE VISCOUS FLOW. IN PART 3 HANDBOOK
OF NUMERICAL ANALYSIS, IX, NORTH-HOLLAND, AMSTERDAM (2003) (ED. P. G. CIARLET AND J. L.
LIONS).
Fast computation methods are discussed in
• N. SHARMA AND N. PATANKAR. A FAST COMPUTATION TECHNIQUE FOR THE DIRECT NUMERICAL
SIMULATION OF RIGID PARTICULATE FLOW. J. COMP. PHYS. (2004).
• N. PATANKAR. A FORMULATION FOR FAST COMPUTATION OF RIGID PARTICULATE FLOW. CENTER FOR
TURBULENCE RESEARCH ANNUAL REPORT, STANFORD 185-196 (2001).
(1) Extension of DLM to different element bases
Even more recent advances in the DLM method are given in the papers listed below.
Extension of DLM calculations from finite element to spectral element bases were
carried out and applied by
• S. DONG, D. LIU, M. MAXEY, G. KARNIADAKIS. SPECTRAL DISTRIBUTED LAGRANGE MULTIPLIER

METHOD: ALGORITHM AND BENCHMARK TESTS. J. COMP. PHYSICS, 195, 695-717 (2004).
Abstract: We extend the formulation of the distributed Lagrange multiplier (DLM) approach for particulate
flows to high-order methods within the spectral/hp element framework. We implement the rigid-body motion
constraint inside the particle via a penalty method. The high-order DLM method demonstrates spectral
convergence rate, i.e. discretization errors decrease exponentially as the order of spectral polynomials increases.
We provide detailed comparisons between the spectral DLM method, direct numerical simulations, and the force
coupling method for a number of 2D and 3D benchmark flow problems. We also validate the spectral DLM
© 2002 Daniel D. Joseph C-1 1/25/2005

method with available experimental data for a transient problem. The new DLM method can potentially be very
effective in many-moving body problems, where a smaller number of grid points is required in comparison with
low-order methods.
Extension of DLM calculations from finite element to finite-difference bases were

carried out and applied by
• Z. YU, N. PHAN-THIEN AND R. TANNER. DYNAMIC SIMULATION OF SPHERE MOTION IN A VERTICAL

TUBE. J. FLUID MECH, 518, 61-93 (2004).
They did a very thorough study and obtained results in uniformly good agreement with
experiments. This paper treats sedimentation of a single sphere in tube flow. The
migration problem was considered by Yang et al (2005) for a neutrally buoyant using
both ALE and DLM methods (see addendum). Their results agree with those of Yu,
where they overlap. Yu et al 2004 gave correlations for their results like those proposed
earlier in various papers from our group but they get all their results from unconstrained
rather than constrained simulations used to generate correlations in the work of Yang et
al. The two papers mentioned here were done independently. The stress formulation of
the DLM method in strong form has been implemented in the animation work of
• M. CARLSON, P. MUCHA AND G. TURK. RIGID FLUID: ANIMATING THE INTERPLAY BETWEEN RIGID
BODIES AND FLUID. CONFERENCE PROCEEDINGS ACM TRANSACTIONS ON GRAPHICS, 23, 377-384 (2004).
Abstract: We present the Rigid Fluid method, a technique for animating the interplay between rigid bodies and
viscous incompressible fluid with free surfaces. We use distributed Lagrange multipliers to ensure two-way
coupling that generates realistic motion for both the solid objects and the fluid as they interact with one another.
We call our method the rigid fluid method because the simulator treats the rigid objects as if they were made of
fluid. The rigidity of such an object is maintained by identifying the region of the velocity field that is inside the
object and constraining those velocities to be rigid body motion. The rigid fluid method is straightforward to
implement, incurs very little computational overhead, and can be added as a bridge between current fluid
simulators and rigid body solvers. Many solid objects of different densities (e.g., wood or lead) can be combined
in the same animation.
The goal of their work focuses on computer generated animation. Their research follows
most closely the DLM work of Patankar et al 2000 and especially of the fast method of
Patankar 2001. They program the equations in strong form using finite differences rather
than finite elements. They apply their formula to several rigid bodies and not just to
spheres and their animations include free surfaces which are computed using the method
of level sets. Their paper is well organized and easy to follow.
(2) Fast DLM methods
N. Patankar 2001, introduced a fast method for producing rigid motions on the places
Ω s occupied by solids. His method is motivated by the Lagrange multiplier formulation
and can be expressed in this frame. However, in the actual implementation the
multipliers are not seen. Roughly, the computation proceeds as if there were no rigid
bodies. Then the places Ω s occupied by these bodies are identified and on them we can
assure that D[u] = 0 if the velocity u = v of the mass center of the rigid body and
angular velocity ω of the body around its mass center is selected as averages satisfying
the principles of conservation of linear and angular momentum
© 2002 Daniel D. Joseph C-2 1/25/2005

Mv= ∫ ρ p u ⋅ dΩ s
Ωs
and
Iω = ∫ x × ρ p u ⋅ dΩ s .
Ωs
The rigid motion u r = v + x × ω does not match the fluid velocity at every point on Ωs
but this discontinuity is corrected in Patankar’s scheme and the correction is smeared on
the length scale of the grid size.
A different fast method for the body force formulation of DLM has been given by
• R. GLOWINSKI, T. PAN, T. HESLA, D. JOSEPH AND J. PÉRIAUX. A FICTITIOUS DOMAIN APPROACH TO THE
DIRECT NUMERICAL SIMULATION OF INCOMPRESSIBLE VISCOUS FLOW PAST MOVING RIGID BODIES:
APPLICATION TO PARTICLE FLOW. J. COMP. PHYSICS, 169, 363-426 (2001).
A misprint in the fast method equations (7.33) and (7.34) in this paper was corrected in
Glowinski (2003). (see pages 724-726)
A comparison of the stress based DLM fast algorithm of Sharma and Patankar (2004)
and the body force based DLM fast algorithm of Glowinski et al (2001, 2003) is given in
the paper of Sharma and Patankar.
(3) Explicit methods avoiding Poisson equations for the pressure using an equation
of state for weak compressibility
A Poisson equation for the pressure
∇ 2 p = − div[(u • ∇)u]
arises from the Navier-Stokes equations for incompressible fluids for which div(u) = 0.
This is an elliptic problem and it generates difficulties for efficient computation which do
not arise when the pressure is given by an equation of state. One method of dealing with
this problem is to introduce and equation of state for weak compressibility perturbing
compressibility. The method of pseudo-compressibility, associated with the names of A.
Chorin and R. Temam, is well known to CFD experts, but it has not been applied to DNS
of particulate flow. The MacCormack scheme is an explicit method for compressible
fluids. It is essentially a predictor-corrector scheme, similar to a second order Runge-
Kutta method commonly used to solve ordinary differential equations. It is very easy to
program and it runs fast and well, is widely used and is widely respected.
• R. W. MACCORMACK “THE EFFECT OF VISCOSITY IN HYPERVELOCITY IMPACT CRATERING.” AIAA

PAPER 69-354, CINCINNATI, OHIO (1969).
Some progress has been made to adapting this scheme to particulate flow in weakly
compressible fluids.
© 2002 Daniel D. Joseph C-3 1/25/2005

• H. HU AND A. PERRIN. SIMULATIONS OF PARTICULATE FLOWS USING EXPLICIT MACCORMACK SCHEME.
IUTAM SYMPOSIUM ON “COMPUTATIONAL APPROACHES TO DISPERSED MULTIPHASE FLOW” ARGONNE
NATIONAL LABORATORY, ARGONNE OCT 4-7 (2004).
Abstract: Most incompressible flow solvers are implicit. Large systems of equations are constructed and solved,
demanding large amounts of memory and special schemes for parallelization. We can evade these difficulties by
solving flow problems based on small Mach number compressible Navier-Stokes equations with explicit finite-
differences on a fixed, uniform grid at each time step. This explicit scheme eliminates the Poisson equation for
pressure by relating pressure to density through an equation of state, then updating density with the compressible
continuity equation. The sound speed in the media imposes a constraint on the time step for the simulation. We find
empirically that for moderate Reynolds numbers (up to 200 based on the particle size), the CFL condition (based on the
sound speed) applies.
Advantages similar to the MacCormack scheme are enjoyed by the Lattice Boltzmann
method. This method can be described as a Galerkin approximate method based on the
Bhatnager, Gross, Krook 1954 (BGK) approximation for the Boltzmann equation. The
LBE method is a pseudocompressible method. Using a method of multiple scales, the
continuum equations implied by this method derived; these equations are compressible,
they have an equation of state and they are not the compressible Navier-Stokes equations,
but are Navier-Stokes like.
See equations (89) in
• R. R. NOURGALIEV, T. N. DINH, T. G. THEOFANOUS AND D. D. JOSEPH 2003. THE LATTICE

BOLTZMANN EQUATION METHOD: THEORETICAL INTERPRETATION, NUMERICS AND IMPLICATIONS.
INTERNATIONAL JOURNAL OF MULTIPHASE FLOW 29, 117-169 (2003).
The LBE method sometimes gives good results for solid-liquid flow because the particles
move under computed rather than modeled forces.
(4) Hybrid methods
A hybrid method is a numerical method which makes use of results from

mathematical analysis. J. Brady’s Stokesian dynamic simulations of particulate flow
embed analytical results from lubrication theory for near collisions of particles with
“Stokeslet” representations for far field effects. (See Brady 1993 for a recent review).
Another hybrid method is implemented in Physalis proposed by
• A. PROSPERETTI AND H. N. OGUZ. PHYSALIS: A NEW O(N) FOR THE NUMERICAL SIMULATION OF
DISPERSE SYSTEMS: POTENTIAL FLOW OF SPHERES. J. COMP. PHYS. 167, 196-216 (2001).
Z. Zhang and A. Prosperetti have recently reported the results of simulations of

sedimentation of 1024 spheres in a periodic domain at Reynolds numbers of about 40
(NOV. 21-23, 2004 MEETING OF DFD OF APS IN SEATTLE, WASHINGTON)
Physalis uses an analytical solution in the neighborhood of each particle and matches
this solution to the external field calculated numerically. For the sedimentation problem
the Stokes flow solution is used in the neighborhood of each sphere; locally because of
no slip, the fluid velocity is nearly zero in the frame of the moving particle. A boundary
© 2002 Daniel D. Joseph C-4 1/25/2005

layer would reduce the size of the Stokes flow layer at high global values of Re, but the
use of an analytical solution probably needs fewer mesh points on the sphere than a direct
method at any Re. It may be difficult to use this method on solids whose shape does not
allow a convenient analytical representation.
(5) Colloids and nanoparticles
Brady’s (1993) Stokesian dynamic method has been applied successfully to many
colloid problems at low global values of Re.
A promising new method which is not restricted to Stokes flow or to random forces
has been proposed by
• N. SHARMA AND N. PATANKAR. DIRECT NUMERICAL SIMULATION OF THE BROWNIAN MOTION OF

PARTICLES USING THE FLUCTUATING HYDRODYNAMIC EQUATIONS. J. COMP. PHYS. 201, 466-486 (2004).
Abstract: In this paper, we present a direct numerical simulation scheme for the Brownian motion of particles.
Solving the fluctuating hydrodynamic equations coupled with the particle equations of motion result in the Brownian
motion of the particles. There is no need to add a random force term in the particle equations. The particles acquire
random motion through the hydrodynamic force acting on its surface from the surrounding fluctuating fluid. The
random stress in the fluid equations is easy to calculate unlike the random terms in the conventional Brownian
dynamics type approaches. We present a three-dimensional implementation along with validation.
© 2002 Daniel D. Joseph C-5 1/25/2005

Applications of DNS
V Applications of DNS
The particle movers created in our Grand Challenge project were designed to (1) simulate the
remarkably different flow microstructures which arise from particle-particle and particle-wall
interactions in Newtonian and viscoelastic fluids, and (2) determine the effects of these
microstructures on anisotropic and other properties of flowing suspensions. These are computa-
tional studies of scale and structure; how do the effects of microstructure at the particle level
scale into flow effects at the slurry level?
§ Studies of Microstructure
There is a microstructure which is induced by the fluid dynamics of moving bodies and is
governed by a very simple principle: Long bodies are stable across the stream in Newtonian
fluids but along the stream in viscoelastic fluids (figure V.1). A key to understanding
microstructure in flowing suspensions of spherical bodies is the stable orientation of long bodies
and pairs of spherical bodies in momentary contact, Fortes, Joseph and Lundgren 1987, Hu 2001,
Huang, Feng and Joseph 1994, Huang, Hu and Joseph 1998, Joseph 1993, Joseph 1996, Joseph,
Fortes, Lundgren and Singh 1987, Leal 1980, Liu and Joseph 1993, Singh, Caussignac, Fortes,
Joseph and Lundgren 1989.
(a) (b)
Figure V.1: Cylinders falling in (a) Newtonian fluid (glycerin), and (b) viscoelastic fluid (2% polyox in
water). In glycerin, the cylinder is turned to the horizontal by inertia; in polyox, it is turned vertical
by viscoelastic pressures.
Particle pair interactions are fundamental mechanisms that enter strongly into all practical
applications of particulate flows. They are due to inertia and normal stresses and they appear to
be maximally different in Newtonian and viscoelastic liquids. The principal interactions between
neighboring spheres can be described as drafting, kissing, and tumbling in Newtonian liquids
(figure V.2a) and as drafting, kissing, and chaining in viscoelastic liquids (figure V.2b). The
drafting and kissing mechanisms involved are distinctly different, despite appearances, Fortes et
al 1987, Glowinski, Pan, Hesla, Joseph and Périaux 1998, Hu 2001, Joseph 1993, Joseph 1996,
Joseph, Fortes, Lundgren and Singh 1987, Joseph and Liu 1995, Joseph, Liu, Poletto and Feng
1994, Liu and Joseph 1993.
In Newtonian liquids, when one falling sphere enters the wake of another, it experiences
reduced drag and drafts downward into kissing contact with the leading sphere. The two kissing
printed 03/11/02 17 • Interog DNS SLF-2.doc
Applications of DNS
spheres momentarily form a single long body aligned parallel to the stream. But the parallel
orientation for a falling long body is unstable: hydrodynamic turning couples tend to rotate it to
the broadside-on orientation (perpendicular to the stream). The pair of kissing spheres therefore
tumbles to a side-by-side configuration. Two touching spheres falling side-by-side are pushed
apart until a stable separation distance between centers across the stream is established, Joseph
1996, Joseph, Liu, Poletto and Feng 1994, Singh et al 1989; they then fall together without
further lateral migrations.
(a) (b)
Figure V.2: (a) Spheres in Newtonian Fluids. Spheres fluidized in glycerin draft (i–ii), kiss (iii), and
tumble (iv–vi). They tumble because a pair of kissing spheres acts like a long body, which is unstable
when its long axis is parallel to the stream. The forces in a Newtonian fluid are dispersive; the
tumbling spheres are pushed apart (v–vi). (b) Spheres in non-Newtonian Fluids. Spheres falling in
2% polyox in water draft, kiss, and chain. They chain because the forces in a viscoelastic fluid are
aggregative. A chain of spheres acts like a long body, which is stable with its long axis vertical. The
chained spheres turn just like the solid cylinder in figure V.1b (i–vi). Reversing time, we see that
chaining, kissing and drafting in V.2b (vi–i) are like drafting, kissing, and tumbling in V.2a (i–vi)
This local rearrangement mechanism implies that globally, the only stable configuration is
the one in which the most probable orientation between any pair of neighboring spheres is across
the stream. The consequence of this microstructural property is a flow-induced anisotropy, which
leads ubiquitously to lines of spheres across the stream; these are always in evidence in two-
dimensional fluidized beds of finite-size spheres, Joseph 1993, Singh, Caussignac, Fortes, Joseph
and Lundgren 1989 (see figure V.3a). In viscoelastic liquids, on the other hand, two spheres
falling one behind the other will be pushed apart if their initial separation exceeds a critical
value. However, if their initial separation is small enough, they will attract (“draft”), kiss, turn
and chain, as shown in figure V.2b. One might say that we get dispersion in a Newtonian liquid
and aggregation in a viscoelastic liquid, Hu 2001, Hu, Joseph and Fortes 1992b, Joseph 1993,
Joseph 1996, Joseph and Liao 1994, Riddle, Narvaez and Bird 1977.
Many stable arrangements—i.e., steady particulate flows—like the “birds in flight” shown in
figure V.3c, and some bizarre steady arrangements of 2, 3, and 4 long cylinders (stable doublets,
triplets, and quadruplets, Joseph 1993, Joseph 1996) are seen in thin fluidized beds. The
Applications of DNS
arrangements displayed in figures V.1–3 have never been acknowledged in any two-fluid or
mixture-theory model of particulate flow. These models cannot predict such arrangements
because no provision is made for the forces that turn long bodies across or along the
stream; hence, two-fluid and mixture theory are silent about microstructure. This certainly
is a motivation for DNS, and suggests new challenges for mathematical analysis and two-phase
flow modeling.
(a) (b) (c)

Figure V.3: Stable arrays of spheres in (a) Newtonian, and (b) viscoelastic fluids. (a) Fourteen spheres
line up in a robustly stable array across the stream of a Newtonian fluid in a fluidized bed. (b) Seven
and four spheres fall in a viscoelastic fluid with their long axis vertical and parallel to the stream,
stably and permanently chained. (c) Spherical particles in a Newtonian fluid form like birds in flight.
When 22 < Re < 43, the spheres do not draft, kiss, and tumble. Three and four of them form a
permanent nested wake structure in which each successive sphere is nested in the wake of the one
before and rotates in a shear field there.
Particle-wall interactions also produce anisotropic microstructures in particulate flows, such

as clear zones near walls, and the like. If a sphere is launched near a vertical wall in a Newtonian
liquid, it will be pushed away from the wall to an equilibrium distance at which lateral
migrations stop. There is also an equilibrium distance for viscoelastic liquids to which spheres
will always migrate; this distance is often so close to the wall that spheres appear to be sucked all
the way to the wall, Huang Feng, Hu and Joseph 1997, Joseph, Liu, Poletto and Feng 1994.
These microstructural features ought ultimately to enter into understanding and control of the
lubrication of slurries.
We showed by mathematical analysis Huang, Hu and Joseph 1998, Joseph and Feng 1996,
and Joseph 1996 that the normal stresses on a rigid body in a viscoelastic fluid give rise to a
“viscoelastic pressure” proportional to the square of the shear rate at the particle boundary,
which is large where the velocity is large and zero at stagnation points where the pressure due to
inertia is largest. Thus, the viscoelastic pressure is large where the inertial pressure is small and
vice-versa, so that the turning couple on a long body has the opposite sign in a viscoelastic fluid
than it has in a Newtonian fluid—the body aligns parallel to the stream. Viscoelastic pressures
are compressive everywhere and tend to impel nearby bodies together (figure V.2b) whereas the
stagnation point pressures act at points toward the inside of near bodies and tend to push them
apart (figure V.2a). The pressures due to viscoelasticity are aggregative, gluing together the long
chain of spheres in figure V.3b. The chain is stable because it is a long body aligned with the
stream. The pressures of inertia are dispersive so that the array of spheres across the stream in
figure V.3a are permanently separated.

Applications of DNS
Figure V.4: Planes of falling spheres across the stream in a tube computed by the ALE particle mover. (a)
Five spheres (Re = 45); four are in a horizontal plane. The center sphere oscillates on the center line
relative to the plane. (b) Sixteen spheres (Re = 20) in two rings in a plane across the flow.
Most of the microstructural features just mentioned have been simulated by the ALE particle
mover. Some new arrangements are predicted. In the simulation of 5 spheres sedimenting in a
tube shown in figure V.4a, 4 of the spheres fall stably in a planar array perpendicular to the
stream, which fits well with the fact that the stable orientation of a long body is perpendicular to
the stream. The center sphere, however, does not remain in the plane; it either escapes by
dropping through the others, or oscillates back and forth across the plane of the other 4, along the
center line. This is a particle-flow realization of leapfrogging vortex rings. The sedimentation of
16 spheres leads to two rings of spheres in a plane perpendicular to the stream (figure V.4b).
Chaining appears in our simulations, and it does not destroy the mesh between particles, even
in some cases when no special collision strategy is used, Feng, Huang and Joseph 1996, Feng,
Joseph and Huang 1996 and Hu 2001. A periodic array of spheres will be sucked together to
form a long chain like that in figure V.3b. Prior to this computation the cause of chaining was in
dispute; many persons thought that the mechanism responsible for chaining was shear thinning
rather than the normal stresses which cause chaining in our simulations.
Cross stream migration and stable orientations of elliptic particles falling through an
Oldroyd-B fluid in a two-dimensional channel were studied using the ALE particle mover. There
are two critical numbers: the elasticity number and the velocity. For elasticity numbers below
critical, the fluid is essentially Newtonian and the ellipse falls broadside-on down the channel
centerline. For elasticity numbers larger than critical, the stable orientation depends on the the
velocity: if the velocity is below the shear-wave speed, so that the viscoelastic Mach number is
less than one, the ellipse falls down the channel centerline with its long axis parallel to the
stream. For larger Mach numbers, the ellipse flips to the broadside-on orientation again, Huang,
Hu and Joseph 1998. (An animation may be found in our Web site, http://www.aem.umn.edu/
Solid-Liquid_Flows.)
§ Sedimentation and Fluidization

Particle collision strategies may be put to a severe test in a sedimentation column. In the
sedimentation problem, we start with a crystal of close-packed particles at the top of the column;
then they settle under gravity resting in a crystal of close packed particles with defects on the
bottom. The defect structure ought to be related to the collision strategies used, but this has not
yet been tested. Fluidization is done in the same column. The particles first form a fixed bed at

Applications of DNS
the bottom; the inflow velocity is stepped up. First we find the so-called fluidizing velocity in
which the fixed bed breaks up and the particles fluidize. The bed height is an increasing function
of the inflow velocity; this gives the bed expansion by DNS. We have such a column in our
laboratory; it is 3:2″ wide, 0:3″ deep and 20″ high. In simulations and experiments we use 0.25″
spheres so that the motion of the particles is confined basically to two dimensions though the
fluid flow is in 3-D.
The fluidization in water of 1204 particles in three dimensions has been designated as a
benchmark problem for DLM particle movers in parallel implementation. The simulations and
experiments are discussed in Section VII. Animations of sedimentation and fluidization can be
found on our web page.
§ Mechanisms of Cross-Stream Migration

One of the uniquely useful features of DNS as compared with experiments is the ability to
isolate effects. Typically in a real viscoelastic fluid the effects of viscosity, shear thinning and
elasticity are all present. In simulations we may examine these effects one at a time: Newtonian,
generalized Newtonian (shear thinning), Oldroyd-B, and Oldroyd-B with shear thinning. Results
of 2-D simulation of 56 circular particles in a pressure driven spatially periodic channel flow,
computed with the ALE particle mover are shown in figure V.5, Huang and Joseph 2000. The
velocity profiles with and without particles are shown for each of the four cases. The difference
between these two velocity profiles can be attributed to an effective two-fluid effect in which the
particle-laden flow is regarded as an effective fluid with a higher effective viscosity. This higher
viscosity is partly due to a slip velocity in particles lag the fluid, holding the fluid back. The slip
velocity is the difference at a point between the average fluid and the average particle velocity
there.

Applications of DNS
10 10
(a) Fluid (n = 1.0) (b) Fluid (n = 0.5)
Particles Particles
8 8
6 6
Y(cm)
Y(cm)
4 4
2 2
d = 1 cm d = 1 cm
0 0
50 40 30 20 10 0 100 80 60 40 20 0
Velocity Profile U (cm/s) Velocity Profile U (cm/s)
y
10 10
(c) Fluid (n = 1.0) (d) Fluid (n = 0.5)
Particles Particles
8 8
6 6
Y(cm)
Y(cm)
4 4
2 2
d = 1 cm d = 1 cm
0 0
0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0
Velocity Profile U (cm/s) Velocity Profile U (cm/s)
y y
x x
Figure V.5: Cross stream migration of 56 neutrally buoyant circular particles in simulations of particle-
laden Poiseuille flow using the ALE particle mover. The velocity profiles without and with particles
are shown at the top of each sub-figure; underneath is a snapshot of the particle distribution. n is a
power law index. (a) Newtonian fluid. The particles remain well dispersed, but a lubrication layer is
evident. (b) Generalized Newtonian fluid. The particles migrate away from the center when the fluid
is shear thinning. (c) Viscoelastic fluid (Oldroyd-B). The particles migrate toward the center with no
shear thinning. (d) Viscoelastic fluid (Oldroyd-B). A clear annulus of fluid develops in the flow of an
Oldroyd-B fluid with shear thinning (n = 0.5).
Applications of DNS
§ Slot Problems for Particle Transport in Fractured Reservoirs

We are going to focus on the problem of proppant transport in hydraulic fracturing
applications. All of the features of slurry transport occur in hydraulic fracturing except that
transport under turbulent conditions is not common. To understand proppant transport it is
necessary to understand sedimentation, fluidization, particle migration and lubrication, lift-off
and resuspension, slip coefficients, and strategies for handling contacting bodies.
Hydraulic fracturing is a process often used to increase the productivity of a hydrocarbon
well. A slurry of sand in a highly viscous, usually elastic, fluid is pumped into the well to be
stimulated, at sufficient pressure to exceed the horizontal stresses in the rock at reservoir depth.
This opens a vertical fracture, some hundreds of feet long, tens of feet high, and perhaps an inch
in width, penetrating from the well bore far into the pay zone. When the pumping pressure is
removed, the sand acts to prop the fracture open. Productivity is enhanced because the sand-
filled fracture offers a higher-conductivity path for fluids to enter the well than through the bulk
reservoir rock, and because the area of contact for flow out from the productive formation is
increased. It follows that a successful stimulation job requires that there be a continuous sand-
filled path from great distances in the reservoir to the well, and that the sand is placed within
productive, rather than non-productive, formations.
Under the flow conditions expected within the fracture during pumping, the sand particles
migrate rapidly towards the center plane of the fracture, leaving a clear fluid layer at the fracture
walls, Karnis and Mason 1966, Nolte 1988a, Tehrani, Hammond and Unwin 1994. This clear
layer lubricates the motion of the slurry, and so increases the rate of gravity driven settling and
density currents. The net result of these processes is to cause sand to accumulate at the bottom of
the fracture and good vertical filling to be lost, Unwin and Hammond 1995. This in turn reduces
well productivity and can also interfere with the fracture growth process by blocking downward
extension.
The phenomenon of proppant migration is not currently controlled or exploited in the
fracturing industry because the relationship between migration and fluid properties is not
understood. DNS can give us this under standing (see figure V.5). The comparison of
simulations with experiments is essential when the suspending fluid is viscoelastic because the
constitutive equation for the fluid used in the experiments is never known exactly; it may be
adequate for some flows and not for others. This is to be contrasted with the situation for
Newtonian fluids, where a single constitutive equation applies in all the usual situations. It is
therefore extremely important to develop particle movers for the viscoelastic fluids which are
actually used in the fracturing industry and in other applications.
A typical vertical crack may be 3 meters high, 30 meters long and 2 cm wide. The diameter
of a typical sand grain is 2 mm, so that the width-diameter ratio h(x, y, t)/d is about 10. The sand
density is 2.4. Because of geological features related to the overburden, the preferred crack
orientation is vertical. Moreover, the dimensions of the fracture are not known a priori since the
crack opens and shuts in response to local changes of pressure; fracture dynamics determining
the slot dimensions is coupled to proppant transport.
In a slot problem a particle laden (say 20% solids) fluid is driven by a pressure gradient and
the particles settle to the bottom as they are dragged forward. Sand deposits on the bottom of the
slot; a mound of sand develops and grows until the gap between the top of the slot and the
Applications of DNS
mound of sand reaches an equilibrium value; this value is associated with a critical velocity. The
velocity in the gap between the mound and the top of the slot increases as the gap above the
mound decreases. For velocities below critical the mound gets higher and spreads laterally; for
larger velocities sand will be washed out until the equilibrium height and velocity are
reestablished, Kern et al 1959 (see figure V.6). The physical processes mentioned here are
settling and washout. Washout could be by sliding and slipping; however, a more efficient
transport mechanism is by advection after suspension which we studied by direct simulation.
Well Bore
Upper Fracture Boundary
Fluid
Sand Injected Sand Injected

Early Late
Figure V.6 Sand transport in a fractured reservoir.
§ Lift-off, Resuspension, Equilibrium Height, Slip Velocities, Lift-Force

Ratios
Using the ALE particle mover, we did lift-off and slip-velocity resuspension studies in
Newtonian and viscoelastic fluids. A heavier particle is resting on the bottom of a channel in the
presence of a shear (Poiseuille) flow at a certain critical speed. Depending on the weight and
diameter of the particle, the fluid properties, and the aspect ratio of the channel, the particle rises
from the wall to an equilibrium height at which the buoyant weight just balances the upward
thrust of fluid forces. We found that the upward thrust (lift) is due to inertia; over 70% of the
thrust is due to pressure. At equilibrium we compute the difference between the forward velocity
and angular velocity of the particle and these same velocities in the fluid (at the center of the
particle) when no particles are present. This gives the slip velocity and angular slip velocity
which are needed for Richardson-Zaki 1954 type of correlations discussed in Section 4.3.
The problem of inertial lift on a moving sphere in contact with a plane wall in shear flow has
been analyzed as a perturbation of Stokes flow with inertia in Cherukat and McLaughlin 1994,
Krishnan and Leighton 1995, Leighton and Acrivos 1985. These studies lead to specific and
useful analytic results expressed in terms of translational and rotational velocity and shear rate.
The lift on a stationary sphere on a wall in a shear flow varies as the fourth power of the radius
and the square of the shear rate. If the shear Reynolds number is sufficiently large, the lift force
exceeds the gravitational force and the sphere separates from the wall.
The Stokes flow analysis is not valid for lift off of proppants; for these heavy particles our
numerical results show lift off at shear Reynolds numbers in the hundreds. The perturbation
analysis are of considerable value because they are analytic and explicit even though they are
only valid well below the values characteristic of our applications.

Modeling and DNS
VI Modeling and DNS

One aim of direct numerical simulation (DNS) of solid-liquid flow is to guide the
construction of models and to generate correlations which in days gone by were always
generated by experiment. The single feature of such models is that the force that the fluid exerts
on the particles is modeled rather than computed as it is in DNS. There are two types of models:
(1)Mixture (two-fluid) theories which are generated by averaging or by ad hoc procedures.
These theories give rise to continuum theories in which the fluid phase and solid phase are
defined at the same point; they are widely used in industry. In the construction of such theories
interaction terms arise and their form is unknown. This is where the modeling comes in.
(2) Lagrangian numerical simulation (LNS). The fluid motion is resolved by DNS but the
particles are moved by modeled forces as in mixture theory and are not computed.
§ Averaging
In the past modeling was the main theoretical approach to solid-liquid flows. More recently
high speed computers and improved software have made possible the development of direct
numerical simulation of solid-liquid flow which can be used to validate models, to suggest new
models and in some situations to replace models. It is necessary to develop techniques to
interrogate direct simulation for useful descriptions, including those which impact modeling. It is
almost always the case that model equations are field equations defined at each space-time (x,t)
point. These equations give rise to the notion of interpenetrating continua in which actual
material points are no longer identifiable, the solid and liquid phases are both present at each and
every material point. The equations for such interpenetrating continua can be generated by
averaging. Averaged equations generate averaged variables but the resulting equations are not
simpler than the original equations from which they were derived. The averaging process
generated more variables, averages of products, etc. than the number of equations; it is necessary
to close the systems of equations and it is not possible for closure to apply to all flow. An
extensive literature has grown up around the process of averaging the equations of motion,
Anderson & Jackson 1967, Whitaker 1969, Wallis 1969, Delhaye 1969, Drew 1971, Drew &
Segel 1971, Hinch 1977, Nigmatulin 1979, Drew 1983, Arnold, Drew & Lahey 1989, Joseph &
Lundgren 1990, Wallis 1991, Drew & Lahey 1993, Zhang & Prosperetti 1994, 1997, Jackson
1997 and Drew and Passman 1999. Some authors have used local spatial averages taken over
regions small in extent compared to macroscopic length scales of interest, others have averaged
at each point of space over an ensemble of ‘macroscopically equivalent’ systems. Multiphase
turbulent models have been discussed by Roco & Balakrishnam 1985 and by Roco 1990; these
models give rise to predictions of the difference between the particle phase velocity and the
composite velocity (slip velocity) for a five phase mixture.
The averaged equations of motions introduce interaction terms between the solids and fluid
which and these terms must be modeled. The modeling of interaction terms in the averaged
equations of motion is a form of guided and intelligent guessing. This modeling rather than the
formal process of averaging is where the difficult part of the subject lies and compromises with
mathematical rigor are inevitable.
It is at first glance curious that a “two-fluid” model arises from averaging solid-liquid flows.
We have here continuum equations for the averages of the fluid, the fluid phase equations, and
Modeling and DNS
for averages over the solid, solid phase equations. The generation of such equations by ensemble
averaging is particularly transparent. Ensemble averages have the added advantage that they can
be formally related to discrete time averages which are readily generated by direct simulation on
fixed grids.
To generate ensemble averaged “two-fluid” equations we follow Joseph and Lundgren 1990.
We define an indicator function
ì0 if x is in the solid
H (x) = í (VI.1)
î1 if x is in the fluid
and let < > designate the operation of taking the average. The average is over many identical
trials. We think of an experiment which is started at a certain time. At a later time and at a certain
place, we record the value of some flow variable. We repeat the experiment, wait the same time,
look at the same place and record again. After many trials we average the values by summing
and dividing by the number N of trials, then we let N→∞. In this manner, we generate a function
< > (x, t). Now we get some identities using ensemble averaging and the indicator function. First
<H> = ε(x, t) = 1 – φ(x, t)

is the fluids fraction and
<1 – H> = 1 – <H> = φ(x, t)

is the solids fraction. Recall that V (x, t) is the true velocity. We define an average fluid velocity
< HV > < HV >

Vf (x, t) = = (VI.2)
<H > ε
and an average solid velocity
<(1–H)V> <(1–H)V>
Vs (x, t) = <(1–H)> = (VI.3)
φ
The composite velocity is
Vc (x, t) = <V> = <HV> + <(1–H)V> = ε Vf + φ Vs (VI.4)
We may define composite averages and mass averages of any quantity f by
fc = <f> = ε ff + φ fs ,
<ρf> (ρf)c
fm = = ,
<ρ> ερf + φρs
In particular the mass averaged velocity is
<ρV> ρfVf ε + ρs Vs φ
Vm = = (VI.5)
<ρ> ε ρf + φ ρs

Modeling and DNS
We next note H(x,t) is a material variable for materials which do not change phase, always
one following fluid particles, always zero following solids; i.e.
∂H
∂t + V • ∇H = 0 (VI.6)
Using this, and div V = 0, we find

∂H ∂H
0 = < ∂t + V • ∇H> = < ∂t + div HV>
∂<H>
= ∂t + div <HV>
∂ε
= ∂t + div ε Vf . (VI.7)
In the same way, we may show that
∂φ
∂t + div φVs = 0 . (VI.8)
It follows that
div Vc = 0 . (VI.9)
The reader can prove that
∂ρc
∂t + div (ρc Vm) = 0 . (VI.10)
We turn next to the momentum equations. Since

∂H
+ (V • ∇) H = 0 and div V = 0 (VI.11)
∂t
we have the identity
æ ∂V ö ∂HV
Hç + [V • ∇]V ÷ = ∂t + div H VV . (VI.12)
è ∂t ø
The momentum equation for the fluid and the solid is
æ ∂V ö
ρç + [V • ∇]V ÷ = div T* + ρb . (VI.13)
è ∂t ø
Multiply (VI.13) by H and ensemble average, using (VI.6),

∂
ρf ∂t <HV> + ρf div <H VV> = <H div T*> + ρf bf ε . (VI.14)

Modeling and DNS
Now we differentiate by parts
<H div T*> = div <HT*> – <∇H • T*> (VI.15)

where
∇H = δΣ (x) n , ∇<H> = <∇H> = <δΣn> = ∇ε (x, t) , (VI.16)
and δΣ (x) is a one-dimensional Dirac's delta function across the solid-fluid interface, n is the
outward normal to the solid. We next note that n • T* = t is the traction vector at a point xΣ on
the interface. From the definitions of Tf* we have
<HT*> = ε Tf* (x, t) (VI.17)
Using these relations, we may write (VI.14) as
é∂ ù
ρf ê∂t εVf + div <HVV>ú = div εTf* – <δΣ (x)t> + ρf ε bf . (VI.18)
ë û
Using the same method, we find a momentum equation for the solid in the form
æ∂φVs ö
ρs ç ∂t + div <(1–H)VV>÷ = div φ Ts* + <δΣ (x)t> + ρs φ bs . (VI.19)
è ø
Let us assume now that the fluid phase is Newtonian,
T* = –p1 + 2 η D [V] in the fluid, (VI.20)
and the solid phase is a rigid body for which

D [V] = 0 on solids, (VI.21)
where
1
D [V] = (∇V + [∇V]T) (VI.22)
2
is the rate of strain. The stress for the fluid phase is given by
Tf = εTf* = <HT*> = <H (–p1 + 2 η D [V])>
= –ε pf 1 + 2 η <(H–1) D [V]> + 2 η <D [V]>
= –ε pf 1 + 2 η D [<V>]
= –ε pf 1 + 2 η D [Vc] (VI.23)
where we have <V> = Vc = vc, and
<(H–1) D [V]> = 0 (VI.24)

Modeling and DNS
because H–1 is zero in the fluid and D[V] = 0 in the solid. The step <D[V]> = D[<V>] is true
because V is continuous, D[V] is uniformly bounded. In the modeling of multiphase flows of
rigid particles, it is convenient to write
T* = –p1 + τ (VI.25)
where p is the mean normal stress. The ensemble average of this is
Ts = φ Ts* = –φ ps 1 + <(1–H)ττ > (VI.26)
where
φ ps = <(1 – H) p > . (VI.27)
The expression (VI.26) with τ = 0 is frequently postulated in mixture theories (e.g., Nunziato,
Passman, Givler, MacTigire and Brady 1986). Givler 1987 argues that ps may be interpreted as
the average of the local pressure field around an isolated particle. Our DLM formulation
following (VI.19) suggests that p = 0. If we assume (VI.23), (VI.24), and (VI.26), and
manipulate the inertia terms to a more elegant form, we get the following system of ensemble
averaged equations.
∂ε
∂t + div ε vf = 0 , (VI.28)
∂φ
∂t + div φ vs = 0 , (VI.29)
æ∂vf ö
ρf ε ç ∂t + vf • ∇vf÷ + ρf div <H (V – vf) (V – vf)>
è ø
= –∇ (εpf) + η ∇2vc – <δΣ (x) t > + ε ρf bf , (VI.30)
æ∂vs ö
ρs φ ç ∂t + vs • ∇vs÷ + ρs div <(1–H) (V – vs) (V – vs)>
è ø
= –∇ (φ ps) + <δΣ (x) t > + div <(1–H) τ > + φ ρs bs . (VI.31)
The boundary conditions between the fluid and the particle takes form in the traction vector
term in (VI.30) and (VI.31). The addition of a constant pressure to the system as a whole, to pf,
ps and t simultaneously, has no dynamic consequence. The proof of this uses (VI.16) and it
works even if ps = 0.
When we add (VI.28) and (VI.29), we get
div vc = 0 .
When we add (VI.30) and (VI.31) we get an equation for the total momentum in which the
mutual forces (the traction vector) vanishes as it does in the weak solution formulation for DNS
(chapter II).
The existence of two fluid equations even when one of the fluids is solid is perfectly justified
by ensemble averaging. These equations, like other two fluid models, are not closed and methods
Modeling and DNS
of closure, or constitutive models for the interaction terms, are required to put the equations into
a form suitable for applications. Moreover, since averaging over repeated identical trials is not a
realizable proposition, the ensemble averaged variables are conceptually abstract and their
relation to more physically intuitive variables, like the ones which arise from spatial averaging, is
uncertain.
There are two kinds of two-fluid models of solid-liquid flow; one (more difficult) requires
the modeling of forces on the particles and a second in which the particle-laden portions of the
fluid are regarded as another fluid with different fluid properties, like an effective density and
viscosity. The first kind of theory can be framed as ensemble averaged equations, those just
given, in which there are more unknowns than equations. These equations must be closed by
modeling unknown terms; this modeling is difficult and can never cover all possible situations; it
is better to compute the forces. The second, effective theory is obviously a rather severe form of
modeling but it can work well in restricted problems as in the sedimentation of 6400 circles
studies in chapter VIII.
There are many problems of mechanics in which averaging of any kind is inappropriate,
leading to correct but irrelevant statements, like "the average weather is widely scattered
showers” or “the average gender is slightly female."
§ Fluidization by Drag and Fluidization by Lift

We are going to develop procedures for converting data from DNS to models for problems of
fluidization. Two kinds of problems are under consideration: fluidized beds in which the
particles are fluidized by drag forces opposing gravity and slurry beds in pipes and conduits in
which the flow is perpendicular to gravity and the particles are fluidized by lift forces.
Flows in which particles are suspended in a balance of drag and weight are fluidized
suspensions. Flows in which particles are suspended by lift forces perpendicular to the motion of
the fluid are fluidized by lift.
§ Fluidization and Sedimentation

In a fluidized suspension, the particles are stationary on the average and the fluid moves up
through them. It can be said that the particles are stationary under weight and drag. Fluidized
beds are amazing because of the action of hindered setting; a single particle will be held
stationary under a balance of weight and drag at one velocity and only one velocity. If the
velocity is smaller, the particle will fall to the bottom of the bed and if the velocity is larger, the
particle will be blown out of the bed. In the case of many particles, the situation is different; a
unique velocity is not required for fluidization. After the particles are fluidized, an increase in the
velocity will increase the height of the bed, the void fraction will increase, the center of mass of
the suspension will rise. The opposite is true of when the velocity is decreased.
It is not possible to fluidize a bed of heavy particles at low velocities. The bed fluidizes all at
once at the critical fluidization velocity; after this, the raising and lowering of the fluidizing
velocity raises and lowers the center of gravity of the fluidized suspension.
The remarkable properties of fluidized suspensions are associated with a hindered settling
function. Settling of a suspension at terminal velocity is like a fluidized bed in a coordinate
Modeling and DNS
system moving with the fluidization velocity; in this system the particles fall. The hindered
settling is controlled by the changes in the solid or void fraction. The velocity of the fluid
through the particles is greater than the input velocity because the flow must go between the
particles and the effective density and viscosity of the suspension is different than the viscosity
and density of the fluid.
§ Two Differences Between Fluidization and Sedimentation

The first difference is that in a fluidized bed you are driving the fluid even when there are no
particles. Hence there is a shear stress at the wall of the bed even when there are no particles. In
the case of sedimentation, there is no driving force when particles are absent, hence the wall
stress is absent in this case apart from that due to the falling particles. Sedimentation and
fluidization may be quite different when the wall effects are large.
A second difference is the contrary behavior of the top of a fluidized bed and the bottom
of a sedimenting bed. The top of a fluidized top is flat, perpendicular to gravity, more or less.
The less part can be seen when a particle is impelled out of the top by some hydrodynamic event;
it will always fall back because the drag on the single particle is less than the drag on the
particles in suspension. “Out of flat” dynamics is unstable and a top for an expended bed is a
robust feature of fluidization.
On the other hand, the bottom of a suspension of settling particles is always unstable
because the particles that fall away from the bottom experience a decreased drag, and will not
back into the suspension. At the bottom of the sedimenting bed you see the formation of fingers
like those in Rayleigh-Taylor instability when heavy fluid is over light.
§ Uniform Fluidization
By a state of uniform fluidization we mean a constant state in which time and space
derivatives of averaged quantities vanish. In this case, we may assume that all the terms in
equations (VI.30), (VI.31) vanish except the pressure gradient and the body force term. After
adding (VI.30) and (VI.31) written as
def
p = p c = εp f + φ p s , b = − e z g (VI.32)
we get
∆p
= ρ s φ g + ρ f εg (VI.33)
h
and when h = He
p1 − p 2
= ρ sφ g + ρ f εg = (ρ s Ω s + ρ f Ω f ) g Ω (VI.34)
He
where, neglecting wall friction,
p 2 − p a = ρ f g (L − H e ) .

Modeling and DNS
After noting that Ω = HA, Ω f = Ω − Ω s , we get
p1 − p 2 = ρ f gH e + (ρ s − ρ f )Ω s A . (VI.35)
L
p2
g h
He
p1
Figure VI.1 Cartoon of a fluidized bed expanded to height He. The total volume under Z = He is Ω = HeA
where A is the cross section of the bed; p is the pressure and pa is atmospheric pressure.
4
Since the volume, Ω s of solids under He is fixed (given by Ω s = πa 3 N for N uniform
3
spheres of radius a equation (VI.35) shows that the bed height He is a linear function of the
pressure difference; the bed height (or fluid fraction) increases with the applied pressure
difference.
The pressure difference, ∆p , between fixed stations h apart given by (VI.33) must decrease
because (ρ s − ρ f ) φ − ρ f will decrease with the solids fraction φ in an expanding bed.
§ Incipient Fluidization
If the pressure difference across the bed is not great enough to lift the particles plus fluid, the
particles will rest one on another in a fixed bed, a porous media satisfying Darcy’s law. The
height Hc when the spheres are close packed is given by
α = Ωs Hc A (VI.36)
where for a monosized dispersion of spheres α = 3/4. For incipient fluidization put He = Hc in
(VI.35), using (VI.36), to find that
æ1 ö
A( p1 − p 2 )c = gΩ s {ρ f ç − 1÷ + ρ s } (VI.37)
èα ø
The force (VI.37) is just what we need to overcome the overburden of fluid plus solids.

Modeling and DNS
Equation (VI.34) predicts what one expects to read as the pressure difference between
transducers at Z = 0 and Z = He. An equivalent formulation can be based on the dynamic pressure
gradient which is obtained from (VI.32) by splitting the total pressure
dp æ dp ö
=ç ÷ + ρf g (VI.38)
dz è dz ø f
into a dynamic part and a part balancing the liquid head. This gives
æ dp ö
[ ]
ç ÷ = ρ sφ + ρ f (ε − 1) g = (ρ s − ρ f )φ g. (VI.39)
è dz ø f
Richardson and Zaki 1954 note that the validity of (VI.39) was verified experimentally by
Wilhelm and Kwauk 1948 and by Lewis, Gilliard and Bauer 1949; it has been used frequently up
to this time but is valid only when the pressure drop due to the confining walls is negligible.
§ Hindered Settling
In one-dimensional approximations of fluidized suspensions, the drag is balanced by the
buoyant weight (VI.13). This is like pipe flows where the frictional pressure gradient is balanced
against the drag of the shear stresses at the tube wall; here it is the drag against the particles that
is reckoned to be most important.
Modeling of drag forces in one dimensional sedimenting or fluidizing suspensions is still a
controversial subject; many models have been proposed and there is no consensus. Some of the
more popular models are due to Barnea and Mizrachi 1973, Foscolo and Gibalaro 1984, 1987,
Batchelor 1974 and others.
The drag models proposed so far require expressions for the hindered sedimentation of
fluidization velocity of suspensions of particles. The composite velocity of a fluidized suspen-
sion of volume fraction φ
V (φ ) = Vc = Vsφ + V f (1 − φ ) (VI.40)
is the volume flux divided by the total area which is independent of Z, ∂Vc ∂Z =0, by (8), and
hence is equal to the superficial inlet velocity. The velocity
V(0) = Vf
is the velocity required to fluidize a single sphere. V(φ) and V(0) are also sedimenting velocities
for many and one single sphere, but here we shall carry out the analysis only for fluidization to
avoid confusion. We also define
f (φ ) drag on a sphere in a suspension of volume fraction φ,

f (0) drag on a single sphere φ = 0, ε = 1 ,
F (0) = 6φ f (φ ) πd 3 drag per unit volume.

Modeling and DNS
To explain the factors which lead to hindered settling we can consider the special case
described by Stokes flow. All of the above quantities depend on the Reynolds number R.
V ( 0) dρ f
R= .
ηf
For the Stokes flow R → 0 and [V0 (φ ), f 0 (φ )] = lim [V (φ , R), f (φ , R)]

R →0
In steady flow, the drag f 0 (0) balances the buoyant weight;
πd 3
f 0 (0) = 3πη f dU 0 (0) = (ρ s − ρ f ) g (VI.41)
6
where η f is the fluid viscosity.
To generalize (VI.41) for hindered settling effects we first note that there increase of velocity
through the particles (back flow in a sedimenting suspension), as that the velocity past the
particles is increased to

V0 = V0 (φ ) ε . (VI.42)
The viscosity of the suspension is increased to
η (φ ) = η f θ (φ ) (VI.43)
where θ (φ ) > θ (0) = 1 and the buoyant weight is reduced from the value on the left of (10) to
πd 3 ( ρ s − ρ c ) g 6 = πd 3 ( ρ s − ρ f ) ε g 6 = εf 0 (0) . (VI.44)
The replacement of ρ f by ρ c is controversial; different credible scholars have different

opinions. The experiments of Poletto and Joseph 1995 support the replacement of ρ f with
ρ c = ρ sφ + ρ s (1 − φ ) as long as the sedimenting sphere is not too much larger than the spheres in
the suspension.
Now rewrite the drag law (VI.41) replacing the viscosity η f with the effective viscosity
(VI.43) and the fluid density with ρ c to get

f 0 (φ ) = 3πdη f θ (φ )V0 = 3πdη f θ (φ )V0 (φ ) ε = εf 0 (0) = ε 3πdη f V0 (0) . (VI.45)
From (VI.44) it follows that
V0 (φ ) = (1 − φ ) 2 V0 (0) θ (φ ) = h(φ )V0 (0) (VI.46)
where h(φ ) is the hindered settling function.

Modeling and DNS
Equation (VI.46) rests on many unproven assumptions; if valid, it is valid only for slow and
only when wall effects are negligible. Barnea & Mizrachi 1973 give
h(φ ) = (1 − φ ) 2 (1 + φ 1 3 ) exp[5φ 3(1 − φ )]
where the exponential represents effective viscosity and φ 1 3 is used to represent the effect of
walls by close packed spheres.
A very accurate expression for hindered settling, which is not restricted to Stokes flow and
accounts for wall effects in cylindrical tubes of diameter D was obtained from extensive
experiments on the sedimentation and fluidization of monosized spheres by Richardson & Zaki
1954. They found that the composite fluidization velocity (or fall velocity of sedimenting
suspension) is given by
V (φ ) = V (0)(1 − φ ) n (VI.47)
where
n = 4.65 + 19.5 d when R = V (0) d ν < 0.2 ,

D
n = (4.36 + 17.6 d ) R −0.03 when 0.2 < R < 1 , (VI.48)

D
n = 4.45 R −0.1 when 1 < R < 500 ,
n = 2.39 when 500 < R < 7000 .
The terminal velocity V (0) of an isolated sphere should also be retarded by the effect of
nearby walls. One frequently used empirical formula
( ρ s − ρ f )d 2 æ dö
2.25
V0 (0) = ç 1 − ÷ (VI.49)
18η f è Dø
valid for slow flow, is due to Francis 1933.

The Richardson & Zaki correlation (VI.47) shows that the largest fluidization velocity V(0) is
for a single sphere or the equivalent which is an infinitely expanded bed in which one sphere
exerts no influence on another. When V > V(0) the spheres are dragged out of the bed, blown
away. In general V(φ) decreases as φ increases. The bed won’t rise if V > V(φ MP) where φ MP is
the packing of a fixed bed and V(φ MP) is the incipient fluidization velocity. In an expanded bed
V (φ MP ) ≤ V (φ ) ≤ V (0) .
The correlation (VI.47) and (VI.48) of Richardson & Zaki can be said to be the empirical
foundation of many engineering theories of fluidization and sedimentation.

Modeling and DNS
§ Drag and Hindered Settling

The relation between the drag f (φ) and composite velocity V (φ) is obtained by generalizing
the relation between the drag f (0) and the velocity V (0) for a single particle in an otherwise
particle free fluid in much the same way that we carried out this correlation for Stokes flow. A
popular empirical formula relating drag and velocity for a single particle is due to Dallavalle
1948,
2
æ 0.63 + 4.9 ö 8 f ( 0)
CD = ç ÷ = (VI.50)
è R ø ρ f πd 2V 2 (0)
where
V ( 0) d
R= (VI.51)
ν
and
πd 3
f (0) = (ρ s − ρ f ) g (VI.52)
6
balances the buoyant weight. Replacing ρ f with ρ c (φ ) for the suspension, we find that
2
ε æ ηf ö
f (φ ) = εf (0) = ρ f πd 2V 2 (0)ç 0.63 + 4.9 ÷ . (VI.53)
8 ç ρ f V ( 0) d ÷
è ø
After replacing V (0) = V (φ ) (1 − φ ) n from (VI.47), putting V (φ ) = Vc from (VI.40), (VI.53)

becomes
2
ρf 2 æ η f (1 − φ ) n ö
f (φ ) = πd 2 V c ç 0.63 + 4.9 ÷ . (VI.54)
8 (1 − V φ ) 2 n −1 ç
2
ρ f Vc d ÷
c
è ø
An important step the Richardson-Zaki correlation (VI.48) is all important. Expanding

(VI.54), we find, using n in (VI.48), that
ìï3.3πdη f Vc (1 − φ ) −3.78 small R ,

f (φ ) = í (VI.55)
ïî0.0463πρ f d Vc (1 − φ )
2 2 − 3.65 large R ,
Foscolo, Gibilaro and Waldman 1983 note that if we replace 2.39 with 2.4 and 4.65 with 4.8
in (VI.48) and 4.9 in (VI.54) with 4.8, then
ìï3πdη f Vc laminar,
f (φ ) = ε −3.8 í (VI.56)
ïî0.0463πρ f d Vc turbulent,
2 2

Modeling and DNS
Evaluating (VI.56) for φ = 0(ε − 1) we have
ìï3πdη f V (0)
f (0) = í . (VI.57)
ïî0.0463πρ f d V (0)
2 2
Since ε = (V (0) Vc )
1
n from (VI.47)
4.8
éV (0) ù n
ε 4.8
=ê ú . (VI.58)
ë Vc û
Obviously
4.8
é Vc ù n ìVc V (0) when n = 4.8
êV (0) ú =í 2 2 (VI.59)
ë û îVc V (0) when n = 2.4 .
Hence, using (VI.57) and (VI.59), we find that

4.8
é Vc ù n ìï3πη f Vc when n = 4.8
ê V ( 0) ú f ( 0) = í (VI.60)
ïî0.0463πρ f d Vc when n = 2.4
2 2
ë û
and
4.8
é V ù n
f (φ ) = ε −3.8 ê c ú f ( 0) (VI.61)
ë V ( 0) û
where n is the Richardson-Zaki exponent with 2.4 replacing 2.39 and 4.8 replacing 4.65, f (0)
and V (0) are related by the Dallavalle formula (VI.50) with 4.8 replacing 4.9.
f (φ) is a drag law which satisfies (VI.53); that is, f (φ ) = εf (0) which must hold when the
drag is balanced by the buoyant based on ρ c , and correctly reduces to (VI.56) for small and
large Reynolds numbers.
The drag law (VI.61) relies heavily on empirical data for “steady flow.” Dallavalle’s
expression is an empirical formula to fit experimental data for steady flow over a sphere and the
Richardson-Zaki correlation is for “steady” fluidization which certainly is steady only in a
statistical sense whose main consequence is that the average particle velocity Vs = 0 and
Vc = εV f . The reliance on good empirical correlations in making two-fluid models of solid-
liquid flow is to be commended. Approaches in this subject which purport to be from first
principles usually miss the mark; they are rigorous but remain at a level too general for practical
results or they give practical results only after making assumptions which are not rigorous. The
excellence of such theories ought to be validated by how well they correspond to what is known
for sure. For this validation direct simulation is a useful tool.

Modeling and DNS
§ Dynamic and One Dimensional Models

Foscolo and Gibilaro 1984 developed a one-dimensional model for the analysis of the
stability of fluidized beds. Batchelor 1988 developed a similar theory; he gave different
plausibility arguments to justify the terms, one-by-one in his equations. Both authors present a
one-dimensional equation for the particle phase; nothing whatever is said about the fluid phase.
This procedure is equivalent to a form of “pre-closure”, since the fluid phase is neglected, the
constraint ∂Vc ∂Z = 0 coming from (VI.9) is lost. In this case the composite pressure (or ps)
cannot be considered to be an unknown in the dynamical system for incompressible constituents,
but must be modeled as an “equation of state” associated with a compressible solids fraction
satisfying
∂φ ∂φVs
+ = 0. (VI.62)
∂t ∂Z
The momentum equation of Foscolo and Gibilaro 1984 is presented in the form
æ ∂V ∂V ö ∂π
φρ s ç s + Vs s ÷ = F + s (VI.63)
è ∂t ∂Z ø ∂Z
where π s = p s − P is the dynamic part of p s (due to motion) and ∂P ∂Z = −φρ s g balances the
weight of the particles. To close this system Foscolo and Gibilaro model the two terms on the
right side of (VI.63). They argue that
2
π s = − dF (VI.64)
3
and construct F from (VI.61) using the following plausibility argument. A particle will be
dragged up if
Vc − Vs = ε (V f − Vs ) > 0 . (VI.65)
In solid liquid models slip velocities may be defined in different ways; V f − Vs is often
called a slip velocity but Vc − Vs can also be called a slip velocity which reduces properly to Vc
for steady flow, when Vs = 0 . Foscolo and Gibilaro 1984 introduce and unsteady drag force
4.8
éVc − Vs ù n
f (φ ) = ε − 3.8
ê V ( 0) ú f ( 0) (VI.66)
ë û
such that f (φ ) = f (0) when. Vs = 0 In unsteady flow the force F on a single sphere is the
unsteady drag f (φ ) minus the buoyant weight given by (VI.44). After replacing f (0) with
(VI.52) they get

Modeling and DNS
æ 4.8
ö
πd 3 ç éVc − Vs ù n −3.8 ÷
ℑ= (ρ s − ρ f )gç ê ú ε −ε÷. (VI.67)
6 ç ë V ( 0) û ÷
è ø
The unsteady force F (φ ) per unit volume is Nℑ Ω , where N is the number of spheres in the
volume Ω = Ω s φ where φ = Ω s Ω and Ω s = Nπd 3 6 . Hence
æ 4.8
ö
ç éVc − Vs ù n −3.8 ÷
F = φ (ρ s − ρ f )gç ê ú ε −ε÷. (VI.68)
ç ë V ( 0) û ÷
è ø
Foscolo and Gibilaro 1984 argue that the same forces of the fluid on the particles acts at the
boundary to keep the particles from dispersing; to get this force we need to multiply the force per
unit volume by the ratio 2d 3 of the volume to the surface area of the sphere leading to (VI.67).
§ Particle Phase Pressure and the Stability of Uniform Fluidization

There is ever so much talk about the particle phase pressure. It boggles the mind to know
how such a quantity might be defined on a rigid body (of equation V.20). The choice of the
expression for the particle phase pressure in one-dimensional one-phase Foscolo and Gibilaro
1984, Batchelor 1988 and two-phase models Wallis 1969 determines whether or not the constant
state Vs = 0 and ε = ε 0 is stable. In a pioneering work Wallis 1969 derives a linear equation for
the perturbation ε ′ of ε 0 which he puts into the form
∂ ε′ ∂ 2ε ′ ∂ 2ε ′ æ ∂ε ′ ∂ε ′ ö
+ 2V 0 + A + Bç + Vω ÷ = 0. (VI.69)
∂t 2
∂Z∂t ∂Z 2
è ∂t ∂Z ø
This is a hyperbolic equation with forward and backward “elastic” waves associated with 2nd
order terms and a continuity wave associated with the first order terms. Instability occurs when
the square of the speed of the continuity wave is greater than the square of the speed of the
elastic wave. However, the coefficients in the Wallis equations are not precisely given as they
are in theories of Foscolo and Gibilaro 1984 and Batchelor 1988. Foscolo and Gibilaro argue that
∂F N ∂ℑ ∂ε
= (VI.70)
∂Z Ω ∂ε ∂Z
and they derive a criterion for the stability of the uniform state which they show agrees
reasonably well with observations of the formation of particle free three-dimensional regions
which resemble bubbles rising through liquids. Of course, strictly speaking a three-dimensional
disturbance cannot be described by a one-dimensional equation.
Batchelor 1988 does not argue his theory in the same way as Foscolo and Gibilaro 1984. He
does not believe in two-fluid models and argues out the terms of his equations by physical
reasoning. The term in his equation corresponding to the particle phase pressure is regarded by
him as arising from the diffusion of particles as in the kinetic theory of gases in which particles
diffuse into empty spaces. In the kinetic theory such diffusion arises from random walk
Modeling and DNS
dynamics arising from collisions. In liquid fluidized beds collision dynamics do not dominate
because there is a substance between the particles which keep them from touching. There is a
dispersive action to hydrodynamic interaction of particles in Newtonian liquids which is obvious
in experiments and direct simulation of particles large enough to see; but there does not seem to
be a tendency for particles to diffuse into empty places, in fact they fall in clusters. Probably
collisions do dominate in gas fluidized beds of heavy particles at high fluidization velocities; the
border between collision dominated and hydrodynamic dominated particle-particle interactions
has yet to be defined. Batchelor acknowledges that hydrodynamic interactions will keep the
particles from actually colliding but believes that velocity fluctuations might arise from
variations in the configuration of particles and the resulting hydrodynamic interactions or, in the
case of high Reynolds number flow around a particle, from turbulence in the fluid. Whether or
when the fluctuation level is large enough to produce diffusion of particles into empty spaces is
an open question.
After all the approximations are made, Foscolo and Gibilaro 1984 and Batchelor 1988 derive
an equation like (VI.69) with explicit but apparently different coefficients. However, the
diffusion parameter in Batchelor’s theory is unknown; Batchelor says it is known only up to an
order of magnitude and this parameter enters directly into his criterion for instability so that
Batchelor’s theory can be compared with experiments only in a qualitative sense. The predictions
of the Foscolo-Gibilaro theory are quite definite and are compatible with observations of
bubbling but these observations do not define the onset of bubbling sharply.
Joseph 1990b noted that if Foscolo-Gibilaro 1984 particle-phase pressure (VI.64) is
interpreted literally, then
∂π s 2 ∂F 2 é ∂F ∂ε ∂F ∂Vs ù
=− d = − dê + ú (VI.71)
∂Z 3 ∂Z 3 ë ∂ε ∂Z ∂Vs ∂Z û
leading to a term proportional ∂Vs ∂Z which was not considered in other theories. When this
term is factored into a stability calculation, uniform fluidization is always unstable. In fact, for
particles large enough to see in water, void fraction waves perpendicular to the flow rather than
bubbles can be seen and explained ultimately by the stable orientation of long bodies across the
stream. I would not say that such flows are uniformly fluidized.
The purpose of this rather long digression into modeling of fluidized beds is not to promote
one or the other model, but to show the reader how the models rest on unproven assumptions
which might be tested by direct numerical simulation.

Modeling and DNS
V Applications of DNS .....................................................................................................................17

§ Studies of Microstructure.............................................................................................................................. 17
§ Sedimentation and Fluidization .................................................................................................................... 20
§ Mechanisms of Cross-Stream Migration ...................................................................................................... 21
§ Slot Problems for Particle Transport in Fractured Reservoirs ...................................................................... 23
§ Lift-off, Resuspension, Equilibrium Height, Slip Velocities, Lift-Force Ratios........................................... 24
VI Modeling and DNS .......................................................................................................................25

§ Averaging ..................................................................................................................................................... 25
§ Fluidization by Drag and Fluidization by Lift .............................................................................................. 30
§ Fluidization and Sedimentation .................................................................................................................... 30
§ Two Differences Between Fluidization and Sedimentation.......................................................................... 31
§ Uniform Fluidization .................................................................................................................................... 31
§ Incipient Fluidization.................................................................................................................................... 32
§ Hindered Settling .......................................................................................................................................... 33
§ Drag and Hindered Settling .......................................................................................................................... 36
§ Dynamic and One Dimensional Models ....................................................................................................... 38
§ Particle Phase Pressure and the Stability of Uniform Fluidization ............................................................... 39
REMOVE THIS PAGE (only used to generate table of contents)

Direct Numerical Simulation of Fluidization of 1204 Spheres
VII Direct Numerical Simulation of Fluidization of 1204 Spheres
§ Experiments
Pan, Joseph, Bai and Glowinski 2002 have carried out simulations and experiments of the
fluidization of 1204 balls in a slit bed.
One goal pursued there is to demonstrate that numerical experiments on fluidization can be
processed in log-log plots for straight lines leading to power laws as did Richardson and Zaki
1954 for real experiments. As far as we know, we are the only group of researchers to carry out
this program. There is no prior literature, in which power laws are obtained from numerical
experiments. On the other hand, there are a number of numerical packages for particles in fluids
that might be used in this way. The methods of Stokesian dynamics (see Brady 1993) can be
recommended for problems in which inertia is absent. Hofler, Muller, Schwarzer and Wachmann
1999 introduced two approximate Euler-Lagrangian simulation methods for particle in fluids. In
one method, the particle surface is discretitized in grid topology; spheres are polygons on flat
places between nodes. In the second method, a volume force term is introduced to emulate rigid
body motion on the particle surface; this method is similar to the force coupling methods,
introduced by Maxey, Patel 1997. Hofler et al 1999 calculated sedimentation of 65,000 spheres
but at Reynolds numbers so small that it is essentially Stokes flow. Johnson and Tezduyar 1999
used a fully resolved DNS/ALE method to compute sedimentation of 1,000 spheres at Reynolds
numbers not larger than 10. A fully resolved method which is based on matching explicit Stokes
flow representations of flow near particles with computations on a grid has been proposed by
Ory, Oguz and Prosperetti 2000. The problem of particulates in turbulent flows has been
considered by a few authors: Crowe, Chung, Troutt 1996, McLaughlin 1994, Maxey, Patel,
Chang, Wang 1997; these approaches use point particle approximations because fully resolved
computations in turbulent flow are not presently possible.
The correlations of Richardson and Zaki 1954 are an empirical foundation for fluidized bed
practice. They did very many experiments with different liquids, gases and particles. They
plotted their data in log-log plots; miraculously this data fell on straight lines whose slope and
intercept could be determined. This showed the variables follow power laws. This same method
works for numerical experiments on fluidization and on the lifting of particles across streamlines
in Poiseuille flow (see Patankar, Ko, Choi, Joseph 2001, Patankar, Huang, Ko, Joseph 2001).
The existence of power laws can be regarded as a consequence of similarity Barenblatt 1996; it is
not an obvious consequence of the physics of flow of particulates or of the equations of motion.
The possibility that power laws underlie the flows of dispersions generally could be considered.
§ Numerical Method
To perform the direct numerical simulation of particulate flow, Glowinski et al 1997, 1998,
1999 have developed a methodology that is a combination of a distributed Lagrange multiplier
based fictitious domain method (DLM) and operator splitting methods. The basic idea is to
imagine that fluid fills the space inside as well as outside the particle boundaries. The fluid-flow
problem is then posed on a larger domain (the "fictitious domain”). This larger domain is
simpler, allowing a simple regular mesh to be used. This in turn allows specialized fast solution
techniques. The larger domain is also time-independent, so the same mesh can be used for the

entire simulation, eliminating the need for repeated remeshing and projection (see Figure VII.1).
This is a great advantage, since for three-dimensional particulate flow the automatic generation
of unstructured body-fitted meshes in the region outside a large number of closely spaced
particles is a difficult problem. In addition, the entire computation is performed matrix-free,
resulting in significant savings.
Figure VII.1. Part of a 2D example of a fixed triangular grid used in DLM computation. The same grid
covers the fluid and solid. The fluid in the circle is bordered by Lagrange multipliers to move as a
rigid body.
The velocity on each particle boundary must be constrained to match the rigid-body motion
of the particle. In fact, in order to obtain a combined weak formulation with the hydrodynamic
forces and torques eliminated, the velocity inside the particle boundary must also be a rigid-body
motion. This constraint in enforced using a distributed Lagrange multiplier, which represents the
additional body force per unit volume needed to maintain the rigid-body motion inside the
particle boundary, much like the pressure in incompressible fluid flow whose gradient is the
force required to maintain the constraint of incompressibility.
Concerning the space approximation of the problem by finite element methods, we use
P1-iso-P2 and P1 finite elements for velocity field and pressure respectively in the space
approximation like in Bristeau, Glowinski and Periaux 1987. Then we apply an operator-splitting
technique of Marchuk-Yanenko discussed in Marchuk 1990 for discretization in time.
(Operating-splitting schemes have been used for solving the Navier-Stokes equations by many
authors, starting, to our knowledge, with A. Chorin 1967, 1968, and 1973.) The linearly
constrainted quadratic minimization problems which arise from this splitting are solved using
conjugate gradient algorithms, yielding a method that is robust, stable, and easy to implement.
For further details, see Glowinski, Pan, Hesla and Joseph 1999. The immerse boundary methods
of C. Peskin and his collaborators, Peskin 1977, 1981, Peskin and McQueen 1980, on the
simulation of incompressible viscous flow in regions with elastic moving boundaries also uses a
fictitious domain method, but without Lagrange-multipliers.

The statement that DNS fully resolves the solid-liquid flow should be qualified to say
resolved up to the treatment of collisions. In the older versions of DNS previously mentioned,
particles were prevented from colliding by an artificial repulsive force. If we consider the
particular case of circular particles in 2-D or spherical particles in 3-D, and if Bi and Bj are such
two particles, with radiuses Ri and Rj and centers of mass Gi and Gj, we shall require the
H
repulsive force Fijp between Bi and Bj to satisfy the following properties:
→
ì0, if d ij > Ri + R j + ρ ,
ï
F =í 1
( )
ï ε (Gi − G j )(Ri + R j + ρ ) − d ij , if d ij ≤ Ri + R j + ρ
p
ij 2 2 (VII.1)
î p
where di,j = |Gi - Gj|, ρ is the force range, and εp is a given small positive “stiffness” parameter.
Gi
di Γj
G'i
Figure VII.2. Imaginary particle
For the particle-wall repulsive force, we take
→
ì0, if d i > 2 Ri + ρ ,
ï
F =í 1
( )
p
(VII.2)
ï ε (Gi − Gi′ ) (2 Ri + ρ ) − d i , if d i ≤ 2 Ri + ρ
ij 2 2
î w
where dj = |Gi – G'i| is the distance between the center of Bi and that of the imaginary particle at
the other side of the wall Γj (e.g., see Figure VII.2) and εw is another (small positive) “stiffness
parameter.” This force does not belong to the problem's description.
Fortunately the artificial repulsive force does not seem to have a big effect on the global
motion (see Figure VII.5). However, the implementation of a security zone has the unfortunate
consequence that particles cannot close pack. Though this does not seem to effect fluidized flows
greatly, we must be able to generate close packing if we are able to accurately model the
frictional resistances between close packed solids and walls.
Singh, Hesla and Joseph 2001 developed and applied a method in which repulsive forces are
activated only when particles touch; this strategy allows for hydrodynamic action to within the
tolerance of the mesh.
The DLM approach uses uniform grids for two and three-dimensional domains, and relies on
matrix-free operations on the velocity and pressure unknowns in the domain. This simplifies the
distribution of data on parallel architectures and ensures excellent load balance (see Pan, Sarin,

Glowinski, Sameh and Periaux 1999). The basic computational kernels, vector operations such
as additions and dot products and matrix-free matrix-vector products, yield excellent scalability
on distributed shared memory computers such as the SGI Origin 2000. A multilevel parallel
elliptic solver, Sarin and Sameh 1998b, has been incorporated into the DLM algorithm for two-
dimensional fluidized bed problems. This has yielded speedup of about 6 on 16 processors
compared with the elapsed time on 2 processors on an SGI Origin 2000 at the NCSA. In
addition, this represents an impressive eight-fold increase in speed over the best serial
implementation. Even though there is a serial component of the 3D code, we have still observed
that the speedup of 1.6 on 4 processors compared with the elapsed time on 2 processors on an
SGI Origin 2000. But no more speedup can be gained if we increase to 8 processors from 4
processors. All numerical results reported in this article are obtained on 4 processors on an SGI
Origin 2000 at the Minnesota Supercomputing Institute.
§ Experiments
We have carried out experiments of fluidization of 1204 spheres in a slit bed whose
dimensions are
[depth, width, height] = [0.686cm, 20.30cm, 70.22cm].
The cross-sectional area of the bed is
A = 0.686 × 20.30 = 13.32cm2.
We could not measure variations of the gap size inside the bed, but the glass plates were
pressured against 0.686cm spacers by aluminum screw clamps. The nominal diameter of the
spheres was
d = 0.635cm (1/4 inch).
The sphere diameters varied from 0.635 to 0.6465 (see Figure VII.3). The averaged diameter of a
sphere was
d = 0.639826cm.
The density of the nylon sphere is
ρs = 1.14 gm/cc.
The sphere was fluidized in water (we did not monitor the room temperature) whose density and
viscosity are
ρ f = 1 gm/cc, η f = 0.01 poise.

Average = 0.639826
25
20
Number (%)
15
10
0
0.634 0.636 0.638 0.64 0.642 0.644 0.646 0.648
Diameter (cm)
Figure VII.3. Distribution of diameter of spheres used in the fluidization experiment.
The Reynolds number based on the fluidization velocity for a single sphere is
V (0)d
R= ≈ 730 . (VII.3)
ηf ρf
Local Reynolds numbers in a fluidized suspension can be larger because of the back flow
through construction formed by nearly spheres. The velocity
V2 = 3.00 cm/sec (VII.4)
for incipient fluidization was identified roughly to within 0.1 cm/sec as the value in which
spheres more loosely packed in the fixed bed lifted slightly away from nearly spheres. According
to the Richardson-Zaki formula (VII.23), the velocity should vary between Vi and V (0) .
3 = Vi ≤ V (φ ) ≤ V (0) = 11.5 cm / sec . (VII.5)
Our experiments were consistent with this inequality.

The water is injected at the bottom of the bed through an array of plastic tubes under a
distributor and an eddy dampening screen. The resulting fluidizing velocity is not uniform but
there is no evidence of systematic anisotropy as the flow passes through the distributor screen.
Large eddies of hydrodynamic origin just exactly like the one shown in the 1204 sphere
animation on http://www.aem.umn.edu/Solid_Liquid_Flows/ are always present in the
experiments. The fluidizing velocity is computed from values of the mass flow rate measured by
collecting the weight of the overflow in a beaker over a fixed period of time. The mass flow rate
is

⋅
ρ f Q = W ∆t (VII.6)
⋅
where Q is the volume flow rate; W is the weight of fluid in the beaker collected over time ∆t .
The fluidizing (superficial) velocity is
⋅
V (φ ) = Q A (VII.7)
where A is the cross-sectional area 13.32cm2.

The height of the bed is measured by averaging the height of the top layer of particles. Stable
bed heights with large fluctuations were typical. The measured values of the bed height as a
function of the fluidization velocity V (φ ) are presented in Figures VII.4 and VII.17 where they
are compared with numerical simulation.
The solids fraction at a fluidization velocity V (φ ) is given by inverting the height H (φ )
Ω s 1204π d 3 6 4.437
φ= = = . (VII.8)
Ω AH H
§ Numerical Simulation
The calculation was carried using the distributed Lagrange multiplier method (DLM)
described in Section 2. The mesh size for velocity is
hV = 0.06858 cm
so the number of nodes is 3,348,675 (11 × 297 × 1025). The mesh size for the pressure h p = 2hυ
(458,622) nodes. The time steps used in the computation are either 0.001 or 0.0005 seconds, with
smaller time steps taken at times when the bed is fully expanded. The main parts of the
computation were carried out with a partially parallel code; the computation time for running it is
115 seconds per time step on four R12000 processors in a SGI Orgin2000 at the Minnesota
Supercomputing Institute. For example, the in case V = 4.5 cm/s, it took about 1660 hours to
reach time τ = 26 seconds in the simulation; this is 63.84 hours of computation time for 1 second
of real time.
In the simulation, the initial configuration of particles for the case of V = 3 cm/s is a square
lattice. The initial flow filed is zero everywhere for the case of V = 3 cm/s. Then we used the
results of V = 3 at t = 13 (resp., t = 19.5) as the initial conditions for the case of V = 3.5 (resp.,
V = 2). For the case of V = 4 (resp., V = 4.5), the initial conditions are obtained from the case of
V = 3.5 (resp., V = 4) at t = 16.15 (resp., t = 2). And finally for the case V = 5, the initial
conditions are obtained from the case of V = 4.5 at t = 27.2. The above choices of initial
conditions explains why the starting values of the bed height for different V are different in
Figure VII.4. For the parameters in (VII.1) and (VII.2), we took
εp = 5 × 10-7, εw = εp, ρ = hv = 0.06858cm. (VII.9)

The force range 0.06858cm is larger than the distance (0.686-0.635)/2 = 0.0255cm between a
centered ball and a side wall. Hence, in the simulation the balls are effectively centered between
the close walls by the particle-wall repulsive force. This centering mechanism is artificial; in the
experiments the balls can go closer to one wall or another. Therefore, the drag on the balls in the
experiment is larger than the drag in the simulation.
Figure VII.4 gives the bed height H(t) as a function of time for different fluidizing velocities.
The bed height is the average height of the top layer of spheres. The height rise curves have been
extrapolated to terminal rise for large times by a least square fit to a + b exp{-ct}. For the case
V = 3 cm/s we fit H(t) to a + b exp{-c(t-2)} for t > 2 because the bed height first decreases. The
terminal values
a = lim H (t )
t→0
are given by
V (cm/s) a (inches)
3 13.33
3.5 16.84
4 19.10
4.5 21.29
5 25.52

The history of the height of the fluidized bed

30
28
26
24 V = 5 cm/sec
Bed Height (inch)
22
(d) V = 4.5 cm/sec
20
(c) V = 4 cm/sec
18
16
(b) V = 3.5 cm/sec
14
(a) V = 3 cm/sec
V = 2 cm/sec
10
0 10 20 30 40 50 60 70 80
Time (sec)
Figure VII.4. Bed height H vs. time for different values of the fluidizing velocity V. H is the average
height of the top layer of 1204 spheres. The letters (a) … (d) index the snapshots shown in Figures
VII.9-10. The dashed lines - - - are least square fits to H(t) = a + b exp (-ct).
To test the effect of changing the size of the security zone in a relatively short time we
studied the fluidization of 150 spheres rather than 1204 spheres. The size of the security zone
was reduced to hv/2. Figure VII.5 shows that the change in the height rise is modest, the smaller
security zone allows for an increase in the bed height of about 2 or 3 percent.

The history of bed heights of 150 balls

18
16
14
12
Bed Height (cm)
10
0
0 5 10 15 20 25 30 35 40 45
Time (sec)
Figure VII.5. Bed height H vs. time for V = 4 cm/s. H(t) is the average height of the top layer of 150
spheres. The lines are least square fits to H(t) = a + b exp (-c(t-2)). Dashed - - - security zone 0.5hv;
solid — security zone hv.
§ Qualitative comparison of experiment and simulation

The simulation was carried out in a fluidization column whose coordinates are shown in
Figure VII.6.
X2,V2
d = 0.635 cm (1/4")
X3,V3
0.686 cm
X1,V1 20.30 cm
Figure VII.6. Coordinates and velocity components in the fluidization column. The fluidization velocity
is V ≡ V2.
In Figures VII.7 and VII.8 snapshots of simulation of 1204 spheres are shown in perspective
to emphasize that the simulation is three-dimensional. The video animations of these simulations,

which can be found at www.aem.umn.edu/Solid-Liquid_Flows, cannot be distinguished from real

experiments. In Figures VII.9-10 we compare snapshots of simulations in frontal view with
snapshots from experiments under equivalent conditions. We also compare snapshots of
simulations in frontal view of V = 4 cm/s with snapshots from experiments under equivalent
conditions for the case of V = 4.037 cm/s in Figures VII.11 and VII.12. We can find that the
simulation results do have features shown in the snapshots from experiments.
Figure VII.7. Snapshot of simulation of 1204 spheres with V = 4 (left) and a blow-up (right) at t = 32.
Figure VII.8. Snapshot of simulation of 1204 spheres with V = 4.5 (left) and a blow-up (right) at t = 31.

Figure VII.9. Cf. (a) in Figure VII.4 (left); cf. (b) in Figure VII.4 (right).
Figure VII.10. Cf. (c) in Figure VII.4 (left); cf. (d) in Figure VII.4 (right).

Figure VII.11. Comparison of snapshots of simulations of V = 4 at t = 2 (left) and 4 (right), and

snapshots from experiments under equivalent conditions for the case of V = 4.037.
Figure VII.12. Comparison of snapshots of simulations of V = 4 at t = 6 (left) and 9 (right), and

snapshots from experiments under equivalent conditions for the case of V = 4.037.
§ Numerical computation of averaged quantities

DLM produces huge amounts of data at each one of millions of nodes. The problem is how to
structure this data to extract useful information; we must decide beforehand what data to collect
as values to store for post processing. The fixed node property of DLM is particularly adapted to
the collection of data in a form suitable for averaging methods used to construct models.

To define a data structure we define a data string; this is a sequence of numbers produced at
that node. We call the number of values in the data string the number of hits. In these simulations
in which time steps are ∆t = 0.001 sec for start up and 0.0005 sec for the later times we may get
10,000 hits or more in a long simulation. Sometimes a particle is at the node, at the other times
fluid is there. By processing data at hits we can create time averages without significant cost in
computational cost. Suppose a solid is at the node M times and the fluid is there M ′ times,
M + M ′ = 10 4 . Then φ = M 10 4 and ε = M ′ 10 4 , ε + φ = 1 gives the volume fractions. If V(x,t)
is the component of velocity parallel to gravity and x is at a node and V1(x,t) are a data string of
hits. Then
M
1
V s ( x, t ) =
M
å V ( x, t )
i =1
i (6.1)
where x is at a node; obviously
1 M′
V f (x, t ) =
M′
åVi (x, t ) . (6.2)
Then, at the same node, we ought to find
V (φ ) = Vsφ + V f ε . (6.3)
After transients have disappeared in the fluidization we ought to find that Vs = 0 .
In the same way, we can find that a data string of values of the angular velocity as a
difference of the velocity at a point in the solid and its mass center and so on for other averages.
The time averages formed from data strings can be thought to be ensemble averages on non-
transient flow and may be assumed to be local time averages on intervals with a sufficiently large
number of hits which is small relative to the length of time of transients. By repeating
simulations we could actually initiate the procedure used to generate ensemble averages. There is
a mathematical literature on the relation of time averages to ensemble averages which is rather
theoretical and involves assumptions of a mathematical nature which are difficult to verify.
In our simulation of 1204 spheres we created such data strings at 38 nodes on a line across
the width of the bed at a height of 10.179 in. in a plane in the center of the depth, 0.135 in. from
each wall. The width of the bed is 8 in. and the 38 nodes plus 2 wall points means that the nodes
are separated by 1/10 in. The number of hits was M = 7365 for the case V = 4.5 cm/sec taken
every 4 time steps. In Figure 13 we present the solid fraction φ and the three components of the
average solid Vs, liquid Vf and composite velocity V(φ) = Vsφ + Vf (1-φ) for the case V = 4.5 cm/s.
The velocity components V1, V2 and V3 correspond to coordinates in the cross section of the
fluidization column shown in Figures 13(b)—13(d).
The data presented in Figure 13 shows that the dynamics of the bed are strongly two-
dimensional, the velocity component V2 is in the vertical direction; in the one-dimensional
approximation the composite velocity
V2C (φ ) = V2 S (φ ) + V2 f (1 − φ )

would equal V(φ). Obviously V2 = V2 (x3) because of a large circulating eddy which is apparent
on the video animations on http://www.aem.umn.edu/Solid-Liquid_Flows and in experiments. It
is also apparent that the difference between averaged fluid velocity V2f and the averaged solid
velocity V2S is positive; the solid lags the fluid by 3 to 5 cm/s. The transverse components of
velocity V1 and V3 are basically zero which is an assumption one makes in a one-dimensional
theory. The fluctuation level of V1 is very low because of the collision strategy interacting with
nearby side walls, but the particles are not rigorously centered by these artificial collision forces.
It would be desirable to have many more points in our data strings. Greater computational
efficiency and speed is a challenge for the future.
Figure VII.13 shows a strong two-dimensional variation of average equations across the slit
column which is suppressed in one dimensional studies.
Volume Fraction
(a) 1 1.0
(b)
0.9 Averaged fluid velocity
0.8
Averaged solid velocity
0.8 Composite velocity
0.6
0.7 0.4
f (solids fraction)
0.6 V1 (cm/sec) 0.2
0.5 0
0.4 -0.2
0.3 -0.4
0.2 -0.6
0.1 -0.8
0 -1.0
0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8
Channel width (inches) Channel width (inches)
(c) 20 5
(d)
Averaged fluid velocity 4
15
Composite velocity 3
2
10
1
V3 (cm/sec)
V2 (cm/sec)
5 0
-1
0
-2 Averaged fluid velocity
-3
-5 Composite velocity
-4
-10 -5
0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8
Channel width (inches) Channel width (inches)
Figure VII.13. Volume fraction φ and averaged components of velocity V1, V2, V3 in coordinate
direction x1 , x2 , x3 at 38 nodes spaced at 1/10 in. across the 8 in. side of the fluidization column
when V = 4.5 cm/s. (a) φ, (b) V1 , (c) V2 , (d) V3. The difference between the averaged fluid and
averaged solid velocity is the slip velocity. This is the first exact calculation of the slip velocity in a
fluidized suspension.

§ Dynamic and wall friction pressure in a fluidized bed

The stress in an incompressible Newtonian fluid is given by
σ = -P1 + 2µD[u] (VII.13)
where D[u] is the symmetric part of the velocity gradient ∇u and P, the total pressure, is the
mean normal stress. The equation of motion in the z direction is
dw dp
ρ = − + µ∇ 2 w (VII.14)
dt dz
where w = ez•u and p = P + ρfgz is the dynamic pressure that we compute in our DNS simulation.
dp
We may compute an average dynamic pressure gradient by averaging over cross sections
dz
and time. This quantity then is given by DNS and
dp dP
= + ρf g . (VII.15)
dz dz
We may decompose
p = pw + p s (VII.16)
where pw is the wall friction pressure and ps is the pressure required to fluidize the spheres. In
fluidized bed practice it is assumed that when wall friction is negligible the pressure gradient
= −(ρ Pφ + ρ f ε )g = − ρ c g
dP
(VII.17)
dz
balances the composite weight of fluid plus solids. In this case
dP dp dp
= − ρ f g = s − ρ f g = −ρc g (VII.18)
dz dz dz
Hence
= −(ρ P − ρ f )gφ
dp s
(VII.19)
dz
This expression was verified in experiments of Wilhelm and Kwauk 1948 and Lewis, Gilliand
and Bauer 1949. Combining now (VII.16) and (VII.19) we get
− (ρ P − ρ f )gφ
dp dp w
= (VII.20)
dz dz
When φ = 0, the dynamic pressure is equal to the wall friction pressure.

Equation (VII.20) is an equation for the wall friction pressure gradient with values of dp dz
and φ given by DNS. Table VII.1 gives the values of terms in this equation for different
fluidizing velocities. The table shows that the wall friction pressure gradient is about 1/4 of the
pressure gradient needed to fluidize the spheres. As V increased, φ decreases and the pressure
gradient − (ρ P − ρ f )gφ to fluidize spheres decreases. The small decrease in the wall friction
pressure gradient with fluidizing is surprising; the wall friction should go up as the speed
increases. We conjecture that the decreased friction is due to a decrease in the effective viscosity
µ(φ) of the mixture as φ is decreased; the viscosity of densely packed mixtures is greater.
100 0.4
0.35
80
0.3
Pressure gradient
60 0.25
Solid fraction
Dynamic
f
pressure gradient 0.2
40
0.15
Wall friction 0.1
20 pressure gradient
0.05
0 0
2.5 3 3.5 4 4.5 5 5.5
V (cm/s)
Figure VII.14. Solid fraction f dynamic pressure gradient and wall friction pressure gradient vs.
fluidizing velocity.
The calculation given here points a new direction for the interrogation of DNS for new
results in continuum engineering descriptions. Our conclusions are tentative because the results
about the wall friction pressure gradient depend on the accuracy of our numerical simulation near
walls. It is certainly true that the activation of a repelling force when the particle approaches the
wall reduces accuracy. We are confident that the size of the security zone can be reduced to zero
by techniques under development so the full hydrodynamics of lubricating flow can be captured
up to mesh resolution.
Table VII.1.Terms in the dynamic pressure gradient equation (VII.20). ρp = 1.14 g/cc, ρf = 1 g/cc.
V (cm/s) dp φ ρcg (ρp - ρf)gφ dp w dp

- (dynes/cm2) − =− - (ρp - ρf)gφ
dz dz dz
3 63.570 0.3582 1050.1 49.177 14.393
3.5 52.050 0.2956 1021.2 41.246 11.463
4 44.694 0.2439 1014.2 34.155 11.202
4.5 39.171 0.2119 1009.8 29.758 10.076
5 34.990 0.1842 1006.0 25.957 9.705

§ Sedimentation and fluidization velocity of single spheres

We did simulations and experiments of the sedimentation and fluidization velocity of single
spheres. It is sometimes assumed that these two velocities are the same but in general this is true
only in very special cases.
The flow of fluid up the slit bed is close to a developing Poiseuille flow in which the velocity
of the fluid vanishes at the wall but not at the center. In the sedimentation case the fluid does not
move unless disturbed by a falling particle. The flow of fluid in a fluidized bed need not be fully
developed. The flow profile can change from station to station in the fluidized bed case, but not
in the sedimentation case.
In the ideal case there is one and only one fluidization velocity and, if we ignore the flow
variations just mentioned, this velocity is the sedimentation velocity; in both cases the drag
balances the buoyant weight. For steady flow fluidized and sedimenting particle velocities are
equivalent under Galilean transformation. In fact we do not verify this ideal case in
sedimentation or fluidization; a unique velocity does not emerge as can be seen from the
experimental results given in Figure VII.15.
- Escape fluidization
- Start fluidization - Sedimentation
12
11
Velocity (cm/s)
10
7
6.36 6.38 6.4 6.42 6.44 6.46 6.48
Particle size (mm)
Figure VII.15. Sedimentation and fluidization velocities from experiments using different size
spheres. The particle lifts off the distributor when V is greater than and is dragged out of the
column only when V is greater than .
In Figure VII.15 we have plotted fluidization and sedimentation velocities for different
spheres. Focusing first on sedimentation we note that even when we drop the same sphere in the
quiet slit bed, the sedimentation velocity differs from trial to trial. Variations of as much as 7%
are observed. How do we account for such variations?

As a practical matter our slit bed has a nominal gap size of 0.686cm. The gap size is certainly
not uniform; perhaps the gap size varies between 0.6778 and 0.7239cm. Obviously this variation
will lead to a variation in the sedimentation velocity.
A more fundamental reason for the variability in the sedimentation rate is that the motion at
Reynolds numbers in the thousands is not steady and is probably chaotic in the sense of
dynamical systems due to vortex shedding. The falling spheres do not center rigorously but are in
some kind of unsteady off-center motion that is not well understood. How and why this kind of
unsteadiness leads to different settling velocities is not understood.
Table VII.2 lists the values of the velocity of sedimenting spheres of different diameter
falling in a 0.686cm gap between the walls of the sedimentation column. The sedimentation
velocity appears to increase rather markedly with mesh refinement; extrapolated to a fine mesh
h→0 would seem to imply a fall of about 10 cm/s rather than the 8.7 cm/s value observed in the
smaller mesh. We do not see a consistent variation with diameter. This may be due to some kind
of unsteadiness to which we alluded in the previous paragraph. There is a discrepancy between
the values in Table VII.2 and those reported in the experiments of between 10% and 20% if
8.7 cm/s is taken as the representative simulation value. We think that it is probable that the
faster fall velocity in the experiments may be due to channeling through places where the gap
size is larger.
Table VII.2. Averaged vertical terminal velocities of balls of different diameter and ρp = 1.14 g/cc
computed by DNS in the volume of Figure VII.6.
Diameter
0.6300cm 0.6350cm 0.6398cm 0.6500cm
h = 0.027" -8.085378 -8.130019 -8.155760 -7.790286
h = 0.018" -8.722692 -8.738451 -8.689694 -8.724728
The fluidization results are of great interest. We do not get a unique fluidization velocity; for
each sphere there is a rather large interval of velocities for which the sphere does not fall to the
bottom or blow out of the bed; the interval may range, say from 6 cm/s to 11 cm/s. In fluidization
velocities less than, say 6 cm/s the sphere will not rise and for large velocities, say about 11 cm/s
the sphere will blow out of the bed.
An important and practically useful result arising from this study of fluidization is that the
height of the sphere above the bottom increases with the fluidizing velocity. This positioning
property is such that the position of the particle in the bed may be controlled at the dial setting of
the inlet flow rate. The positioning property also arises form our simulation and is clearly evident
in the rise curves shown in Figure VII.16. We have seen such positioning hydrodynamics in
experiments on the fluidization of sensors in round pipes where the ability to position the sensor
is of practical importance.

T h e H e ig h t o f th e c e n te r o f th e b a ll (i
T h e h is to ry o f th e h e ig h t o f a flu id iz e d b a ll
3
I n f lo w s p e e d s ( c m /s e c ) a r e 6 , 6 .5 , 7 .5 , 9 , 9 .5 , 1 0 , 1 0 .5 ,
fo r th e fo llo w in g c u rv e s fro m th e b o tto m to th e to p re s
2 .5
1 .5
0 .5
0
0 0 .5 1 1 .5 2 2 .5 3 3 .5 4 4 .5
T im e (s e c )
Figure VII.16. H vs. t for a single particle d = 0.635cm from numerical simulation. For V < Vm ≈ 6 cm/s
the sphere remains at the distributor; for V > Vm ≈ 10.5 cm/s the sphere is dragged out of the bed.
The aforementioned positioning hydrodynamics is not understood. A promising explanation

follows from the observation that the average distance between the particle and the wall is a
function of the flow speed. Loosely described, we could say that the Segré-Siberberg position of
equilibrium between the centerline of close walls and the walls is a function of flow speed. The
overall drag on the sphere, which in any case must balance the buoyant weight, is a function of
flow speed and particle position. Evidently the flow speed produces a particle position such that
the drag-weight balance is preserved at different heights in the bed.
In the simulation an exactly zero velocity of the sphere was not achieved. For V ≤ 6 cm/s the
particle will not rise. The particle velocity in the interval of fluidization is very nearly zero and
rather random as shown in table VII.3. The particle rises out of the bed when V > 10.5.

Table VII.3. The averaged vertical velocities of a ball of diameter 0.635cm fluidized in a 2D-like bed of
dimension D × W × H = 0.27" × 7.992" × 8.1" are in the following table. The ball was initially
located at the center of the bottom of the bed.
In-flow velocity Averaged vertical speed

(cm/s) (after stabilized)
6 -0.00522
6.5 0.00827
7.5 0.00219
9 -0.00178
9.5 0.00631
10 0.0006997
10.5 0.00521
10.75 0.276
11 0.265
12 1.260
§ Richardson-Zaki correlations from DNS

Here we introduce an application of DNS that we call the method of correlations. The
method is inspired by the work of Richardson and Zaki 1954. They processed their data in log-
log plots and found straight lines leading to power laws. The method of correlations follows the
same procedure using numerical data from DNS rather than experimental data. Data from our
real and numerical experiments are shown in Figure VII.17. The experiment and simulations do
not quite match. The 1204 spheres used in the experiments are polydisperse (Figure VII.3) with
an average diameter of 0.6398cm rather than the 0.635cm diameter used in the simulation.
Moreover the lowest data point for the simulation may be inaccurate because the artificial
repulsive force which is activated to keep particles apart makes accurate calculations near close
packing at incipient fluidization less accurate.

40
35 Experiment
Simulation
30
Bed height (inches)

25
20
15
10
0
0 1 2 3 4 5 6 7 8
V (cm/s)
Figure VII.17. The bed height vs. fluidizing velocity for both experiment and simulation.
The Richardson-Zaki correlation relates the fluidization velocity V(φ) to the solid fraction
φ = 1 - ε where ε is the fluid fraction, in a factored form in which V(0), the blowout velocity for a
single sphere, is multiplied by a hundred settling function. When V > V(0) all the particles will be
dragged out of the fluidized bed. For the nondispersed case studied in simulation
Hs = 4.564/(1-ε) (VII.21)
The mean sphere size for the polydisperse case studied in the experiments is slightly larger and
He = 4.636/(1-ε) (VII.22)
The Richardson-Zaki correlations is given by
V(φ) = V(0)ε n(Re) (VII.23)
where V(0) is V when ε = 1,
æ dö
n = ç 4.65 + 19.5 ÷ when Re = V (0)d / v < 0.2,
è Dø
æ dö
n = ç 4.36 + 17.6 ÷ when 0.2 < Re < 1, (VII.24)
è Dø
n = 4.45 Re −1 when 1 < Re < 500,
n = 2.39 when 500 < Re < 7000
and D is the tube radius. The structure of the RZ correlation deserves consideration. The
exponent n does not depend on Re when Re is large or when Re lies between 500 and 700. In
these two regimes V(φ) /V(0) is a power law. For other values of Re, V(φ) /V(0) can be regarded
as a transition between power laws. This transition may be described empirically by a "logistic
dose curve" which is discussed in Appendix A of this monograph. In our experiments and
simulations Re is confined to the range for which n = 2.39.
The data shown in Figure VII.17 is plotted in a log-log plot in Figure VII.18 as V vs. ε. We
draw a straight line with slop n = 2.39 through both sets of data. The fit is not perfect but we
think rather encouraging. From the straight lines we determine the blow-out velocities Vs(0) =
8.131 cm/s and Ve(0) = 10.8 cm/s and find the power laws
Vs(φ) = 8.131ε 2.39 cm/s

and (VII.25)
Ve(φ) = 10.8ε 2.39 cm/s .
Experiment
Simulation
Fluidization velocity V (cm/s)
10
1
0.5 0.6 0.7 0.8 0.9 1
e
Figure VII.18. Data from Figure VII.17 plotted in a log-log plot. The slopes of the straight line are given
by the Richardson-Zaki n = 2.39. The blow-out velocities Vs(0) and Ve(0) are defined as the
intercepts at ε = 1.
Different reasons could be considered for the discrepancy between numerical and
experimental blow out velocities. The following argument suggests that the discrepancy is due to
the difference in the diameter 0.635cm of the sphere in the simulation and average diameter
0.6398cm of the 1204 spheres used in the experiments. This means that the walls will increase
the drag more in the experiments than in the simulations. To estimate the effect we may say the
wall correction formulas of Francis 1933, which was derived for Stokes settling of a sphere of
diameter d in a tube of diameter D
ρ s − ρ f 2 æ d ö 2.25
V ( 0) = d ç1 − ÷ (VII.26)
18η f è Dø
If we are allowed to consider D = 0.686cm, which is the distance between plane walls rather than
a tube diameter, then the ratio of velocities corresponding to the nominal d1 and the average
diameter d2 is given by

2 2.25
Vs (0) æ d1 ö æ D − d1 ö
2 2.25
æ 6.35 ö æ 51 ö
=ç ÷ çç ÷÷ =ç ÷ ç ÷ = 1.233 (VII.27)
Ve (0) çè d 2 ÷ø è D − d 2 ø è 6.398 ø è 46.2 ø
The value 1.233 is very close to the shift ratio
10.8
= 1.248 (VII.28)
8.131
necessary to bring the straight lines in Figure VII.18 together. After shifting by (VII.28) we
reverse the log-log plot to view the shifted plot in Figure VII.19.
35
30 Experiment
Simulation
25
Height (inch)
20
15
10
5
0 1 2 3 4 5 6
V (cm/s)
Figure VII.19. Bed height vs. fluidizing velocity after shifting by the ratio (VII.28) of blow-out velocities
obtained from the intercepts at ε = 1 in Figure VII.18.
§ Discussion
The fluidization of 1204 spheres at Reynolds numbers in the thousands was studied using the
method of distributed Lagrange multipliers. The results of the simulation are compared with a
real experiment designed to match. This is the first direct numerical simulation of a real fluidized
bed at the finite Reynolds number encountered in the applications. It is the first attempt to match
a real experiment to a fully resolved simulation. The numerical method used is presently very far
in advance of competitors for fully resolved CFD approaches to solid-liquid flow. The
experiments are carried out in a slit fluidization column in which the gap between close walls is
slightly larger than the fluidized spheres. The match between theory and experiment is very good
but not perfect; the spheres in the experiments are polydisperse with an average diameter of
0.6398cm whereas all the spheres in the simulation are 0.635cm. When these differences are
factored into the comparison the simulations and experiments are in good quantitative
agreement; the qualitative agreements are compelling. The emphasis of this paper is on the
interrogation of DNS for results in multiphase fluid mechanics and introduces four new
directions. First we have shown how our numerical method can be used to generate averaged
values of the solid fraction, the average of velocity components of both the solid and fluid and
possibly other averaged values used in multiphase models; for example we give the first ever
numerical simulation of the slip velocity, which might be used in drift flux models. In a second
application we obtain the contribution from wall friction that is usually neglected in one-
dimensional models of the dynamic pressure in a fluidized flow. We find that in our slit bed the
contribution of wall friction is about 1/4 of the total and that the contribution decreases modestly
as the bed expands. Our comparative study of sedimentation and fluidization of single spheres
revealed an unanticipated result that the balance between drag and buoyant weight can be
achieved in an interval of velocities. We framed this result as a positioning property; the particle
may be moved up and down the column by changing the fluidization velocity. The fluid
mechanics here are not understood; we conjectured that the drag can be maintained as the
velocity changes by a simultaneous change in the stand-off distance from the wall. The fourth
and most important application was framed as the method of correlations inspired by the way
that Richardson and Zaki 1954 processed their experimental results for power law correlations.
The idea is to plot results of experiments in log-log plots. Remarkably, straight lines emerge.
Apparently the flow of dispersions is governed at the foundation by similarity rules which are
not at all evident. We have the idea that we may implement this method using numerical rather
than real experiments and we think that our results establish this concept. In fact the results in the
experiments and simulations do follow the Richardson-Zaki correlation in a compelling if not
perfect match. We prefer to frame our result as a demonstration that we can generate power laws
by processing DNS data, rather than confirming correlations already obtained. We intend to
promote this approach strongly in the future.

VII Direct Numerical Simulation of Fluidization of 1204 Spheres.................................................41

§ Experiments .................................................................................................................................................. 41
§ Numerical Method ........................................................................................................................................ 41
§ Experiments .................................................................................................................................................. 44
§ Numerical Simulation ................................................................................................................................... 46
§ Qualitative comparison of experiment and simulation ................................................................................. 49
§ Numerical computation of averaged quantities ............................................................................................ 52
§ Dynamic and wall friction pressure in a fluidized bed ................................................................................. 55
§ Sedimentation and fluidization velocity of single spheres............................................................................ 57
§ Richardson-Zaki correlations from DNS ...................................................................................................... 60
§ Discussion..................................................................................................................................................... 63

Interog-4a.tex 65
VIII Modeling Rayleigh-Taylor Instability of

a Sedimenting Suspension of Several Thousand
Circular Particles in Direct Numerical Simulation
In this chapter we study the sedimentation of several thousand circular particles in 2D using the
method of distributed Lagrange multipliers for solid-liquid flow. The simulation gives rise to fin-
gering which resembles Rayleigh-Taylor instabilities. The waves have a well defined wavelength
and growth rate which can be modeled as a conventional Rayleigh-Taylor instability of heavy fluid
above light. The heavy fluid is modeled as a composite solid–liquid fluid with an effective com-
posite density and viscosity. Surface tension cannot enter this problem and the characteristic short
wave instability is regularized by the viscosity of the solid liquid dispersion. The dynamics of the
Rayleigh-Taylor instability are studied using viscous potential flow generalizing work of Joseph,
Belanger, and Beavers 1999 to a rectangular domain bounded by solid walls; an exact solution is
obtained.
The data in this chapter is generated by the direct numerical simulation of solid–liquid flow
using a distributed Lagrange multiplier/fictitious domain method (see Glowinski, Pan, Hesla &
Joseph 1999, Glowinski, Pan, Hesla, Joseph & Periaux 2000). The calculation is carried on fixed
triangular mesh on which fluid equations are satisfied everywhere. Rigid motions of the portions
of the fluid occupied by solids are accomplished by a strategic choice of a Lagrange multiplier
field there. The method has a certain elegance in that the rigid motion constraint on the fluid is
associated with a multiplier field in a manner analogous to the way in which the pressure in an
incompressible flow is a multiplier field associated with the constraint on incompressibility. The
details of the computation have been given in the cited references and will not be repeated here.
Simulation Data
The specific simulations discussed in this chapter concern the sedimentation of several thousand
disks settling down in a 2D rectangular box filled with water of density 1 = 1 g/cm3 and viscosity
1 = 0:01 poise. The disks of same diameter are initially arranged like those shown in Figure
VIII.2(a), Figure VIII.3(a) and Figure VIII.4(a). We also have shown in Figure VIII.1 the relative
position of disks in three different lattices mentioned above and call them square, hexagonal, and
rectangular respectively. In a square lattice the gap sizes in the horizontal and vertical directions
are the same. Similarly in a hexagonal case the gap size in the horizontal direction and the one
between rows in the vertical direction are the same. But in the rectangular case the gap size in the
horizontal direction is 1.2642 times of the one in the vertical direction.
The diameters of disks are 8=192 cm, 10=192 cm, 11=192 cm, 12=192 cm, 13=192 cm, 14=192
cm, 15=192 cm, and 16=192 cm. The density of disks is p = 1:1 g/cm3 . The volume fraction of
disks in the initial lattice is the ratio of the area Ap of the disks to the total area AT of the initial
lattice
Ap N d2 =4
= = (VIII.1)
AT HW
Interog-4a.tex 66
Square Hexagonal Rectangular
Figure VIII.1: Initial lattices.
where N is the total number of the disks, d is the diameter of the disks, H is the height of the initial
lattice, and W is the width of the box. In the simulation we have chosen 4, 6, 8, 10, and 12 cm as
the width W and the height of the box is always 12 cm.
In the cases where the initial lattice is square, there are 60 rows of disks in most of the cases. In
each row, there are 42, 63, 84, 105, and 126 disks as the width is 4, 6, 8, 10, and 12 cm, respectively.
To test whether more rows of disks can have different effect, we also tested 80 rows cases in a 2D
box of width 10 cm and height 12 cm with diameters varying from 10/192 cm to 16/192 cm.
In hexagonal cases there are 6270 disks staggered at the top of the box (see Figure VIII.3(a)).
There are 60 rows and in each row there are either 104 or 105 disks. The width and the height of
the box are 10 cm and 12 cm respectively. The diameter of disks varies from 10/192 cm to 16/192
cm. In Figure VIII.3(b), the snapshot of the sedimentation of 6270 disks of diameter 10/192 cm in
a 2D box is shown.
In rectangular cases, there are 80 columns at the top of the box. We tested two sets of cases in
which the number of rows is either 60 of 80 in order to probe the effect of the number of rows. The
diameter of disks varies from 10/192 cm to 16/192 cm. In Figure VIII.4, there are snapshots of the
sedimentation of 4800 and 6400 disks of diameter either 10/192 cm or 16/192 cm in a 2D box.
In all simulations the averaged particle Reynolds number at each time step is less than 3. The
maximal individual particle Reynolds number among all simulations is about 11. In each case,
simulation gives rise to fingering which resembles Rayleigh-Taylor instabilities (Figures VIII.2,
VIII.3, and VIII.4). The waves have a well defined wavelength and growth rate which we shall
model as a conventional Rayleigh-Taylor instability of heavy fluid above light. The arrangement
of sedimenting particles is asymmetric, flat on the top (most of portion on the top in the hexagonal
case) and corrugated at the bottom. The drag on a single disk is smaller than when it is among
many so that isolated disks on the bottom fall out of the lattice and isolated disks on the top fall
into the lattice.
Interog-4a.tex 67
(a). t = 0
(b). t = 0.476
(c). t = 0.476
Figure VIII.2: Snapshots of the sedimentation of 5040 (top, W = 8 cm) and 7560 (bottom, W = 12
cm) disks of diameter 14/192 cm in 2D. The initial lattice is square.
Interog-4a.tex 68
(a). t = 0 (b). t = 0.476
Figure VIII.3: Snapshots of the sedimentation of 6270 disks of diameter 10/192 cm in 2D (W = 10

cm). The initial lattice is hexagonal.
Two-fluid Model
We turn next to the two-fluid modeling of the instability of the sedimenting suspension just de-
scribed. The basic idea is to regard the particle laden portion of the sedimenting suspension (shown
in Figures VIII.2, VIII.3, and VIII.4) as an effective fluid with an effective viscosity 2 and an ef-
fective density 2 = (1 )1 + p and, of course, zero surface tension ; then we have two fluids,
an effective one above and water below. The dynamics of this two-fluid problem can be analyzed
using viscous potential flow (Joseph and Liao 1994). Joseph, Belanger and Beavers 1999 showed
that the wavelengths and the growth rates obtained with viscous potential flow differ from those
obtained from a fully viscous analysis by only a few percent. The success of the potential flow
analysis arises from the fact that main action of viscosity is in the viscous part of the normal stress
acting here in our problem through the effective viscosity of the solid–liquid suspension. Surface
p which
tension can not enter into this problem so that the effective viscosity is the only mechanism
regularizes an otherwise ill–posed problem in which the growth rate increases like 1= , tending
to infinity with ever shorter waves (Joseph and Saut 1990).
The analysis of Rayleigh-Taylor instability using viscous potential flow can be carried out in
an infinitely extended domain using the method of normal modes with disturbance proportional to
ent ei(kx x+ky y) eqz (VIII.2)
where, for viscous potential flow
q
q = k = kx2 + ky2 (VIII.3)
Interog-4a.tex 69
(a). t = 0
(b). t = 0.476
(c). t = 0.476 (d). t = 0.476
Figure VIII.4: Snapshots of the sedimentation of 4800 (top) and 6400 (bottom) disks in 2D (W = 8
cm). The diameter of disks in (a), (b) and (c) is 10/192 cm and the diameter of disks in (d) is 16/192
cm. The initial lattice is rectangular.
Interog-4a.tex 70
z
H2
W
O x
- H1
Figure VIII.5: Initial configuration.
where the z increases against gravity g = 980:6635 cm/sec and the sign k chosen so that the
amplitude decays at infinity. The analysis leads to the following dispersion relation (equation (25)
of Joseph, Belanger and Beavers 1999)
k k 3 2k 2
2 + 1 = (2 1 )g (2 + 1 ): (VIII.4)
n2 n2 n
Equation (VIII.4) depends on kx and ky only through k in (VIII.3); hence (VIII.4) is valid in both
two and three dimensions and it applies to the planar problem under discussion.
To get k which maximizes n (with zero surface tension, = 0), we differentiate (VIII.4) with
respect to k , set dn=dk = 0 and find that
(2 1 )g
k = : (VIII.5)
4n(2 + 1 )
Substituting (VIII.5) into (VIII.4), we get the growth rate

(2 1 )2 g 2
n3 = ; (VIII.6)
8(2 + 1 )(2 + 1 )
and the associated wave length is given by

(22 21 )g
k3 = : (VIII.7)
8(2 + 1 )2
We also carry out similar analysis in the rectangular domain of the computation as shown in
Figure VIII.5, in which we can construct the viscous potential flow. Let W be the width of the
domain, H2 is the height of the fluid-solid mixture above water of height H1 . Then the velocity
obtained from a potential is u = r . Let z = (x; t) be the interface. The normal stress balance
applied on z = 0 may be reduced, using Bernoulli’s equation, to
@ 2

j @@t j j @@z2 j
2
2
@x
= +[ jj] g + 2 (VIII.8)
Interog-4a.tex 71
The kinetic equation of motion of the perturbated free surface ! = @=@t implies that
@ @ 1 @ 2
= = (VIII.9)
@t @z @z
The normal derivative of the potential (x; z; t) must vanish on the solid wall, @ =@x = 0 on
x = 0 and x = W , @ 1 =@z = 0 on z = H1 (< 0), and @ 2 =@z = 0 on z = H2 (> 0). The
normal mode solutions corresponding to (VIII.8) on the bounded domain are
8
>
< = A ent cos kx cosh k(z + H ); for z < 0;
1 1 1
= A ent cos kx cosh k (z H ); for z > 0;

>
: = A ent cos kx
2 2 2 (VIII.10)
3
where
k = (m + 1)=W: (VIII.11)
It is convenient to treat k as a continuous variable.

After inserting (VIII.10) into (VIII.8) and (VIII.9) we find the dispersion relation
1 2
1 2

(2 1 )gk k n + 2nk :
3 2 2
= + +
tanh kH1 tanh kH2 tanh kH1 tanh kH2
(VIII.12)
The analysis of (VIII.12) proceeds along conventional lines, we find the k which maximizes n;
this k is such that kH1 and kH2 is never smaller than 29 in those simulation cases and two tanh’s
in (VIII.12) are almost one giving rise to (VIII.4). The comparison of computation and the model
may then proceed on the basis of (VIII.4).
Comparison of Two-fluid Model and Simulation
The “effective” viscosity 2 of the solid-liquid dispersion is unknown and may be defined by our
stability analysis using the following procedure. We first select the associated wavelength k0 =
2= where is the wavelength determined by the numerical experiment. From (VIII.7) we can
obtain the value of the effective viscosity 2 and then the associated growth rate by (VIII.6). The
determination of a growth rate from numerical simulation is carried out by fitting the growth in
the wave amplitude to bent . The wave amplitude is the distance of the wave crest defined by a line
through the centers of disks in the bottom row. The time step is 0.001 sec and the firest record
of distortion of the line of centers through the bottom row of disks is t = 0:026. Values of the
amplitude used in the curve fitting are at 0.025 second intervals from t = 0:026 to t = 0:476. In
Figure VIII.6, a set of amplitudes and the curve are shown for the case in Figure VIII.4(c).
The aforementioned procedure for determining the effective viscosity of sedimenting disks
from a stability calculation has been implemented for a large number of cases and the results
obtained are presented in Tables 1 through 12 and Figures VIII.9 through VIII.11. The data has
Interog-4a.tex 72
0.07
0.06
0.05
Amplitude (cm)
0.04
0.03
0.02
0.01
0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Time (sec)
Figure VIII.6: Amplitude vs. time for the simulation of instability of 6400 disks shown in Figure
VIII.4(c). The dashed line is obtained by curve fitting.
(a). t = 0.476 (b). t = 0.476
Figure VIII.7: Snapshots of the sedimentation of 9628 disks (left) and 11340 disks (right) of diam-
eter 8/192 cm in 2D (W = 8 cm). The initial lattice is square.
Interog-4a.tex 73
(a). t = 0 (b). t = 0.276
Figure VIII.8: Snapshots of the sedimentation of 5040 disks of diameter 10/192 cm in 2D (W =8

cm). The initial lattice is disordered.
been sorted by the geometry of the initial lattice of particles; square, hexagonal or rectangular
(Figure VIII.1). In each of these three categories we vary the width and height of the initial lattice
by adding or subtracting rows and columns of disks. Each table was indexed by the volume fraction
of disks; different volume fractions were created basically by changing the size of the particles.
The procedure we have adopted requires that we assess the success of the modeling by com-
paring the growth n(K0 ; 2 ) with growth rate from simulation. The largest discrepancy (Table
VIII.1) is of the order 30% but the error is less than 10% in the most of cases. In general the errors
are greater for small volume fractions and the largest errors occur at small volume fraction in the
narrowest (W = 4 cm in Table VIII.1).
The data for the square lattice in Table VIII.1 through 6 can be compared with the hexagonal
lattice in Table VIII.7 and in the rectangular lattice in Tables 8 through 10. The initial lattice of
disks is not modeled and it must have an effect. Obviously if most of the disks were in a ball,
it would be necessary at least to prescribe the distribution of volume fraction at the initial instant.
Results shown in Table VIII.4 (square) and Table VIII.7 (hexagonal) under otherwise identical con-
ditions do not reveal any important differences; the columns giving the effective viscosity for these
two cases match closely and can be seen graphically in Figure VIII.10. However, an examination
of the details of the instability in the region near the wall shows a dramatic difference between
the square (Figure VIII.2), the hexagonal (Figure VIII.3) and the rectangular (Figure VIII.4). The
staggered disks in the hexagonal case are such that only every other row of disks is near the wall, so
the wall region has a lower solid fraction and is “weaker” there in such a way the whole bed sinks
even as Rayleigh-Taylor waves develop. It is clear that the major effects of the wall are confined
Interog-4a.tex 74
to a wall layer.
When adding more rows to some existing cases, the results shown in Tables 4 and 6 (square)
and Tables 8 and 9 (rectangular) again do not reveal any important differences; the columns giving
the effective viscosity for these two cases match closely as can be seen graphically in Figures
VIII.9 and VIII.11.
Since rectangular lattices are clearly more anisotropic than square ones, much greater differ-
ences in the effective viscosity when calculated at the same volume fraction in simulations starting
from square and rectangular lattices of disks are evident in comparisons of Tables 3 and 8 and
summarized in Figure VIII.11.
To test whether different size particles with same volume fraction would give rise to the same
result, we used smaller disks of diameter 8/192 cm in simulations with square lattice. The results
are shown in Figure VIII.7 and Table VIII.11. Comparing the results in Table VIII.3 with same
dimension and about the same volume fraction and H2 , again we found that the effective viscosity
is not uniquely determined by the volume fraction.
We also considered a disordered initial lattice which was generated from a square lattice by
moving disks in the horizontal and vertical directions randomly within a given distance except
those disks in the bottom row which were only allowed to move in the horizontal direction. This
initial configuration is not really a random one. (To generate a random initial configuration with a
sharp and flat interface, we have to have many much smaller particles. This is beyond the capability
of our code for now.) The results are shown in Figure VIII.8 and Table VIII.12. We still can find
the development of waves in Figure VIII.8. We believe that the distribution of disks in these rows
right above the bottom row is a perturbation which has strong influence to the development of the
interface. The viscosity is consistent with those of square and hexagonal cases with wider width
(greater than or equal to 8 cm) and about the same solid volume fraction.
Discussion and Conclusions
The direct 2D simulation of the sedimentation of a close packed array of circular particles into
a rectangular box filled with water gives rise to fingers of particles with a wave structure which
resembles that which arise from Rayleigh-Taylor instability of heavy fluid above light. The wave
length and growth rate of falling particles can be compared with a two-fluid model of the Rayleigh-
Taylor instability using viscous potential flow. The particle-fluid mixture is modeled as a heavier
than water fluid with an effective density and viscosity. The effective density function of the solid
fraction is given by the formula () = p + f (1 ) which can be justified by ensemble and
other kinds of averaging. Interfacial tension can not enter at the nominal water-mixture interface.
The two-fluid model used here to study the Rayleigh-Taylor instability then is fixed when the
hindered settling function f (), which determines the effective viscosity () = w f () where w
is the viscosity of water, is known. Many formulas have been proposed (see, for example, Figure
3 in Poletto & Joseph 1995) and none are perfect.
The effective viscosity of a suspension is a way of describing the flow resistance due to internal
friction in a slurry. The resistance can depend on factors like wall proximity, particle size, particle
Interog-4a.tex 75
distribution and other factors even when the solid fraction is fixed. Clumped particles fall faster
than well mixed particles, particles near walls fall more slowly. The flow type is also a factor, the
effective viscosities of settling, shear and extensional flows are in general different even when the
volume fraction is fixed. It is necessary to think of an effective viscosity of a dispersion under well
specified conditions; one suit will not fit all.
In our study we focus our attention on the effective viscosity functions of the volume fraction
which gives rise to arrangements between computational experiments and two-fluid theory in a
restricted situation; we choose the viscosity function to obtain the same number of waves from
theory and numerical experiments. The theory may then be used to predict the growth rate and
this theoretical value can be checked against numerical experiments. Theoretical and experimental
values are listed in last two columns of Tables 1–12 and the agreements are satisfactory.
We have already argued that effective property models live only in well prescribed situations.
This is a negative for modeling because besides the model we must specify the situations in which
such model exists. In this chapter we determined viscosity function
(; W; H1 ; H2 ; d; I )
where different functions of are obtained when the bed geometry W , H1 and H2 (Figure VIII.5),
the particle size d and the initial lattice I (Figure VIII.1) are varied.
The asymptotic case of a semi-infinite bed in which W , H1 and H2 tend to infinity is of specific
interest since in this case there is no length to compare with the circle diameter d; the instability
must be independent of d but could depend on the initial lattice I of particles. This asymptotic case
might be the most universal, the one closest to our two-fluid model on the semi-infinite domain
leading to (VIII.4).
The periodic box of length W used in our exact solution (VIII.10) can be repeatedly extended
onto the infinity, but the ratio d=W then is a solution parameter. The asymptotic limit of a semi-
infinite domain mentioned above would be achieved when
d=W ! 0:
Unfortunately, we can’t compute when there are very many small particles d ! 0 at finite or
W ! 1 at finite . Our data shows that the observed wave length in the numerical simulation is
decreasing function of d=W when the initial arrangement of particles is fixed (see Tables 1-6 and
8-11). The data suggests that there is a limiting value, depending on the initial arrangement, also
for the two-fluid model as is indicated by the convergence of viscosity function exhibited in Figure
VIII.9.
Figure VIII.9 shows a very strong effect of the walls; this arises as a consequence of the varia-
tion of d=W for a fixed d and by a perhaps serious mismatch of theory and numerical experiment
which is amplified by reducing W . Nearby walls have a big effect when the no-slip condition at
the side-walls is enforced, as in the numerical simulation. The no-slip condition is not enforced
in the viscous potential flow theory; the retardation due to the walls is apparently realized in the
model by a higher value of effective viscosity. This value is larger when the solids fraction is small
because the shielding from the wall by other particles works less well when there are few than
Interog-4a.tex 76
when there are many particles. The viscous potential flow model, which apparently works well
as a fully viscous two-fluid model for Rayleigh-Taylor instability, may not be good approximation
when the walls are close. A fully viscous two-fluid model of Rayleigh-Taylor instability would
probably give better results.
The initial lattice of particles is an important parameter in our effective property model of
Rayleigh-Taylor instability. Rather large differences in the effective viscosity are demonstrated
between square and rectangular lattices are exhibited in Figure VIII.11. Figures VIII.4(b) and
VIII.4(c) exhibit an increase of local solid fraction near the interface. However, when we increased
the diameter of the disks from 10/192 cm to 16/192 cm, this increase of local solid fraction is
reduced dramatically (see Figure VIII.4(d)). An effective property model might be expected to
work best in a statistically homogeneous media. The square and hexagonal arrangements are
periodic in x and y with same period but the rectangular arrangement is doubly periodic.
The overall conclusion of this study is that effective two-fluid models can be made to work
in particulate flow but such theories require a prior prescription of the domain of arrangements of
particles to which the theory might apply. The greatest predictive value of such effective theories
is for statistically homogeneous media.
Table VIII.1. The dimension of the box is (W; L) = (4; 12) and H2 varies from 5.6541 to 5.6831.
The number of disks is 42 60 = 2520 (60 rows). The averaged particle Reynold number at the
final time step varies from 2.226 to 1.692. Initial lattice is square.
Wave length Effective

Disk Solid and associated viscosity 2 for n(K0 ; 2 ) Growth
diameter fraction wave number which n(K0 ; 2 ) rate from
K0 (in cm ) 1
is maximum simulation
10/192 23.74% 1.0, 2 0 0.14409 6.012 9.201
11/192 28.69% 0.8, 2.5 0.11135 7.380 10.237
12/192 34.10% 0.667, 3 0 0.09078 8.801 11.185
13/192 39.97% 0.615, 3.25 0.08690 9.904 12.716
14/192 46.29% 0.571, 3.5 0.08346 11.044 14.645
Interog-4a.tex 77

K0 (in cm 1 ) is maximum simulation
10/192 23.66% 0.600, 3.33 0.06149 7.748 9.941
11/192 28.60% 0.545, 3.67 0.05821 8.924 10.451
12/192 34.01% 0.500, 4.00 0.05537 10.150 11.402
13/192 39.88% 0.4615. 4.33 0.05287 11.423 12.565
14/192 46.21% 0.4615, 4.33 0.05778 12.278 14.512
15/192 53.00% 0.429. 4.67 0.05506 13.623 14.536
16/192 60.25% 0.4138. 4.83 0.05593 14.756 13.917
Interog-4a.tex 78

K0 (in cm ) 1
10/192 23.61% 0.500, 4.00 0.04433 8.480 10.056
11/192 28.56% 0.471, 4.25 0.04462 9.600 10.712
12/192 33.96% 0.444, 4.50 0.04474 10.758 11.377
13/192 39.83% 0.432, 4.625 0.04698 11.795 12.503
14/192 46.17% 0.421, 4.75 0.04903 12.849 14.377
15/192 52.96% 0.41025, 4.875 0.05091 13.919 13.849
16/192 60.22% 0.41025, 4.875 0.05507 14.816 15.039

10/192 23.59% 0.455, 4.4 0.03707 8.889 9.979
11/192 28.53% 0.435, 4.6 0.03848 9.983 10.544
12/192 33.93% 0.4167, 4.8 0.03967 11.107 11.313
13/192 39.81% 0.4167, 4.8 0.04388 12.012 12.283
14/192 46.14% 0.4000, 5.0 0.04464 13.179 13.795
15/192 52.94% 0.39215, 5.1 0.04691 14.233 14.366
16/192 60.20% 0.37736, 5.3 0.04739 15.446 14.081
Interog-4a.tex 79

K0 (in cm ) 1
10/192 23.57% 0.429, 4.667 0.03308 9.151 10.022
11/192 28.51% 0.414, 4.833 0.03500 10.230 10.570
12/192 33.92% 0.400, 5.000 0.03671 11.333 11.308
13/192 39.79% 0.400, 5.000 0.04066 12.257 12.308
14/192 46.12% 0.387, 5.167 0.04201 13.394 14.250

10/192 23.59% 0.47629, 4.2 0.04047 8.685 9.788
11/192 28.53% 0.45454, 4.4 0.04183 9.763 10.435
12/192 33.93% 0.43478, 4.6 0.04295 10.873 11.255
13/192 39.81% 0.41667, 4.8 0.04388 12.012 12.259
14/192 46.14% 0.40816, 5.0 0.04464 13.179 13.527
15/192 52.94% 0.38462, 5.2 0.04528 14.372 14.080
16/192 60.20% 0.37037, 5.4 0.04580 15.591 13.857
Interog-4a.tex 80
The number of disks is 105 60 = 6270. The averaged particle Reynold number at the final time
step varies from 2.259 to 2.089. Initial lattice is hexagonal.

K0 (in cm ) 1
10/192 23.48% 0.476, 4.20 0.04035 8.664 8.870
11/192 28.39% 0.455, 4.40 0.04170 9.740 9.257
12/192 33.77% 0.434, 4.60 0.04282 10.848 9.722
13/192 39.62% 0.417, 4.80 0.04374 11.984 10.513
14/192 45.92% 0.400, 5.00 0.04451 13.148 12.226
final time step varies from 2.591 to 1.552. Initial lattice is rectangular.

10/192 23.80% 1.14286, 1.75 0.17849 5.630 8.632
11/192 28.44% 1.0, 2.0 0.15883 6.572 9.723
12/192 33.42% 0.8889, 2.25 0.14358 7.548 10.679
13/192 38.74% 0.8421, 2.375 0.14268 8.338 13.034
14/192 44.39% 0.7619, 2.625 0.13084 9.370 13.508
15/192 50.35% 0.6956, 2.875 0.12105 10.428 15.038
16/192 56.61% 0.64, 3.125 0.11281 11.511 14.518
Interog-4a.tex 81
final time step varies from 2.593 to 1.523. Initial lattice is rectangular.

K0 (in cm ) 1
10/192 23.80% 1.14286, 1.75 0.17850 5.630 8.298
11/192 28.44% 1.0, 2.0 0.15883 6.572 9.591
12/192 33.42% 0.8889, 2.25 0.14358 7.548 10.553
13/192 38.74% 0.8421, 2.375 0.14268 8.338 12.785
14/192 44.39% 0.7619, 2.625 0.13084 9.370 13.136
15/192 50.35% 0.6956, 2.875 0.12105 10.428 14.253
16/192 56.61% 0.64, 3.125 0.11281 11.511 13.900
Table VIII.10. The dimension of the box is (W; L) = (10; 12) and H2 is 7.7109. The number of
disks is 100 80 = 8000 (80 rows). The averaged particle Reynold number at the final time step
is 0.857. Initial lattice is rectangular.

16/192 56.59% 0.6061, 3.3 0.10315 11.827 13.959
Table VIII.11. The dimension of the box is (W; L) = (8; 12) and H2 varies from 5.6988 to
5.7047. The numbers of disks are 116 83 = 9628 (83 rows) and 126 90 = 11340 (90 rows).
The averaged particle Reynold number at the final time step varies from 1.031 to 1.536. Initial
lattice is square.

K0 (in cm ) 1
8/192 28.77% 0.3478, 5.75 0.02484 11.207 11.914
8/192 33.92% 0.4, 5 0.03671 11.333 12.736
Interog-4a.tex 82
Cubic configuration
0.16
W=4
0.14
,h ) is maximum
0.12
0 2
h2 for which n(K
0.1
0.08
The effective viscosity
0.06 W=6
W=8
0.04 W=10(80)
W=10
W=12
0.02
0
0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6
Solid Fraction
Figure VIII.9: The effective viscosity 2 for which n(K0 ; 2 ) is maximum: the width of box is W
varying from 4 cm to 12 cm, and height of box is 12 cm, the initial lattice is square, the number of
rows is 60 (except the case W = 10(80) in which there are 80 rows), the number of column varies
from 42 to 126, and the diameter of disks varies from 10/192 cm to 16/192 cm.
Table VIII.12. The dimension of the box is (W; L) = (8; 12) and H2 is 5.6838. The number of
disks is 5040. The averaged particle Reynold number at the final time step is 3.022. Initial lattice
is disordered.

10/192 23.61% 0.4444, 4.5 0.03553 8.994 9.730
The wave length is an averaged quantity in this case due to the irregular shape of waves.
Interog-4a.tex 83
Comparison between cubic configuration (W=10) and hexagonal configuration (W=10(h))

0.16
0.14
The effective viscosity η for which n(K ,η ) is maximum
0.12
2
0
0.1
0.08
2
0.06
W=10(h)
0.04
W=10
0.02
0
0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6
Solid Fraction
Figure VIII.10: The comparison of the effective viscosity 2 (for which n(K0 ; 2 ) is maximum): the
width of box is W = 10 cm, the height of box is 12 cm, the initial lattice is either square (W = 10)
and hexagonal (W = 10(h) ), the number of rows is 60, the number of column is 105, and the
diameter of disks varies from 10/192 cm to 16/192 cm.
Interog-4a.tex 84
Comparison between cubic configuration (W=8) and rectangular configuration (W=8(r))
0.18 W=8(r)
The effective viscosity η for which n(K ,η ) is maximum
0.16
0.14
2
0
0.12
0.1
2
0.08
0.06
W=8
0.04
0.02
0
0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6
Solid Fraction
Figure VIII.11: The comparison of the effective viscosity 2 (for which n(K0 ; 2 ) is maximum): the
width of box is W = 8 cm, the height of box is 12 cm, the initial lattice is either square (W = 8)
and rectangular (W = 8(r)), the number of rows in square case is 60, those of rectangular case
are 60 and 80, the number of column is 80, the diameter of disks varies from 10/192 cm to 16/192
cm. Cases in which there 60 rows (resp., 80 rows) and 80 columns of disks are marked by r (resp.,
“x”). The “*” is a case in which there 100 rows and 80 columns of disks of diameter 16/192 cm.
Fluidization by lift: single particle studies
IX Fluidization by lift: single particle studies1

In Sections VI, VII and VIII we discussed fluidization and sedimentation in which the flow is
parallel to gravity and is governed by a balance of buoyant weight and drag. Now we turn to
flows perpendicular to gravity in which the particles are levitated against their buoyant weight by
lift forces perpendicular to the flow.
The theory of lift is one of the great achievements of aerodynamics. Airplanes take off, rise
to a certain height, and move forward under a balance of lift and weight. The lift and suspension
of particles in the flow of slurries is another application in which lift plays a central role; in the
oil industry we can consider the removal of drill cuttings in horizontal drill holes and sand
transport in fractured reservoirs. The theory of lift for these particle applications is
undernourished and in most simulators no lift forces are modeled.
Problems of fluidization by lift can be decomposed into two separate types of study: (1)
single particle studies in which the factors that govern lifting of a heavier-than-liquid particle off
a wall by a shear flow are identified and (2) many particle studies in which cooperative effects
on lift-off and hindered settling are important.
Our study of single particle lift starts by looking at a particle on the bottom wall of a plane
channel. The fluid in the channel is driven by a pressure gradient. We want a qualitative
description of the levitation of the particle as the pressure gradient is increased. The description
will be carried out in two dimensions here and then compared with computed results from DNS.
ΩP
Ω
Up He Up γ
d γ Uf Uf
(a) (b) (c) (d)
Figure IX.1 Lift off and levitation to equilibrium.
i. The pressure on a particle is increased
ii. The particle slides and rolls
iii. At a critical speed the particle lifts off
iv. It rises to a height in which the lift balances the weight

v. It moves forward under zero net force and torque; the particle does not accelerate (see
IX.9)
After the particle lifts off, it rises to a height he in which lift balances weight. Then the slip
velocity U f − U p and angular slip velocity Ω f − Ω p = γ / 2 − Ω p are positive and at equilibrium
1
This chapter is based in part on results reported in the paper by Patankar, Huang, Ko and Joseph 2001.
Last printed 03/11/02 11:43 AM 85 ♦ Interog-4B.doc

values as shown in figure IX.1(d). In many of the calculated cases the equilibrium is steady, but
for faster flows we have found periodic motions associated with Hopf bifurcation; surely there
are more complicated dynamics for turbulent regimes.
§ Equations of motion and dimensionless parameters

The problem to be considered is the levitation of a circular particle in a plane Poiseuille flow
in a horizontal channel. A particle of diameter d rests on the bottom wall at y = 0 of a horizontal
channel of width W (see figure IX.2). A flow is induced by a pressure gradient. The pressure
drop must balance the force due to the wall shear stress; γ = du dy is the shear rate, γ w is the
shear rate at the wall y = 0, all when there are no particles in the flow. We use
o
γ d2 η
R = w ,ν = (IX.1)
ν ρf
to index different flows. The equations of motion (II.6, 7, 8) are modified to include the effects
of an applied pressure gradient P given by (III.2).
g
Channel Width W
Poiseuille
Flow
y Particle of Diameter d
x
Figure IX.2. The particle is heavier than the fluid and lifts off the bottom y = 0 as the pressure gradient,
indexed by γ w , is raised past a critical value, lift-off value.
There are features of the problem of levitation of circular particles in a horizontal plane
Poiseuille flow which require explanation. The equations of motion for fluid and particle are
given in general terms by (II.1) through (II.5). In the present case it is necessary to explain how
the pressure is decomposed for computation on a periodic domain. It is understood that the
velocity u is solenoidal, div u = 0 and we have changed p to P where P is the mean normal
stress. We have
é ∂u ù
ρ f ê + (u ⋅ ∇)u ú = −∇P + η∇ 2 u + ρ f g (IX.2)
ë ∂t û
where u satisfies no-slip conditions at solid boundaries. The equations of motion of the solid
particles are given by

dU
ρ pV p = ρ pV p g + ò {− P1 + 2ηD[u]} ⋅ n dΓ
dt
(IX.3)
dΩ
dt ò
Ι = (x − X) ∧ [(− P1 + 2ηD[u]) ⋅ n ] dΓ
where u is the velocity of the mass center and Ω is the angular velocity.
The Poiseuille flow problem studied in the computation is an idealization of a channel flow
problem in which pressure is prescribed at the inlet x1 and outlet x2 , x2 >> x1 where P1 > P2.
The prescribed inlet-outlet conditions are modeled by a prescribed pressure gradient
P1 − P2
p= . (IX.4)
x2 − x1
We decompose the pressure into a periodic part p ( x, y, t ) , hydrostatic part ρ f g ⋅ x and the
constant pressure gradient part − pe x ⋅ x = − px , thus
P = p + ρ f g ⋅ x − pe x ⋅ x . (IX.5)
Combining (IX.5) with (IX.2) and (IX.3) we get
é ∂u ù
ρ f ê + (u ⋅ ∇)u ú = −∇p + η∇ 2 u + pe x (IX.6)
ë ∂t û
dU
ρ pV p = ( ρ p − ρ f )V p g + V p pe x + ò {− p1 + 2ηD[u]}⋅ n dΓ (IX.7)
dt
dΩ
= (x − X) ∧ [(− p1 + 2ηD[u]) ⋅ n] dΓ
dt ò
Ι (IX.8)
where V p = π a 2 is the volume per unit length of the circles and I = ρ pπa 4 / 2 .
The functions u ( x, y , t ), p ( x, y , t ) , U ( X , Y , t ) and Ω ( X,Y,t ) are periodic function of x and X

with period L. For steady flow dΩ Ω/dt = 0 and dU/dt = 0 for which
( ρ p − ρ f )V p g + V p pe x + ò {− p1 + 2ηD[u]}⋅ n dΓ = O . (IX.9)
The vertical component of (IX.9) represents the balance of lift and buoyant weight. The e x
component of (IX.9) balance the forward pressure gradient thrust against the e x resultant of the
periodic stress.
The diameter d is introduced as the scale of length, V is the scale of the velocity, d/V is the
scale for time, ηV/d is the scale of stress and pressure and V/d is the scale for angular velocity.
After introducing these scales into (IX.6~8), we find that

ü
∂u 2d ï
R[ + (u ⋅ ∇)u] = −∇p + e + ∇ 2u
∂t W x ï
ρ p V p dU Vp ï
Ld V p ï
d y W d3 x ò
R 3 = −G 3 e + e + {− p1 + 2ηD[u]} ⋅ ndΓ ý (IX.10)
ρ f d dt ï
ρ p I R dΩ ï
ï
ρ f ρ p d 5 dt ò
= (x − X) ∧ (− p1 + 2ηD[u]) ⋅ n dΓ .
ïþ
Equation (IX.10) holds in two dimensions. The particle equations of motion can be formed
for cylinders of length d; then Vp = d•π d2/4 and Ip = d•ρpπ d4/32 and
ρ p dU d 4
2π
dθ
R = −Ge y + 2 e x + ò {− p1 + 2ηD[u]} ⋅ n (IX.11)
ρ f dt W π 0 2
ρ p dΩ 32 2π dθ
R =
ρ f dt π 0ò (x − X) ∧ {− p1 + 2ηD[u]} ⋅ n
2
. (IX.12)
The flow is determined by four dimensionless groups
d (ρ p − ρ f )g
o
ρ p 2d ρf γ w d2
, ,R= ,G= . (IX.13)
ρf W η o
γ η w
For fixed ρp /ρf and d/W, all flows with the same R and G are dynamically similar. The density
ratio ρp /ρf appears as an independent parameter only as a coefficient of particle acceleration in
equations (IX.11) and (IX.12). This parameter does not enter into steady motion. Heavy particles
accelerate more slowly than light ones.
The gravity Reynolds number
ρ f (ρ p − ρ f )gd 3
RG = RG = (IX.14)
η2
is independent of γ w ; it is based on the sedimentation velocity Used = (ρp - ρf)gd2/18η of a sphere

in Stokes flow and measures the ratio of buoyant weight to viscous effects. The particle will fall
rapidly when RG is large. The ratio
o 2
R dγ w
= (IX.15)
G æ ρp ö
ç − 1÷ g
çρ ÷
è f ø
is independent of η; this ratio, which measures the ratio of lift to buoyant weight, is a generalized
Froude number; particles will rise more when R/G is large.

All the calculations have been carried in dimensional variables using CGS units. The height
of the channel W is 12d (unless otherwise specified) and d = 1 cm. The length of the periodic
channel L is 22d for single particle simulations and 63d for simulations of the fluidization of 300
particles. Computed quantities become independent of L for large L, and 22d and 63d are large
enough. The results of calculations in dimensional variables may be generalized by post-
processing in the dimensionless parameters defined above.
The undisturbed Poiseuille flow is given by
yæ yö du
U ( y ) = 4U m ç1 − ÷, γ = ,
Wè Wø dy
(IX.16)
æW ö W p
2
4U m W p
Um = Uç ÷ = , γ w = = .
è 2ø 8η W 2η
The calculation was carried in two channels: W/d = 12 and W/d = 48. The period of the
periodic solution in the channel W/d = 12 is L/d = 22; for W/d = 48, L/d = 88.
§ Lift-off of a single particle in plane Poiseuille flows of a Newtonian fluid
In figure IX.3 we plot the trajectory of the circular particle as a function of the distance
traveled along the axial direction. The channel dimensions are: W/d = 12 and L/d = 22, where d =
1 cm (figure IX.2). The fluid density and viscosity are 1 g/cc and 1 poise, respectively. The
particle density is 1.01 g/cc and RG = 9.81. The center of the particle is initially at y = 0.6d.
When R < 2.83 the particle falls to the bottom wall. For R > 2.83 it falls or rises to an equilibrium
height he at which the buoyant weight balances the hydrodynamic lift. Thus the critical value of
R for lift-off in this case is 2.83. The equilibrium height increases with R (figure IX.3).

0.9
0.85
0.8 R = 6.67
0.75
h e = 0.78cm
h (cm)
0.7
0.65
R =5
0.6
h e = 0.608cm
0.55 R = 4.17
R = 2.83, he = 0.5012cm h e = 0.543cm
0.5
0 10 20 30 40 50 60 70
x (cm)
Figure IX.3. Cross stream migration of a single particle (ρp > ρf ). A single particle of diameter d = 1 cm
is released at a height of 0.6d in a Poiseuille flow. It migrates to an equilibrium height he.
When the motion of the particle is steady, the acceleration terms in equations (IX.11, 12)
vanish, and the density ratio ρp/ρf can be eliminated from the list (IX.13) of controlling
parameters. We compute the critical shear Reynolds number for lift off from a steady flow. In the
simulations, the smallest allowable gap size is set at 0.005d. The gap between the particle and the
wall can never be zero (Hu and N. Patankar 2001). The smallest shear Reynolds number at which
we observe an equilibrium height greater than 0.5005d is therefore identified as the critical shear
Reynolds number for lift-off in our dynamic simulations. In most cases the smallest equilibrium
height we obtain is around 0.501d.
The non-dimensional equilibrium height, he/d, is a function of R, RG and W/d. Figure IX.4a
shows the plot of he/d as a function of R at different values of RG with W/d = 12. The equilibrium
height increases as the shear Reynolds number is increased at all values of RG. A larger shear
Reynolds number is required to lift a heavier particle to a given equilibrium height. Figure IX.4b
compares the equilibrium height of a particle of given density in channels of different widths
(W/d = 12, 24 and 48). L/d = 44 for a channel with W/d = 24 whereas L/d = 88 for W/d = 48. At a
given shear Reynolds number the dimensionless equilibrium height is larger for the bigger
channel.

1
(a) W/d = 12
R G= 9.81 (rp = 1.01 g/cc)
0.9 R G= 392.4(rp = 1.4 g/cc)
R G= 0.981(rp = 1.001g/cc)
0.8 Lift-off
he (cm)
0.7
Lift-off
0.6 Lift-off
0.5
0 10 20 30 40 50 60
R
(b) 0.8
0.75 RG = 9.81 (r p = 1.01g/cc)
0.7 W/d = 12
W/d = 24
he (cm)
0.65 W/d = 48
0.6
0.55
Lift 0ff
0.5
1 2 3 4 5 6 7
R
Figure IX.4 (a) Lift-off and equilibrium height as a function of the shear Reynolds number at different
values of RG. (b) Equilibrium height vs. shear Reynolds number at different channel widths.

10000
RG = aRn
W/d = 12 & 48 : a = 2.3648, n = 1.3904
1000 W/d = 6 : a = 1.8120, n = 1.3566
W/d = 4 : a = 1.2596, n = 1.3644
100
W/d = 48
RG
W/d = 12
10
W/d = 6
1 W/d = 4
Data from dynamic

simulations (W/d = 12)
0.1
0.1 1 10 100 1000
R
Figure IX.5. The plot of RG vs. the critical shear Reynolds number R for lift-off on a logarithmic scale
at different values of W/d.
We did many simulations of lift-off from the height 0.501d, varying R and W/d, and post
processed the results in terms of R and RG (see figure IX.5). The particle moves only under
forces of the fluid motion and gravity; no axial force or torque is applied. Figure IX.5 shows the
plot of RG vs R at different values of W/d; a larger R is required to elevate heavier particle. The
channel width apparently does not influence the correlation when W/d > 12 (figures IX.4 and
IX.5). The data for W/d > 12 collapse onto the power law curve
RG ≈ 2.36 R1.39 (IX.17)

This may be the first correlation obtained from numerical experiments, based on DNS and it
points to one future direction for interrogation of DNS.
Up and Ωp are the translational and angular velocities, respectively, of the particle in
equilibrium when the particle accelerations vanish. In figure IX.6 we plot the results of dynamic
simulations, the slip velocity, Uf - Up vs. equilibrium height for particles of different densities. A
similar plot for the slip angular velocity, γ 2 − Ω p , is shown in figure IX.7. A larger slip
velocity is required at a given equilibrium height to balance a heavier particle.

10
W/d = 12
8
R G = 0.981
Uf -Up (cm/s) R G = 9.81

6
R G = 392.4
Lift off
0
0.5 0.6 0.7 0.8 0.9
he (cm)
Figure IX.6. Slip velocity vs. equilibrium height for different particle densities.
12
10 W/d = 12
R G = 0.981
8 R G = 9.81
Wf - Wp (/s)
R G = 392.4
Lift off
6
0
0.5 0.6 0.7 0.8 0.9
he (cm)
Figure IX.7. Slip angular velocity vs. equilibrium height for different particle densities.
In table IX.1 we have listed all the computed values of RG, R/G, R, p , he, Up, Uf , Us = Uf –
Up, Ωp, Ωf = γ 2 and Ωs = γ 2 -Ωp at equilibrium, where Uf and γ are as shown in figure IX.1.

The aforementioned quantities define a data structure generated by DNS which can be used to
help in the creation and validation of lift models; the tables give the answers to which the models
aspire.
Table IX.1. Data structure for a freely translating and rotating circular particle levitated by
Poiseuille flow (d = 1 cm, ρf = 1 g/cc and η = 1 poise). Bold numbers represent the critical
condition for lift-off. All the dimensional variables are given in CGS units. (a) W/d = 12, L/d =
22, (b) W/d = 24, L/d = 44, (c) W/d = 48, L/d = 88.
Table IX.1a
R/G RG R p he Up Uf Us Ωp Ωf Ωs
0.1133 0.981 0.3333 0.0555 0.5024 0.0170 0.1605 0.1436 0.0161 0.1527 0.1367
0.7079 0.981 0.8333 0.1389 0.5081 0.0820 0.4055 0.3235 0.0752 0.3814 0.3062
2.8316 0.981 1.6667 0.2778 0.9055 1.2310 1.3953 0.1643 0.6085 0.7076 0.0991
0.8183 9.81 2.8333 0.4722 0.5012 0.1337 1.3608 1.2271 0.1147 1.2983 1.1836
1.1326 9.81 3.3333 0.5556 0.5058 0.3479 1.6149 1.2670 0.2934 1.5262 1.2328
1.7697 9.81 4.1667 0.6944 0.5433 1.1230 2.1613 1.0383 0.8868 1.8947 1.0079
2.2200 9.81 4.6667 0.7778 0.5786 1.6560 2.5699 0.9139 1.2220 2.1083 0.8863
2.5484 9.81 5.0000 0.8333 0.6083 2.0590 2.8873 0.8283 1.4430 2.2465 0.8035
4.5305 9.81 6.6667 1.1111 0.7784 4.3350 4.8527 0.5177 2.3340 2.9009 0.5669
1.5928 392.4 25.000 4.1667 0.5009 2.2820 11.999 9.7178 1.2790 11.456 10.177
2.8316 392.4 33.333 5.5556 0.5074 7.7790 16.198 8.4192 4.2600 15.257 10.997
6.3710 392.4 50.000 8.3333 0.5475 21.730 26.126 4.3960 11.500 22.719 11.219
Table IX.1b
0.8183 9.81 2.8333 0.2361 0.5015 0.1611 1.3912 1.2301 0.1381 1.3575 1.2194
1.6310 9.81 4.0000 0.3333 0.5485 1.2020 2.1439 0.9419 0.9468 1.9086 0.9618
1.7697 9.81 4.1667 0.3472 0.5619 1.3990 2.2864 0.8874 1.0750 1.9858 0.9108
1.9141 9.81 4.3333 0.3611 0.5766 1.6030 2.4386 0.8356 1.2010 2.0626 0.8616
Table IX.1c
0.8183 9.81 2.8333 0.1181 0.5034 0.2491 1.4113 1.1622 0.2134 1.3870 1.1736
1.4979 9.81 3.8333 0.1597 0.5467 1.1640 2.0718 0.9078 0.9247 1.8730 0.9483
1.6310 9.81 4.0000 0.1667 0.5600 1.3590 2.2139 0.8549 1.0540 1.9533 0.8993
1.9141 9.81 4.3333 0.1806 0.5901 1.7750 2.5257 0.7507 1.3070 2.1134 0.8064

§ Data Structure for DNS and experiments

We have identified two distinguished regimes: lift-off and equilibrium. For lift-off we would
like to know
i. The critical condition for lift-off (the pressure gradient, Reynolds or other prescribed
control parameter)
ii. The particle velocity Up at criticality
iii. The particle angular velocity at criticality
iv. The velocity and shear rate at the particle center when there is no particle present
v. The slip velocity and slip angular velocity at criticality
At equilibrium we have a similar list. At equilibrium we wish to know
he (height of mass center at equilibrium)
Up at he
Uf at he
Ωp at he
γ at he
2
Uf - Up at he
γ + Ω at h
p e
2
The aforementioned quantities define a data structure for DNS, which may be used to
develop a theory of fluidization by lift. Tables IX.1 and IX.2 show how to structure data from
DNS to test certain modeling assumptions. Plots of the rise evolution of neutrally buoyant
particles are given in figures IX.8 and IX.9. The neutrally buoyant particles rise rapidly, reaching
an equilibrium value not determined by the buoyant weight, in this case the equilibrium height he
is the place where the lift vanishes. The generation of zero lift for a neutrally buoyant particle is
associated with the curvature of the velocity profile in Poiseuille flow. The freely rotating
particle with R = 5.4 rises to a distance of 2.248 cm from the center of the channel as compared
to 3.6 cm from the center which is the “Segré-Silberberg” radius (0.6 w/2). The nonrotating
particle rises further to 4.999 cm. The heavy particle ρ p = 1.01 also rises further when the free
rotation is suppressed and R = 16.2, but the rise of freely rotating and nonrotating heavy particles
is not greatly different when R = 5.4. It may be true that models, which ignore particle rotation
will overestimate lift.

5
Ws= g°/2+Wp W s =g° /2
4 Freely rotating 0 < W se< g /2
W s= 0
Y = y/d 3
Neutrally buoyant: rp /rf =1.00

2
1 rp /rf =1.01
0
0 50 100 150 200
Time (second)
Figure IX.8 Rise vs. time for Rw = 5.4. Compare rise of freely rotating and nonrotating particles.
Nonrotating ones rise more. A neutrally buoyant, freely rotating particle rises closer to the center
line than the “Segré-Silberberg” experiment; the nonrotating one rises even more. Models which
ignore particle rotation overestimate lift. A yet smaller lift is obtained when the slip velocity is
entirely suppressed (Ωs = 0), but the particle does rise. The greater the slip velocity, the higher the
particle will rise.
4
W s=g° /2+W p
3.5
Wse=g /2
3
2.5
Y = y/d
2
Freely rotating
1.5 0 < W se< g /2
1 W s= 0
0.5
0 10 20 30 40 50 60 70 80
Time (second)
Figure IX.9 Rise vs. time for a circular particle ρp /ρf = 1.01 when Rw = γ w d 2 / η = 16.2 . The rise of
the particle is an increasing function of the slip velocity in the range 0 ≤ Ω s ≤ γ / 2 and is a
maximum when the particle velocity is suppressed. The freely rotating particle has a small positive
slip angular velocity.

There are a small number of experiments on the inertial lift of single particles. The data
obtained falls far short of that necessary for testing ideas about lift-off and levitation to
equilibrium. The list of quantities mentioned at the head of this section as well as data on
stability and bifurcation are not adequately probed in these experiments; more penetrating
experiments should be done.
Table IX.2 Data structure for a circular particle levitated by Poiseuille flow (W/d = 12, L/d = 22, d
= 1cm, η = 1 poise). The data for steady flow after the particles rises to its equilibrium height. Three
cases are considered: Ωp = 0, freely rotating (*) and Ωs = 0.
he Ωp Up γ 2 uf = γ he Ωslip Uslip
ρp /ρf Re
(cm) (sec-1) (cm/sec) -1
(sec ) (cm/sec) (sec-1) (cm/sec)
1.00 5.4 4.999 0.0 15.670 0.450 15.749 0.450 0.080
*1.00 5.4 3.753 0.958 13.780 1.011 13.928 0.053 0.148
1.01 5.4 0.681 1.044 13.630 1.044 13.780 0.0 0.150
1.01 5.4 0.602 0.0 2.125 2.429 3.088 2.429 0.963
*1.01 5.4 0.627 1.560 2.296 2.418 3.208 0.858 0.912
1.01 5.4 0.651 2.407 2.453 2.407 3.325 0.0 0.872
1.01 16.2 3.620 0.0 40.260 3.213 40.953 3.213 0.693

*1.01 16.2 1.211 5.320 17.410 6.465 17.638 1.145 0.228
1.01 16.2 1.199 6.481 17.190 6.481 17.483 0.0 0.293
Eichhorn and Small 1964 suspended a small sphere in Poiseuille flow through an incline tube
and determined the lift and drag coefficients on the particle. By suspending the sphere in the
flow, the lift and drag forces are calculated from the tangential and normal components of the
buoyancy force with the flow. The results are not extensive, but at the time did provide new
information about the drag and lift on spheres in such flows. These experiments showed that the
lift coefficient increases with Reynolds number. However, the accuracy of the measurements was
limited and the variables such as rotation speed and radial position are related by the operating
characteristics of the apparatus and could not be varied independently.
Bagnold 1974 measured the lift and drag forces on spheres and cylinders in the gravity flow
of a liquid in an open channel (the upper, free boundary of the liquid is frictionless). The objects
are placed near the lower, solid boundary of the device and are allowed to translate down the
channel in the shear flow. Bagnold overcame the problem of the limited channel length by
creating a stationary "flow" where the lower boundary is replaced with an endless belt that
translates in the direction opposite the free stream velocity. By producing the proper flow
kinematics to balance the inertial motion of the sphere down the channel, the particle can be
suspended in the flow and the equilibrium height can be measured. In separate measurements, a
linkage assembly is used to measure the drag and lift forces on bodies fixed in the flow field. In
general, Bagnold observed that the lift force decreases rapidly with increasing distance from the
solid boundary and disappears when the clearance exceeds on particle diameter. Unfortunately,
the author admits, "the experiments must be regarded as exploratory only'' because of the
accuracy of the device and the limited scope of the experiments.

Cherukat, McLaughlin and Graham 1994 used a homogeneous shear flow apparatus (HFA)
to measure the shear-induced inertial lift on a rigid sphere. The HFA device creates a uniform
linear shear flow between two timing belts moving in opposite directions. Spheres were injected
onto the mid-plane between the two belts will translate between the belts, migrating laterally
towards one of the walls. A system of cameras recorded the motion of the particle that was then
used to calculated the dimensionless lift force on the sphere. The lift force increases with the
ratio of ε = Re / Re s , where Re is the shear Reynolds number of the flow and Res is the
Reynolds number of the sphere based on the slip velocity. These results, within experimental
error, validated the theory of McLaughlin 1991 and they showed that Saffman's expression
overpredicts the lift if ε >> 1 . However, these experiments are not able to measure the
equilibrium height from the wall, and the rotation velocity of the sphere is not presented.
Experiments by Mollinger and Nieuwstadt 1996 examined the lift forces on particles within
the viscous sublayer of highly turbulent flows (Re ~ 106). Their experimental device measures
the lift force on a small particle permanently affixed to cantilever beam at the surface over which
air flows in a wind tunnel. Optical methods allow for measuring the lift force to an accuracy of
10-9 N. Their results show a substantial difference between experiments and theory for this
regime of Reynolds numbers. However, because the particle is fixed onto the surface,
measurements of important parameters such as equilibrium height, slip velocity, and rotation rate
are unobtainable.
Ye and Roco 1992 measured the velocity and angular velocity for neutrally buoyant spheres
with diameters from 3 to 8 cm in a plane turbulent Couette flow at three shear rates
corresponding to Reynolds number (4.6, 6.8, 9.2) ×104. Fluid and particles were assumed to have
the same velocity; slip velocities were not measured. Particle spin increases rapidly near the
walls. No significant differences between the angular velocity of 6.36 and 7.94mm spheres were
found, but smaller spheres of 4.76 and 3mm rotate faster.
More recently, King and Leighton 1997 used a rotating parallel-plate device to measure
translation velocity and the angular velocity of particles in contact with a wall. The velocities
were determined by timing the particle as it travels through an arc length. A wide range of
Reynolds numbers could be sampled because the local shear rate between the parallel plates is
proportional to the radial distance. Unfortunately, secondary currents drive the particle inwards
at larger Reynolds numbers. This study showed that the particle undergoes three different modes
of motion as the flow Reynolds number is increased: (i) solid body rotation along the wall, (ii)
particle rotation with translational slip along the wall, and (iii) translation and rotation in the free
stream following lift-off. Careful attention was paid to the surface roughness of the sphere and
the frictional coefficient of the wall that plays an important role in the motion when the sphere is
in contact with the wall. The transition from (i) to (ii) was observed to scale with Re/Res ~ 0.1;
while lift-off occurred when Re2/Res ~ 4. These experiments were in good agreement with the
rough sphere model proposed by Krishnan and Leighton 1995 for a range of reasonable surface
roughness values. No values of the equilibrium height were presented.
It is more difficult to specify a data structure for interrogating DNS for the modeling of lift
on particles in a flowing suspension in pipes and conduits in which hydrodynamic interactions
are important. At present there are no satisfactory models for this problem. Perhaps an approach
to this modeling can be developed along the lines followed by Richardson and Zaki. In their case
it was necessary to find an empirical formula for the slip velocity in a suspension in the factored

form (VI.60), the product of the terminal or blow-out velocity for a single particle with the
hindered settling function. We have shown that such an empirical relation can be formed from
data from DNS. All the factors in (VI.60) are empirical; analytical formulas for the terminal
velocity are known only for low Reynolds number flow, Stokes and Oseen flows. It is almost
certain that role of first principle formulas from analysis is even more limited in the problem of
lift.
The Richardson-Zaki experience teaches also that the existence of a good correlation like
(VI.16) does not automatically lead to a drag formula; creative intelligence is still required.
To arrive at an appropriate data structure for lift in suspension it is necessary to identify what
quantifies ought to be modeled and how these quantities enter into models for lift. For example,
we can consider that idea that lift formulas ought to depend on the slip and angular slip velocities
and seek these in a factored form, products of slip and angular slip velocity for a single particle
with to-be-determined hindered settling functions. To see what may be involved in an approach
like this it is instructive to search for ideas in the very restricted situations for which
mathematical analysis is possible.

Analytical models of lift
X Analytical models of lift

We have already noted that the domain of parameters for which mathematical analysis is
possible is rather severely restricted. As in the case of fluidization by drag, the analysis is mainly
restricted to Stokes and Oseen flow Reynolds number drag and rather overly complicated
analysis of perturbation of these with inertia. The main problem with the perturbations is that the
Stokes flow is not uniformly valid, for away from the body inertia and viscosity are both
important and rather complicated analysis involving the matched asymptotic expansions must be
involved. The low Reynolds number far lift from a wall is slightly different, because if the
particle is heavy and not lifted too far away from wall, regular rather than singular perturbations
can be used. Another difference between lift and drag is that you get fine lift formulas in 2D with
viscosity completely neglected as in aerodynamic theory; no such formulas can be obtained for
drag.
Our goal here is to explore some ways DNS can be used to generate engineering models of
lift which almost always have an empirical input, as does the Richardson-Zaki theory. It is to be
hoped that DNS is a source for the empirical input. To achieve our goal we must first come to
realize what quantities will enter into expressions for lift in a slurry. For a single particle, say a
sphere, we would like to understand the role of fluid shear and shear gradients (curvature),
particle velocity and angular velocity, wall effects and fluid material parameters. The brief
discussion of analytical results to follow focuses on how these quantities enter into lift in the
rather special cases in which lift formulas can be written down. These formulas are compositions
of the quantities listed in tables IX.1 and IX.2 and in similar to-be-created tables for levitating
spheres, and they can be evaluated and generalized by DNS.
The lift on a particle moving forward is a force opposing gravity perpendicular to the motion.
More generally we may identify the lift as the component of a force transverse to the motion in
the direction of gravity.
§ Lift in an inviscid fluid

First we can look on formulas in a fluid without viscosity for which viscous drag is
impossible. The most famous formula for lift on a body of arbitrary shape moving forward with
velocity U in a potential flow with circulation Γ was given by
L′ = ρUΓ (X.1)
where ρ is the fluid velocity and L′ is the lift per unit length.
The lift on circular cylinder of radius a is of special interest. A viscous potential flow
solution for a stationary cylinder rotating with velocity Ω which satisfies the no slip boundary is
given by
U ( r ) = e θ Ω a 2 / r , u( a ) = e θ Ω a (X.2)
The circulation for this viscous potential flow is
Γ = ò u • dx = 2π Ωa 2 (X.3)

When this rotating cylinder moves forward it generates a lift L′ per unit length
L′ = 2πρa 2UΩ (X.4)
The direction of the lift can be determined by noting that the velocity due to rotation adds to the
forward motion of the cylinder at the top or bottom of the rotating cylinder according to the
directions of Ω and U. In figure X.1 the velocity is smaller on the bottom of the cylinder; by the
Bernoulli equation, the pressure is greater there and it pushes the cylinder up.
Low
Pressure
U Ω
High
Pressure
Figure X.1 The lift per unit length L′ = 2πρa2UΩ on a cylinder of radius moving forward at speed U
and rotating with velocity Ω in such a way as to reduce the velocity at the bottom and add at the top.
The effect of proximity to the ground is enhanced lift. An aerodynamic formula for this can
be derived by the method of reflections, representing the airfoil by a point vortex (see, e.g.
Kuethe and Chow 1998).
é c Γ æ c 2 öù
L′ = ρUΓ ê1 + − ç 1 + ÷ú (X.5)
ë 16h
2
4hU çè 16h 2 ÷ø û
where c is the chord and h the distance of the vortex from. We can write a ground effect formula
for the lifting of a cylinder by putting c = 2a, Γ = 2πΩa 2 ; the
é a πΩa 2 æ a 2 öù
L′ = 2πρa UΩ ê1 + 2 −
2
ç1 + ÷ú . (X.6)
ë 8h 2hU çè 4h 2 ÷ø û
If U and Ω are given we could compute h from the balance of buoyant weight and lift
L′ = πa 2 ( ρ p − ρ ) g .
Another formula for the lift of on particle in an inviscid fluid in which uniform motion is
perturbed by a weak shear was derived by Auton 1987 and a more recent satisfying derivation of
the same result was given by Drew and Passman 1999. They find that
2
L = πa 3 ρ ω /\ (u − U) (X.7)
3
o
for the lift force on sphere of radius a. In plane flow ω = e z γ , u − U = e x (u − U ) , gravity is
g = − ge y γ = − g (e x /\ e z ) , we have

4
L = − πa 3 ρ Ω f (u − U )e y (X.8)
3
where
o
2 Ω f = − γ = − du . (X.9)
dy
If du/dy > 0 the sphere is lifted against gravity when the slip velocity u – U is positive; if u – U is
negative the sphere will fall. Particles which lag the fluid migrate to streamlines with faster flow,
particles which lead the fluid migrate to streamlines with slower flow.
There are rather striking differences between (X.8) and (X.4); first (X.4) depends on the
angular velocity of the particle but (X.8) depends on the angular velocity of the fluid. Both
formulas leave the slip velocity undetermined, u – U appears in (X.8) because of the shear, in
(X.4), u = 0. The slip velocities have to be prescribed in these theories because the particle
velocity is not determined by viscous drag. Similarly the angular velocity of the particle cannot
arise from torques arising from viscous shears. The effects of particle rotation cannot be obtained
by the method of Auton 1987.
§ Low Reynolds numbers

Other lift formulas have been obtained in the limit of low Reynolds numbers. Inertial effects
are much smaller than viscous effects near a slowly moving particle, but these effects are
comparable far away from the particle. This non-uniformity generates mathematical difficulties,
which are especially severe for perturbations; different regimes of flow must be distinguished
and different cases arise. We shall not present a review of these difficulties in detail; extended
discussion of the subtle problems arising in discussions of lift at low Reynolds number have been
presented by Brenner 1966, Cox & Mason 1971, Leal 1980, Feuillebois 1989, Cherukat and
McLaughlin 1994, and Asmolov 1999, among others. The review of the literature given in the
DNS study of migrations of particles in plane Couette and Poiseuille flow is of value since it
compares analytical studies of migration with restricted domains of validity with direct
numerical solutions of similar problems when the restrictions are relieved.
In unbounded domains Stokes flow is not uniformly valid even though the celebrated Stokes
flow solution giving the drag on a sphere is an achievement of enormous utility. The problem
becomes evident when one perturbs this solution with small inertia to correct the drag. The
perturbation in powers of the Reynolds number (inertia) fails because there are important effects
of inertia in the far field at lowest order which are not present in Stokes flow. The same problem
has even stronger consequence in plane flow; there is no bounded Stokes flow for flow over a
body.
Oseen solutions of the Navier-Stokes equations linearized for uniform flow over a body can
be obtained and matched with the Stokes solution near the body by methods of matched
asymptotic expansions. This leads to a correction of Stokes drag for flow over a sphere in three
dimensions; in two dimensions the method of matched asymptotic using Oseen’s solution gives a
low Reynolds number solution in the case where Stokes flow fails.

The same problem of perturbing Stokes flow for corrections of drag occur when perturbing
Stokes flow for lift; on unbounded domains it is necessary to turn to matched asymptotic
solutions.
It is essential to understand that Stokes flows will not generate lift; to get lift from the
Navier-Stokes equations some effects of inertia must be considered. Saffman 1956 and
Bretherton 1962 showed that no sideways force on a single rigid spherical particle can be derived
from Stokes equations whatever the velocity profile and relative size of particle and tube if the
velocity is unidirectional.
We turn now to some studies of lift in low Reynolds number flows on unbounded domains
obtained by matching inertialess solutions near a body to Oseen’s solution far away. Those cases
could be considered; uniform flow flows with a uniform shear, like Couette flow, and flows like
Poiseuille flow with a shear gradient. Rubinow and Keller (RK) 1961 derived a formula for the
transverse force on a spinning sphere in a viscous fluid which is at rest at infinity, at low
R = Ua/ν
L = πa 3 ρ Ω /\ U(1 + O(Re) ) . (X.10)
If Ω and U are orthogonal, (X.10) reduces
L = ρπa 3UΩ . (X.11)
In RK’s problem the velocity U of the sphere and Ω its angular velocity are maintained by an
external agent and a drag and torque may be computed.
The forms of (IX.1) and (X.11) are identical apart from a prefactor despite the fact that (IX.1)
is for two dimensions in the flow of an inviscid fluid and (X.11) is for a sphere in a highly
viscous (low R) flow. The ratio L/L′ = a/2 differs by a/6 from the ratio of the volume to the
perimeter of a sphere.
The RK lift (X.11) also differs from Auton’s 1987 lift (X.7) in several ways: the prefactor in
(X.7) is 4/3 times that in (X.11); (X.11) is for uniform flow and (X.7) for shear flow; (X.11) is
for low and (X.7) for high Reynolds number with no viscous effects; and RK lift depends on the
spin of the particle but Auton’s lift depends on the “spin” Ωf of the fluid. Apart from these
gigantic differences the formulas look alike. We might ask when the particle spin, the fluid spin
or the difference Ωf -Ω is most important. This comparison brings out the importance of the way
in which the measure of angular velocity enters into lift; we could look at the difference of lift
when the particle rotates freely and when the particle rotation is suppressed. In figure IX.9 we
show that a heavier than liquid cylinder will rise higher when rotation is suppressed than when it
rotates freely, and table IX.2 shows that the slip angular velocity is greatly reduced when the
particle rotates freely.
Application of Auton’s formula (X.7) to a real situation could overestimate the lift due to the
fact that the angular velocity of the particle cannot adjust to the shear stresses which are
generated by shear at the particle surface in a viscous fluid.

§ Lift in shear flows at low R

The experiments of Segré and Silberberg 1961, 1962 have had a big influence on fluid
mechanics studies of migration and lift. They studied the migration of dilute suspensions of
neutrally buoyant spheres in pipe flows at Reynolds numbers between 2 and 700. The particles
migrate away from the wall and centerline and accumulate at 0.6 of a pipe radius.
Rubinow and Keller 1961 discussed the aforementioned experiments and tried to apply their
lift theory to explain the observations. They remark that their theory is for a sphere in a uniform
flow but the Poiseuille flow in a tube is not uniform; it is a shear flow with a parabolic profile.
They say that the shear accounts for the spin and the shear gradient accounts for the lag.
Bretherton 1962 considered the lift and drag force on a cylinder (2D sphere) and Saffman
1965 the lift force on a sphere in an unbounded shear flow. They matched Stokes flow near the
body to Oseen flow far away. Bretherton found that the lift and drag per unit length at small
values of R = γ a 2 / ν is given by
21.16ηU s
L′ = ,
(0.679 − ln( R/ 4 )) + 0.634
2
(X.12)
D′ =
( ( )) ,
4πηU s 0.91 − ln R/ 4
(0.679 − ln( R/ 4 )) + 0.634
2
where Us is the slip velocity, η the viscosity, v = η/ρ. Saffman 1965 found that the lift on a
sphere in a linear shear flow is given by
o
L = 6.46ηa 2U s (γ /ν )1 / 2 + lower order terms . (X.13)
The lower order terms are
é11 o ù é 11 ù
− U s a 3 ρ f + ê γ − π Ωú = − U s a 3 ρ f ê− Ω f − π Ωú (X.14)
ë8 û ë 4 û
and they are lower order as ν → ∞ for small R. When Us > 0 the sign of the first term is positive
when γ > 0 , negative when γ < 0 ; assuming that γ > 0 , the lift is positive when Us > 0 and
negative when Us < 0. The lower order terms depend strongly on the slip angular velocity Ωp -
Ωp = Ωp + γ 2 . Saffman’s result requires that the particle Reynolds number based on the slip
velocity is small 2Us a/ν << 1 and the flow Reynolds number R = γ a 2 / ν is small and another
restrictive condition which was removed by McLaughlin 1991 and Asmolov 1990.
These results cannot explain Segré-Silberberg’s observations, which require migration away
from both the wall and the center. There is nothing in these formulas to account for the migration
reversal near 0.6 of a radius. Moreover the slip velocity Us and angular slip velocity which are
functionals of the solution are prescribed in these studies.

§ Slip velocity and lift

A definite value for the slip velocity may be obtained by preventing lateral migration by
balancing the hydrodynamic lift with the buoyant weight of the particle. Applying this balance to
(X.12) and (X.13) we get
êë
(
πd 2 (ρ p − ρ f )g é 0.679 − ln R/ 4 ( ))
2
+ 0.634ù
úû
Us = for a circular particle,
84.63η (X.15)
πd (ρ p − ρ f )g
2
Us = for a sphere.
9.69η R
The net buoyant weight on a neutrally buoyant particle (ρ p = ρ f ) is zero; hence Us = 0 and
L = 0. The Bretherton and Saffman formulas thus predict that a freely moving neutrally buoyant
circular or spherical particle will have zero slip velocity in a linear shear flow in an unbounded
domain.
Lin, Peery and Schowalter 1970 did a low Reynolds number analysis of simple shear flow
around a rigid sphere by matching Stokes and Oseen flow. They enforced the condition of no net
force or torque and found that
U s = 0,
æ
o 1
ö (X.16)
Ω p = − γ ç − 0.153δ R 3 / 2 ÷
è2 ø
where R = γ a 2 /ν . This shows that the slip velocity in an unbounded linear shear flow vanishes
but the angular slip velocity is proportional to R3/2. Relaxing the assumption of small Reynolds
numbers, Feng, Hu and Joseph 1994 did a direct simulation of a circular particle in plane Couette
flow. They found that the particle migrates to center of the channel where it moves with the fluid
but rotates at 46% of γ when the particle Reynolds number is about 0.63. Cox, Zia and Mason
1968 showed that the angular slip velocity vanishes for a freely moving and rotating particle in
Stokes flow.
N. Patankar and Hu (Patankar 1997) gave a symmetry argument which suggests that
spherical bodies which move with the fluid in a linear shear flow experience no lift even when
they are not neutrally buoyant. This symmetry is displayed in figure X.6 where the imposed
shear is represented in a coordinate system fixed at the shear center. The up-down symmetry
implies that up-down lift are equally possible, hence the lift, and also the drag, is zero.

Us = 0
y Ωp
II I
x
III IV
Imposed shear flow I == III and II = IV Imposed shear flow
Figure X.6 A neutrally buoyant particle in an unbounded linear shear flow.
The symmetric Us = 0 solution for a neutrally buoyant circular or spherical particle moving
freely in an unbounded linear shear flow may be unstable under certain conditions not yet
understood.
§ Non-uniqueness
McLaughlin 1991 generalized Saffman's formula for lift; he found that
6.46 η 2 J (ε )
L= R ,
2.255 4 ρ f ε
(X.17)
R
ε= ,
Rs
where J is a function of ε only and Rs = ρfUsd/η is the slip Reynolds number. The function J has
a value of 2.255 as ε → ∞ (the Saffman limit). Figure X.7 shows the plot of J(ε)/ε as a function
of ε (for ε ≥ 0.025) based on the data provided by McLaughlin 1991. For a neutrally buoyant
particle J(ε)/ε = 0 i.e. ε = 0.218 or ε ~ ∞ (figure X.7). There is probably another value of
ε < 0.025 at which J(ε)/ε = 0 but we do not have that data. Equation (X.17) implies
Rη
Us = or Us = 0 (prediction from the Saffman formula); hence the slip velocity is not
0.218 ρ f d
single valued for a given L. The argument also works for non-neutrally buoyant particles. The
two solutions of McLaughlin’s equation may not both be stable.

2 (a) 2 (b)
1.5 1.5
J/ e
J/ e
1 1
0.5 0.5
0 0
0 5 10 15 20 0 0.5 1 1.5 2
e e
Figure X.7. Graphs of J(ε)/ ε vs. ε for ε > 0.025. The graphs are based on McLaughlin's data for the lift
on a sphere in an unbounded linear shear flow.
§ Validation of lift formulas by DNS

An important application of DNS is to establish if and when analytic expressions for lift are
valid. Most of the expressions found in the literature are for spheres and the formula will be
tested in the future. An example of this kind of testing will now be given for Bretherton's
formula's (X.12) for the lift and drag on a cylinder in a linear shear flow.
Bretherton's analysis does not apply to the case of a freely moving cylinder in equilibrium
under the balance of weight and lift. The condition of zero drag, required for steady motion, is
not respected. Assuming some engine to move the particle with the required drag, we may
compare this formula with the results from DNS.
V1
y
W x
Us
-V2
W
Figure X.8. Computational domain for the simulation of linear shear flows around a circular particle.
Numerical simulations were performed in a square channel of size W × W. The channel size
should be large enough to simulate flows in an unbounded domain. The circular particle was
placed at the center of the channel and the coordinate system at the center of the particle. The
velocity boundary conditions are as shown in figure X.8. The upper wall moves with velocity V1

and the bottom wall with velocity -V2. The shear-rate γ = (V1 + V2 ) W and the slip velocity Us is
as shown in figure X.8. The particle is free to rotate so that the net torque is zero at steady state,
and γ and Us were varied in the simulation.
0.1
0.09
0.08
0.07
Lift (dyne/cm)
0.06
0.05
0.04
0.03
0.02
0.01
0
0 100 200 W/d 300 400 500
Figure X.9. Lift vs. domain size for a particle in an unbounded linear shear flow.
The fluid density is 1 g/cc, viscosity is 1 poise and the particle diameter is 1 cm. At t = 0+ the
flow is started by imposing the boundary conditions. The particle begins to rotate until a constant
angular velocity is reached at steady state. The hydrodynamic lift (in the y-direction) and drag (in
the x-direction) on the particle is calculated. Figure X.9 shows the plot of the lift force on the
particle as a function of W for R = 0.01 and Rs = 0.1. The simulations were carried out on a
sequence of domains of increasing size. If this procedure is to yield a result which is
asymptotically independent of the size of the domain then the curve giving lift vs. domain size
ought to flatten out. Figure X.9 shows just such a flattening. Though the curve is still rising
modestly at W = 450d, we have used this domain for the simulations in table X.1. In this table
the computed values of lift and drag are compared to the analytical values from Bretherton’s
expressions (X.12). The drag force is in better agreement than the lift. Larger domains may lead
to better agreements.
Rs = 0.003 Rs = 0.1
DNS Analytic % Error DNS Analytic % Error
R = 0.01 Lift 0.00347 0.00449 -22.72 0.08593 0.1496 -42.56
Drag 0.01010 0.01041 -2.98 0.3374 0.3471 -2.79
R = 0.02 Lift 0.00436 0.00542 -19.56 0.1239 0.1806 -31.39
Drag 0.01093 0.01145 -4.54 0.3637 0.3818 -4.74
Table X.1. Comparison between the numerical and analytic values (equation X.12) of lift and drag per
unit length (in CGS units). The error is calculated with respect to the analytic value.

§ Wall effects in shear flows

(X.6) shows how the lift on a cylinder rotating the fluid in an aerodynamic approximation
might be increased as it moves to the ground. A similar enhancement of lift in a shear flow near a
wall occurred at low Reynolds numbers. The generation of lubrication layers in flowing
suspensions is most probably a consequence of enhanced wall lift. The analysis of lift at low R is
complicated by the non-uniform way in which viscosity and inertia enter the problem. A
neighborhood ds > d of a sphere moving with a slip velocity Us is dominated by viscous forces
when
U sds
Re ( d s , U s ) = << 1 . (X.18)
ν
A neighborhood lw > 0 of the wall in a shear flow is dominated by viscous forces when
o
o γ l2
Re (l w , γ ) = w << 1 . (X.19)
ν
The neighborhood l > 0 between the particle and the wall is dominated by viscous forces when
l < min(d s , l w ) . (X.20)
The effect of inertia in this region dominated by viscosity enters as a perturbation of Stokes flow
with terms arising from ρ(u•∇)u, where u(x,t) is computed on Stokes flow. The perturbations are
regular rather than singular because the lowest order, Stokes flow, does not need to be corrected
for inertia far away from the particle. Lift formulas arising from regular perturbations are
proportional to ρ and to bilinear and quadratic powers of prescribed data U s , γ, Ω for Stokes
flow; the viscosity η does not enter into the coefficients of the lift as in the singularly perturbed
problems studied by Bretherton 1967 and by Saffman 1965 (X.12-X.14).
There is no need to consider singular perturbations in channels of width less than l; the region
above a wall can also fall in this class, if the particle does not move so far away that the ratio
ρu ⋅ ∇u
(X.21)
η∇ 2 u
is not uniformly small. This fortunate condition could apply to the uniform suspension of heavy
particles against gravity.
An excellent review of literature on lift and migration of single spheres in low Reynolds
number flows can be found in the paper by Cherukat and McLaughlin 1994. All of the papers
except Leighton and Acrivos 1985 and Krishnan and Leighton 1995 require also that the radius
of the sphere be small compared to the distance between the sphere and the wall and treat the
particle as a point or forced doublet. A consequence of this additional assumption is to interdict
the study of the effects of the angular velocity of the sphere on lift.
Cox and Brenner 1968 introduced the idea that inertial perturbations of Stokes flow could be
carried out in regions in which the ratio (X.21) is uniformly small. Cox and Hsu 1977 computed

the lift for inertial perturbations of Stokes flow for a sedimenting sphere and for spheres in linear
and quadratic flows.
The inertial lift on a sphere translating in a shear flow bounded by a single flat infinite wall
was analyzed by McLaughlin 1993. He derived an expression for the lift force by superposition
of the disturbance flow created by the wall and the migration velocity due to an unbounded shear
field. The analysis required asymptotic matching in the far field but it converges to the regular
perturbation formula (X.22) of Cox and Hsu 1977 when the distance from the wall decreases. An
analytical form for the lift
L = 6πηaυm , (X.22)
− 0.2855 aU s
1/ 2
υm = æ γo ν ö
ç ÷ (X.23)
l ∗5 / 3 è ø
where
o 1/ 2
l = l æç γ v ö÷
∗
è ø
⋅ ⋅
and υm is the migration velocity, is valid for large separation, l* >> 1. When γ is negative γ is
⋅
replaced (X.20) by | γ | and the minus sign is replaced by plus; the particle migrates in an
opposite sense.
§ Curvature
Gradients of shear (curvature) produce lateral forces. At the centerline of a Poiseuille flow
the shear vanishes, but the shear gradient does not. To understand the Segré-Silberberg effect it
is necessary to know that the curvature of the velocity profile at the center of Poiseuille flow
makes the center of the channel run unstable position of equilibrium. A particle at the center of
the channel or pipe will be driven by shear gradients toward the wall; a particle near the wall will
lag the fluid and be driven away from the wall. An equilibrium radius away from the center and
wall must exist.
Ho and Leal 1974 were the first to combine these effects in an analysis of the motion of a
neutrally buoyant sphere rotating freely between plane walls so closely spaced that the inertial
lift can be obtained by perturbing Stokes flow with inertia. They treated wall effects by a method
of reflection and found that
a4 2
L = 36 ρ 2 U m (1 − 2 β )[(1 − 2 β ) G1 ( β ) − G2 ( β )]
l
where β = x/e and G1(β), G2(β) are functions whose values are given in their paper; the
equilibrium positions are the center line β = 0.5, which is unstable, and β = 0.2 and 0.8, which is
0.6 a from the center. This expression is valid when the sphere is not close to a wall. Vasseur and
Cox 1976 used another method to treat wall effects and their results are close to Ho and Leal’s
near the center line but rather different than those of Ho and Leal near the wall. The requirement
a/l << 1 prevents them from obtaining results just next to the wall. Feng, Hu and Joseph 1994
studied the motion of solid circles in plane Poiseuille flow by DNS. The circle migrates to the
0.6 of a radius equilibrium position. They compared their 2D results with those of Ho & Leal and
Vasseur & Cox.
The experiments of Segré & Silberberg 1961, 1962 do not satisfy the condition of low
Reynolds number or those required to carry out an inertial perturbation of Stokes flow without
asymptotic matching using Oseen’s solution. Schonberg and Hinch 1989 analyzed the lift on a
neutrally buoyant small sphere in a plane Poiseuille flow using matched asymptotic methods.
The same problem for neutrally buoyant and non-neutrally buoyant small sphere studies using
asymptotic method by Asmolov. The Reynolds numbers based on the slip velocity is small for
both analyses but there is no explicit condition on the channel Reynolds number. Asmolov 1999
found that
… wall induced inertia is significant in thin layers near the walls
where the lift is closed to that calculated for linear shear flow,
bounded by a single wall. In the major portion of the flow,
excluding near-wall layers, the wall effect can be treated as
unbounded parabolic shear flow. The effect of the curvature of the
unperturbed velocity is significant, and the lift differs from the
values corresponding to linear shear flow even at large Reynolds
numbers.
The analyses mentioned in the paragraph above take the effect of inertia (u • ∇ )u into
account only in an Oseen linear system; the comparison of the results of these analyses with
experiments is far from perfect. The analysis is heavy and explicit formulae for lift are not
obtained.
§ Regular perturbation in the wall region

The question addressed here is whether the regular perturbation of Stokes flow with inertia in
regions dominated by viscosity (X.20) is analytic in the Reynolds number. To be precise about
this we may think that the l in (X.20) is the distance between plane walls and consider steady
motion of a sphere in a plane Poiseuille flow as was considered by Cox & Brenner 1968, Ho &
Leal 1974, Cox & Hsu 1977 and Vasseur & Cox 1976. In the neutrally buoyant case these
authors look for solutions of the form
é u ù = é u 0 ù + éu1 ù R + O ( R ) (X.24)
êë p úû êë p 0 úû êë p1 úû
where say R = U m l / v and u0 , p0 are Stokes flow. The Stokes flow is degenerated in that the
sphere can move parallel to a wall at any distance β = x/l, but steady solutions will arise only
when the sphere is at a position of equilibrium and doesn’t migrate. These are the Segré &
Silberberg positions of equilibrium. The little O(R) is not specified in these studies because it is
not needed. It seems probable to me that in the equilibrium case of an R family of steady
solutions of non-migrating spheres the solution is analytic and may be represented by power
series

∞
é u ù = éu n ù R n
êë p úû å êp ú (X.25)
n=0 ë n û
Consider the perturbation problems for the steady motion U = e xU of a sphere parallel to the
stationary walls of a Poiseuille flow u = e x u~( y ) , as in figure IX.2. The equations are made
dimensionless with a and U, R = Ua/v, and they are written in a coordinate system fixed on the
sphere. The equations are
R (u • ∇ )u = div σ[u], div u = 0 (X.26)
u = (Ω / U ) /\ x, x = 1 ü
ï
u = −1 e x on the walls ý (X.27)
u = (u~( y ) / U − 1) e x as x → ∞ ïþ
where
div σ[u] = −∇p + ∇ 2 u . (X.28)
At zeroth order, Stokes flow, we have
div σ[u 0 ] = 0, div u 0 (X.29)
where u0 satisfies the conditions (X.27). Higher order problems are in the form n = 1, 2, …
ìf n (u 0 , u1 ,Ku n −1 ) = div σ[u n ], div u n = 0

ï
ï ìon x = 1
í ï (X.30)
ïu n = 0 íon the wall
ï ïat infinity
î î
where fn depends on only solutions at lower order um, m < n. For example
f (u 0 ) = (u 0 • ∇ )u 0 . (X.31)
The lift-off problem for non-neutrally particles which was studied by DNS in chapter IX is
not conveniently framed as a power series, since lift-off occurs at a finite Reynolds number;
below this the particle slides and rolls and the sliding and rolling might be generated as a power
series. Another possibility would be to develop the steady forward motion of a heavy sphere or
cylinder subject to zero torque and net force, with lift balancing buoyant weight as a power series
in the Reynolds number, which is zero at lift-off.
It is something of a mathematical mystery as just why it is that the solution of the problem of
a sphere moving forward in an infinite domain is not analytic as can be seen already in the lift
formulas (X.12) of Bretherton 1962 and (X.13) of Saffman 1965. The drag formula for a single
sphere of a radius a moving forward with velocity U in an infinite fluid which was derived by
Proudman and Pearson 1957

é 3 9 2 ù
D = 6πηaU ê1 + R + R ln R + O( R 2 )ú
ë 8 40 û
where R = au/v does not reverse sign automatically when the sphere velocity is reversed;
conditions specified by T.B. Benjamin 1993 are required to guarantee the required symmetry.
§ Reciprocal theorem
The lift on a sphere translating parallel to a wall may be assumed to be a power series
L= åL R
n =0
n
n
. (X.32)
The reciprocal theorem of Lorenz can be used to show that the lift coefficient at order n may
be computed on a solutions of lower order m < n. The same kind of analysis can be used to
compute the migration velocity perpendicular to the wall when this migration is not suppressed
as shown first by Cox 1965 and comprehensively by Ho & Leal 1974. The proof of the
reciprocal theorem given below applies to the problem studied by Cherakut & McLaughlin 1994
and generalized the result for L1 to Ln.
Consider an auxiliary Stokes flow problem between flat wall
div σ[q ] = 0, div q = 0 ü

ï
q = e y on x = 1 ý (X.33)
q = 0 on the wall and at ∞ ïþ
Using (X.30), we form the integral
òV (q • div σ[u n ] − u n • div σ[q ]) dV = ò q • f n dV

V
where V is the region occupied by fluid. Using the divergence theorem and the boundary
conditions (u n = 0 on ∂V )
ò q • (n • σ[u ]) d A + I = Vò q • f
∂B
n n dV (X.34)
where A is an element of area on the boundary ∂B of the sphere B and
æ ∂q ∂u ö
I = ò ç − i σ ij [u n ] + ni σ ij [q]÷ dV
ç ∂x ∂x j ÷
Vè j ø
= ò {− Dij [q ](− p[u n ]δ ij + Dij [u n ]) + Dij [u n ](− p[q]δ ij + Dij [q])}dV
V
= ò ( p[u n ] div q − p[q] div u n ) dV = 0.

V
Hence, using q = ey on |x| = 1, we get

Ln = ò e • (n • σ[u ])d A + I = Vò q • f
∂B
y n n dV . (X.35)
Since fn depends on Um, m < n, so does Ln. Using (X.31), we get
L1 = ò q • ((u • ∇ )u 0 )dV . (X.36)

V
In the low R analysis given in the sequel we mean L1 when we write L .
§ Finite size sphere near a wall

Leighton & Acrivos 1985, Cherakut & McLaughlin 1994 and Krishnan & Leighton 1995
have studied the inertial lift of finite size spheres as a perturbation of Stokes flow in the near wall
region.
Leighton & Acrivos 1985 evaluated the lift on a stationary sphere in a shear flow to the
lowest order in the Reynolds when the sphere touches the wall; the lift points away from the wall
and varies as the fourth power of the radius of the sphere and the square of the shear rate.
Cherakut & McLaughlin 1994 derived an expression for the lift of a spherical particle in a
linear shear flow perturbing Stokes flow. They represented the sphere and the wall in a bi-polar
spherical coordinate system (ξ, η, φ) and an associated cylindrical polar coordinate system. The
coordinate ξ = α corresponds to the sphere (the center of the sphere being located at z = l, ρ = 0)
and the coordinate surface ξ = 0 corresponds to a sphere of infinite radius which coincides with
the wall. The bispherical solution method breaks down as the separation distance between the
sphere and the wall vanishes; results were obtained for sphere-wall separations greater or equal
to 0.1 radius.
The Stokes flow solution for steady solutions of a sphere moving parallel to a wall requires
selection of the distance from the wall. Leighton & Acrivos 1985 and Krishnan & Leighton 1995
put this separation to zero, the sphere touches the wall. Cherakut & McLaughlin 1994 leave this
distance arbitrary.
Krishnan & Leighton 1995 studied the inertial lift on a translating and rotating sphere in
contact with a plane wall in a shear flow. The calculation requires three independent Stokes flow
solutions for prescribed values of a translational velocity U; angular velocity Ω and wall shear
γ . The Stokes flow solutions may be superposed.
o
u 0 = u 01U + u 02 Ω + u 03 γ .
The lift is presented in terms of quadratic and bilinear products of these coefficients
o2 o
L /ρ = 9.275a 4 γ + 1.755a 2U 2 + 0.546a 4 Ω 2 + 1.212a 4 γ Ω
(X.37)
o
− 2.038a UΩ − 9.044a U γ .
3 3
This formula is independent of viscosity.

They note that when this lift exceeds the buoyant weight, the sphere will rise ultimately to an
equilibrium height h for which
L = (ρ p − ρ ) πa 3 g
4
(X.38)
3
under the conditions that the net force accelerating particle and the net torque on the particle
vanish. These three conditions for steady motion driven by shear of a particle parallel to a wall
requires Stokes flow solutions not only when the particle touches the wall but when it is a
distance h away from the wall. Given a translating and rotating sphere in Stokes flow in a shear
field at a height h we may select Ω(h, γ ), U (h, γ ) so that the three contributions to the total for
forward force and torque vanish:
o
D0 = D 01U + D 02Ω + D 03 γ = 0 . (X.39)
And torque
o
T0 = T 01U + T 02Ω + T 03 γ = 0 . (X.40)
Then (X.38) determines h (γ ) , the equilibrium height, and the equilibrium solution
U (γ ), Ω(γ ), h (γ ) .
To carry out the program outlined in the paragraph above, Krishnan & Leighton 1995 used
an approximate “lubrication” solution of Stokes equations of Goldman, Cox & Brenner 1967 for
the translation and rotation of a sphere parallel to a plane wall in a semi-infinite fluid when the
gap width h is very small. Krishnan & Leighton 1995 say that the dependence of the inertial lift
on h is relatively weak and use the formula (X.37) to compute the equilibrium h as a function
of γ .
The separation distance h, the forward slip velocity Us of the sphere, the angular velocity
Ω of the fluid and the shear rate γ are prescribed in the Stokes flow solutions used by Cherakut
& McLaughlin 1994. This is more accurate than the lubrication solution of Goldman et al 1967
used by Krishnan & Leighton 1995. They computed the lift L (γ , h,U s , Ω ) using the reciprocal
theorem (X.36) and numerical quadrature. They find that
L = L = ρV 2 a 2 I (κ , Λ γ , Λ Ω, Λ s ) (X.41)
where κ = a / h, Λ γ = γ a / V , Λ Ω , = Ωa / V and Λ s = U s / V . They consider the case of no rotation

Ω = 0 and of a freely rotating sphere taking the values of Ω computed by Goldman, Cox and
Brenner 1967 for zero torque. This reduces their considerations to two families of I depending
on κ , Λ γ , Λ s . They studied the lift force for different prescribed values of Us; when
U s ≠ 0, V = U s , Λ s = 1 and I = I (κ , Λγ ) where Λ γ = γa / U s . The case Us = 0 corresponds to
neutrally buoyant spheres in a linear shear flow when wall effects are not important; there is no
slip. In this case V = γ a, Λ γ = 1, Λ s = 0 .

The analysis of Cherakut & McLaughlin 1994 does not determine the slip velocity Us. They
do extensive numerical studies computing L for different choices of Λγ and κ when Ω = 0 and
when the particle is freely rotating. The computed values are assembled into tables. Since Uo is at
most a linear function of Λγ , the reciprocal theorem (X.36) shows that I is at most a quadratic
polynomial in Λγ . The coefficients were expressed in powers of κ by nonlinear regression from
the tables. They find that when Ω is determined from the analysis of a freely rotating sphere by
Goldman, et al 1967
[
I = 1.7631 + 0.3561κ − 1.1837κ 2 + 0.845163κ 3 ]
é 3.24139 ù
−ê + 2.6760 + 0.8248κ − 0.4616κ 2 ú Λ γ (X.42)
ë κ û
[
+ 1.8081 + 0.879585κ − 1.9009κ + 0.98149κ 3 Λ2γ .
2
]
Whereas when Ω = 0, they get
[
I = 1.7716 + 0.2160κ − 0.7292κ 2 + 0.4854κ 3 ]
é 3.2397 ù
−ê + 1.1450 + 2.0840κ − 0.9059κ 2 ú Λ γ& (X.43)
ë κ û
[
+ 2.0069 + 1.0575κ − 2.4007κ + 1.317κ Λ2γ& .
2 3
]
Recall that L = ρV 2a2I, V = U, or γa when Us = 0.
The analysis of Cherakut & McLaughlin 1994 appear to agree with results obtained by
Leighton & Acrivos 1985 and by Krishnan & Leighton 1995 when κ = a/h is large and with Cox
& Hsu 1977 when κ is small.
Lovalenti 1994 derived an expression for the lift force for small by modifying the results
derived by Cox & Brenner 1968 and Cox & Hsu 1977 in an interesting appendix to the paper by
Cherakut & McLaughlin. He finds that
6π (61) 4 o 2 18π 2 2
L /ρ = a γ + a Us
4 × 144 32
ì 66 æ 1 27 ö 11 üo
+ í− π a 3 ç + ÷ + π a 3 ý γ U s
î 64 è κ 16 ø 8 þ
6π (61) 12 4 o
− ⋅ a Ω γ − π a 4 ΩU s
4 × 144 61 (X.44)
o 2
= 1.99622a 4 γ + 1.76715a 4U s2
æ 3.23977 ö o
+ ç− − 1.14742 ÷a 3 γ U s
è κ ø
o
− 0.39270a 4 Ω γ − 3.1415a 3ΩU s .

This formula can be compared with (X.37); the term proportional to Ω2 is absent and the signs
and magnitude of the coefficients of other terms do not agree.
The expression (X.42) satisfies the condition for a freely rotating sphere. With γ prescribed,
the lift formula (X.43) leave h (γ ) and the slip velocity U s (γ ) undetermined. To determine these
quantities we could apply the lift-buoyant weight balance, as in (X.25) and a zero force balance.

VIII Modeling Rayleigh-Taylor Instability of a sedimenting suspension of circular particles ....65

IX Fluidization by lift: single particle studies .................................................................................85
§ Equations of motion and dimensionless parameters ..................................................................................... 86
§ Lift-off of a single particle in plane Poiseuille flows of a Newtonian fluid.................................................. 89
§ Data Structure for DNS and experiments ..................................................................................................... 95
X Analytical models of lift .............................................................................................................100

§ Lift in an inviscid fluid ............................................................................................................................... 100
§ Low Reynolds numbers .............................................................................................................................. 102
§ Lift in shear flows at low R......................................................................................................................... 104
§ Slip velocity and lift.................................................................................................................................... 105
§ Non-uniqueness .......................................................................................................................................... 106
§ Validation of lift formulas by DNS ............................................................................................................ 107
§ Wall effects in shear flows.......................................................................................................................... 109
§ Curvature .................................................................................................................................................... 110
§ Regular perturbation in the wall region ...................................................................................................... 111
§ Reciprocal theorem..................................................................................................................................... 113
§ Finite size sphere near a wall...................................................................................................................... 114

Slip velocity and lift at finite Reynolds numbers
XI Slip velocity and lift at finite Reynolds numbers

In this chapter we assemble results from direct numerical simulation, analysis and
experiments that enter into an analysis of lift. We define and show that the slip angular velocity
discrepancy is an important quantity that determines the equilibrium of a free particle in flows
with a shear gradient. The equilibrium values of free particles lose stability as the shear Reynolds
number is increased; they undergo turning point bifurcations leading to nonuniqueness and
hysteresis.
§ Equilibrium positions of neutrally buoyant and heavy particles

We have already noted that the experiments of Segré and Silberberg 1961, 1962 have had a
big influence on fluid mechanics studies of migration and lift. They studied the migration of
dilute suspensions of neutrally buoyant spheres in pipe flows at Reynolds numbers between 2
and 700. The particles migrate away from the wall and centerline and accumulate at 0.6 of a pipe
radius.
The lift on heavier than liquid particles is also influenced by the factors that determine the
equilibrium position of neutrally buoyant particles. The heavy particles must reach an
equilibrium that balances the hydrodynamic lift and buoyant weight. If the buoyant weight is
very small, the equilibrium position of the particles will be close to the value for the neutrally
buoyant case. The effect of increasing the weight is to lower the equilibrium position whose zero
is established for the case of zero buoyant weight.
Most attempts to explain the Segré-Silberberg effects have been based on linearized low
Reynolds number hydrodynamics. Possibly the most famous of these attempts is due to Saffman
1965. There are a number of formulas like Saffman's that are in the form of Us times a factor,
which can be identified as a density times a circulation as in the famous formula ρUΓ for
aerodynamic lift. A relatively recent review of such formulas can be found in McLaughlin 1991.
Formulas like Saffman's cannot explain Segré-Silberberg's observations, which require
migration away from both the wall and the center. There is nothing in these formulas to account
for the migration reversal near 0.6 of a radius. Moreover the slip velocity Us, the angular slip
velocity Ωs = Ωp - Ωf = Ωp + γ 2 , the particle velocity and the particle angular velocity, which
are functionals of the solution are prescribed quantities in these formulas.
The fluid motion drives the lift on a free body in shear flow; no external forces or torques are
applied. If there is no shear there is no lift. In Poiseuille flow there is not only a shear but a shear
gradient. Gradients of shear (curvature) produce lateral forces. At the centerline of a Poiseuille
flow the shear vanishes, but the shear gradient does not. To understand the Segré-Silberberg
effect it is necessary to know that the curvature of the velocity profile at the center of Poiseuille
flow makes the center of the channel an unstable position of equilibrium. A particle at the center
of the channel or pipe will be driven by shear gradients toward the wall; a particle near the wall
will lag the fluid and be driven away from the wall. An equilibrium radius away from the center
and wall must exist. The effects of migration and lift which determine the equilibrium position of
a particle in a Poiseuille flow were discussed as a perturbation of Stokes or Oseen flow by Ho
and Leal 1974, Vasseur and Cox 1976, Schonberg and Hinch 1981 and Asmolov 1999.
Printed 03/11/02 118 ♦ Interog-4C.doc

This paper approaches the problem of migration and lift in a different way. Basically we have
used direct numerical simulation (DNS) to formulate and validate a long particle model that
gives a very good, completely explicit analytical approximation to the velocity and slip velocity
of circular particles. DNS is used here as a diagnostic tool to analyze the role of the slip velocity,
and the angular slip velocity on migration and lift. We are able in this way to establish a rather
simple picture of lift and migration that in particular clarifies the role of the angular slip velocity,
and is not restricted to low Reynolds numbers. Our analysis is carried out in two dimensions but
should apply in principle to 3D, which is at present under study.
§ Mechanism for lift

We can look at formulas for the lift on a particle in an inviscid fluid, which were discussed in
chapter X, and which can be viewed as realizations of Rayleigh's lift formula (X.1) L' = ρUΓ. A
circular particle experiences a lift per unit length (X.4) L' = 2πρa2UΩ; this may be compared
with Auton's formula (X.8) for the L on a sphere rotating in a shear flow. L = –4/3 ρa3ΩfUs where
Ωf = –1/2 du/dy is the angular velocity of the fluid.
If du/dy > 0 the sphere is lifted against gravity when the slip velocity Us is positive; if Us is
negative the sphere will fall. Particles that lag the fluid migrate to streamlines with faster flow,
particles which lead the fluid migrate to streamlines with slower flow.
There are rather striking differences between (X.8) and (X.4); first (X.4) depends on the
angular velocity of the particle but (X.8) depends on the angular velocity of the fluid. Both
formulas leave the slip velocity undetermined, Us appears in (X.8) because of the shear, in (X.4),
Uf = 0. The slip velocities have to be prescribed in these theories because the particle velocity is
not determined by viscous drag; similarly the angular velocity of the particle cannot arise from
torques arising from viscous shears. The effects of particle rotation cannot be obtained by the
method of Auton 1987.
The lift formula ρUΓ captures the essence of the mechanism in which the motion of the
particle relative to the fluid is such as to increase the pressure on the side of the particle as it
moves forward.
The lift on a spherical or circular particle in a shear flow is different; there is no exterior
agent to move and rotate the freely moving particle. Instead the particle is impelled forward and
rotated by the shear flow. Previous theoretical studies and the simulations of Joseph and Ocando
2001 show that the relevant velocity is the slip velocity and the relevant circulation is
proportional to an angular slip velocity discrepancy
Γ ∝ Ωs - Ωse (XI.1)
where Ωse is the slip angular velocity in steady flow (equilibrium). This conclusion will be
established in the sequel. For now we simply note that in the simulations the angular slip velocity
discrepancy Ωs - Ωse < 0 when the cylinder is above the equilibrium (Segré-Silberberg) position
and Ωs - Ωse > 0 when it is below the equilibrium (figure XI.8).

§ Numerical simulation of migration and lift

In numerical experiments of solid-liquid flows we can examine physical effects one at a time;
this cannot be done in real experiments. For the present application we look first at the effect on
particle migration of controlling the angular velocity of the particle. In figure IX.8 we plotted the
rise to equilibrium of a neutrally buoyant particle for three different values of the slip angular
velocity
o
ìo ü
γ ïγ ï
Ω s = + Ω p = í , Ω se , 0ý. (XI.2)
2 ïî 2 ïþ
The rise is the greatest when the particle angular velocity Ωp = 0 and the least when the particle
angular velocity is equal to the local rate of rotation Ωp = – γ 2 . The rise of a heavier than liquid
ρp/ρf = 1.01 circular particle is plotted in figure IX.8 for Reynolds number Rw = γ w d 2 ν = 5.4
and for Rw = 16.2 in figure IX.9. The angular slip velocity Ωse > 0 is the equilibrium value that a
free circular particle takes in torque-free motion when the angular acceleration vanishes. We call
attention to the fact that Ωse > 0 is very small, and at equilibrium
o o
γ γ
> Ω pe , Ω pe ≈ − . (XI.3)
2 2
In another constrained motion we fix the y position of the particle and compute the slip
velocities and lift (figure X.6). A fixed particle with non-zero lift forces will migrate if the
constraint is relaxed.
100
80
Relative value (%)
60
40
20
ANGULAR Slip Velocity
0 HORIZONTAL Slip Velocity
Lift Force
-20
6 7 8 9 10 11 12
Centerline y (cm) Wall
Figure XI.1. FIXED PARTICLE: Rw = 20, ρs /ρf = 1. Steady state relative values for the lift force and
the slip velocities. In the region close to the wall, the lift force and the horizontal slip velocity have a
similar non-linear behavior. In the region close to the centerline, the lift force appears to be

proportional to the angular slip velocity. Therefore, the lift force may be expressed as a function of
the slip velocity product L = L(U s , Ω s ) .
§ Long particle model

Joseph 2001 proposed a model problem for the velocity of a long particle in Poiseuille flow
(also see Choi and Joseph, 2001). We replaced the circular particle of diameter d with a long
rectangle whose short side is d. The rectangle is so long that we may neglect the effects of the
ends of the rectangle at sections near the rectangle
y’
hA
p1 > p2
p1 u A (y ’) p2
τA
UA
d
UB τB
u B (y ) y hB
x
Figure XI.2. Sketch of flow field under consideration and variables involved in the long particle model.
W = h A + h B + d is the channel height.
center. In that model the long particle was assumed to be rigid but it was noted that a more
realistic model could be obtained by letting the long particle shear. We could choose this shear to
be the same as the shear rate of the circular particle in the approximation in which Ω p = − 12 γ
(figure XI.2).
The long particle model is meant to represent the constrained forward motion with y p fixed
after transients have decayed and steady flow is achieved. The model may be compared with the
numerical simulation satisfying the fluid equations (IX.6) with ∂u ∂t = 0 , the x-component of
(IX.7) with du d t = 0
ò {− p1 + 2ηD[ y]}⋅ ndΓ = 0

1
pe x + e x (XI.4)
Vp
and (IX.8) with dΩ d t = 0 . The y-component of (IX.7) gives the balance of buoyant weight and
lift; the particle density enters into this balance through the buoyancy term. It follows that
buoyancy and the particle density do not enter into the constrained simulations, which determine
the steady motion of the fluid and the forward motion of the particle, and they do not enter into
the long particle model, which approximates the simulation.
The model leads to an explicit expression for the particle velocity and slip velocity in which
vertical migration is suppressed. Since the simulation and the model do not depend on ρ p , there

is a sense in which the results given here are universal. However, each constrained simulation is
realizable for a density given by ρp = ρf + Le/Vpg in which the y-component of (IX.7) is satisfied.
The forces acting on the long particle are the force due to pressure acting on the sides
perpendicular to the flow, and the force due to shear stress acting on the sides parallel to the flow
(figure XI.2). The former force is always positive, while the latter may be positive or negative
depending if the fluid is faster than the particle or vice versa,
(τ A + τ B )l + ( p1 − p 2 )d = 0 (XI.5)
( p1 − p 2 )
τ A + τ B + pd = 0 , p = (XI.6)
l
where the shear stresses are defined by

du A du
τ A = −η (h A ) , τ B = −η B (hB ) (XI.7)
dy ′ dy
The velocity profiles above and below the long particle are given by
p U y′
u A ( y ′) = y ′(h A − y ′) + A (XI.8)
2η hA
p U y
uB (y) = y (hB − y ) + B (XI.9)
2η hB
where different velocities (U A , U B ) were assumed for the top and bottom walls to take into
account the angular speed of the circular particle. The relation between them is given by
U A − U B = 12 γ (hB + 12 d ) d , (XI.10)
where γ ( y ) is the shear rate for the undisturbed flow (without the particle), given by
du p
γ ( y ) = = (W − 2 y ) . (XI.11)
dy 2η
The shear rate on the particle’s sides parallel to the flow may be evaluated from (XI.8), (XI.9)
and (XI.10),
du A γ (h + 1 d )
(h A ) = − p h A + U B + B 2 d (XI.12)
dy ′ 2η hA 2 hA
du B
(hB ) = − p hB + U B . (XI.13)
dy 2η hB
Substituting, recursively, (XI.12) and (XI.13) in (XI.7), and then the resultant equation in (XI.6),
we find that at the top and bottom of the long particle (diameter d):

( p η )(2d + h A + hB )h A hB + γ (hB + d 2) h A d
UA = (XI.14)
2(h A + hB )
( p η )(2d + h A + hB )h A hB − γ (hB + d 2) hB d
UB = (XI.15)
2(h A + hB )
The average particle velocity is:
( p η )(2d + h A + h B )h A h B − γ (hB + d 2) (h B − h A ) 12 d
Up = 1
(U A + U B ) = . (XI.16)
2
2(h A + h B )
The undisturbed flow field (without the particle) can be written as:
p
u( y) = y (W − y ) (XI.17)
2η
At the position where the center of the particle is located ( y p = hB + 12 d ) , the undisturbed fluid
velocity is:
p
u (hB + 12 d ) = (h B + 12 d ) (h A + 12 d ) (XI.18)
2η
The particle slip velocity can be defined as:
U s = u (h B + 12 d ) − U p (XI.19)
which can be written as:
( p η )[(h A + hB )(hB + 12 d )(h A + 12 d ) − (2d + h A + hB )h A hB ] + γ (hB + 12 d )(hB − h A ) 12 d
Us = (XI.20)
2(h A + h B )
The channel height W and the position hA and hB satisfy the following conditions:
ìh A = W − (hB + d ),
í y p = h B + 12 d
îh A + hB = W − d ,
and the shear rate at the particle center is

p
γ (hB + 12 d ) = (W − d − 2hB ) . (XI.21)
2η
Then the slip velocity can be simplified:
Us =
p
2η (hA + hB )
[(hA + hB )(hB + 12 d )(hA + 12 d ) − (2d + hA + hB )hAhB + 12 (h − d − 2hB )(hB − hA ) 22 d ]
(XI.22)
=
p
2η (W − d )
[( )
d 12 W − y p + 14 d 2 (W − d ) ≥ 0
2
]
when d → 0 , we can get U s → 0 .

Wall Wall
12 12
(a) (b)
10 10
8 undisturbed 8 undisturbed
long particle model
y (cm)
y (cm)
6 6 long particle model

DNS DNS
4 4
2 2
0 0
0 10 20 30 40 50 60 0 10 20 30 40 50 60
Wall u (cm s -1 ) Wall u (cm s -1 )
Wall Wall
12 12
(c) (d)
10 10
8 8 undisturbed
undisturbed
long particle model
y (cm)
y (cm)
6 long particle model 6

DNS
4 DNS 4
2 2
0 0
0 10 20 30 40 50 60 0 10 20 30 40 50 60
Wall -1 Wall -1
u (cm s ) u (cm s )
Wall Wall
12 12
(e) (f)
10 10
8 8 undisturbed
undisturbed
long particle model
y (cm)
y (cm)
6 long particle model 6

DNS
DNS
4 4
2 2
0 0
0 10 20 30 40 50 60 0 10 20 30 40 50 60
Figure XI.3. Velocity profiles through the center y p of a particle and the particle velocity at R = 20 . The
velocity profiles of the undisturbed flow, of the DNS simulation and the long particle model are
compared: (a) centerline y p = 6.0 cm, (b) the unique equilibrium position when ρ p ρ f = 1.005
( y p = 3.834 cm), (c) the higher equilibrium position when ρ p ρ f = 1.01 ( y p = 3.165 cm), (d)
y p = 3.0 cm, (e) the lower equilibrium position when ρ p ρ f = 1.01 ( y p = 1.323 cm), (f) y p = 0.75 cm.
The centerline of (a) is unstable.

Wall Wall
12 12
(a) (b)
10 10
8 undisturbed 8 undisturbed
long particle model long particle model
y (cm)
y (cm)
6 6
DNS DNS
4 4
2 2
0 0
0 10 20 30 40 50 60 0 10 20 30 40 50 60
Figure XI.4. Velocity profiles at steady state on a line through the center of a particle at y p = 0.75 cm,
R = 20 . (a) reference size ( d = 1.0 cm), (b) small particle ( d = 0.5 cm). As the particle is smaller, the
difference between disturbed and undisturbed velocity profiles is smaller.
60
50
40
Up (cm s -1)
30
DNS
20
long particle model
10
0
0 1 2 3 4 5 6
Wall y c (cm) Centerline
Figure XI.5. Particle velocity vs. particle position.
A comparison of the long particle model with DNS for a circular particle is given in figure
XI.3 and XI.4. In these constrained simulations we fix the y position of the particle and compute
the dynamic evolution to equilibrium at R = 20 . The diameter of the particle in figure XI.3 is 1
cm and in figure XI.4 it is 0.5 cm. The profiles in the figures are at equilibrium and on a cross-
section through the center of the particle. The agreement is rather better than might have been
anticipated given the severe assumptions required in the model. The agreement is quite good
away from the centerline, even close to the wall. Equations (XI.6-XI.10) can be recommended
for an analytical approximation for the velocity of a circular particle in Poiseuille flow.
In figure XI.5, we compare the particle velocity from the simulation with the long particle
model. In figure XI.6, we compare the slip velocity, and in Figure XI.7 we show how nearly the
particle angular velocity is given by − 12 γ (y p ) .

0.35
0.3
0.25
Us / u ( yc )
0.2
DNS
0.15
long particle model
0.1
0.05
0
0 1 2 3 4 5 6
Figure XI.6. Slip velocity/fluid velocity ratio vs. particle position at R = 20 . Slip velocity evaluated using
DNS results vs. slip velocity in the long particle model. These are relative values of the slip velocity,
U s , with respect to the fluid velocity on the undisturbed flow at the particle center u ( y p ) . The largest
discrepancy is at about 1
4
the distance from the wall to the centerline.
10
DNS
8
long particle model
6
W p (s -1 )
0
0 1 2 3 4 5 6
Figure XI.7. Particle angular velocity, Ω p at R = 20 . The angular velocity of the particle is
approximated in the long particle model as half the value of the shear rate on the undisturbed flow
evaluated at particle’s center position Ω p LPM = − 12 γ (y p ) .

§ Slip velocities, angular slip velocities and lift for neutrally buoyant circular
particles
Multiple equilibrium solutions do not appear at moderate numbers when ρ p = ρ f , RG = 0 ; the
equilibrium solutions are unique. Figures XI.8, XI.9 and XI.10 show the evolution to
equilibrium, at R = 10 , of a neutrally buoyant particle started at the wall and at the centerline
from an initial condition of rest. No matter where the particle is released it will migrate to a
unique equilibrium solution at y e = 4.18 cm .
4
y p (cm)
3
starting at the centerline
2
starting at the wall
1
0
0 20 40 60 80 100 120
t (s)
Figure XI.8. Migration of a neutrally buoyant particle in an unconstrained simulation at R = 10 .
30
25
20
Us (cm s -1 )

15
10
0
0 20 40 60 80 100
t (s)
Figure XI.9. Evolution of the slip velocity of the particle whose trajectory is shown in figure XI.8. The
slip velocity evolves to zero through positive values whether the particle is started above or below the
equilibrium position. The slip velocity discrepancy U s − U se > 0 , U se = 0.148 cm s -1 .

10
W s (s -1)
1
0.1
0.01
0 20 40 60 80 100 120
t (s)
Figure XI.10. Evolution of the angular slip velocity of a neutrally buoyant particle at R = 10 to
equilibrium (see figure XI.8). The angular slip velocity function evolves without crossing the
equilibrium value. When the angular slip velocity is below the equilibrium value, the particle moves
downward. When the angular slip velocity is above the equilibrium value, the particle moves upward.
In figure XI.9, we show the evolution of the slip velocity to equilibrium. The slip velocity is
positive and of course the greatest for a particle released from rest at the centerline.
In figure XI.10, we show that the angular slip velocity is smaller than its equilibrium value
when the particle is above the equilibrium position and, is larger than its equilibrium value when
it is below the equilibrium position. The angular slip velocity discrepancy
Ω s − Ω se
0.05 1
(a) (b)
0
W s (s -1)
-0.05
F y (dyn cm -1 )
-0.1
0.1
-0.15
-0.2 fixed position above equilibrium
fixed position above equilibrium fixed position below equilibrium
-0.25
fixed position below equilibrium
-0.3 0.01
10 30 50 70 90 110 10 30 50 70 90 110
t (s) t (s)
Figure XI.11. (a) Lift and (b) angular slip velocity in constrained simulations of a particle fixed above
and below equilibrium. The sign of the lift correlates perfectly with the sign of the angular slip
velocity discrepancy. R = 10 , y p = 4.16 and 4.19 cm, y e = 4.18 cm .

0.06
0.04
Fy (dyn cm -1 )
0.02
fixed position above centerline
0
fixed position below centerline
-0.02
-0.04
-0.06
0 20 40 60 80 100
t (s)
Figure XI.12. Evolution of the lift on a particle at a fixed position at R = 10 slightly above,
y p = 6.05 cm and below, y p = 5.95 cm , the centerline. The lift pushes the particle away from the
centerline.
0.06
fixed position above centerline
0.04
fixed position below centerline
0.02
W s (s -1 )
-0.02
-0.04
-0.06
0 20 40 60 80 100
t (s)
Figure XI.13. Evolution of the angular slip velocity in a constrained simulation for particles at
y p = 5.95 cm and 6.05 cm when R = 10 . The evolution is to a steady state with the following
properties: when the angular slip velocity is below the equilibrium value, the particle moves upward.
When the angular slip velocity is above the equilibrium value, the particle moves downward. This
behavior is the opposite of the previous cases, because the previous cases were stable equilibrium
positions, and therefore, the force field around them is the opposite.
where Ω se is the angular slip velocity at equilibrium, changes sign with the lift across the
equilibrium.
In figure XI.11, we carry out a constrained simulation in which the circular particle is fixed at a
position slightly above and slightly below the value of equilibrium. This figure shows that the
sign of the angular slip velocity discrepancy changes with the sign of the lift, which is positive
for particles below and negative for particles above the equilibrium.

In figures XI.12 and XI.13 are shown the evolution of the lift and angular slip velocity,
respectively, from constrained simulations at fixed positions slightly above and slightly below
the channel centerline. Figure XI.13, shows that the angular slip velocity discrepancy is negative
when the particle is above the centerline and is positive when it is below the centerline. The
discrepancy changes sign in both the stable and unstable cases, but the sign of the discrepancy is
opposite in the two cases.
In figure XI.14, we have plotted the resultant constrained dynamical simulation comparing
distributions of the normalized slip velocity, the angular slip velocity and lift for a neutrally
buoyant particle are computed at each fixed position y p for R = 20 .
y e (cm)
ρp ρf Starting at Starting close
centerline to wall
1.000 4.560 4.560
1.005 3.834 3.834
1.010 3.165 1.323
Table XI.1. Position of equilibrium when R = 20 .
0.8
L max
L
0.6
,
W s max , Us max
0.4
W s Us
0.2
angular slip velocity
0 slip velocity
lift force
-0.2
0 1 2 3 4 5 6
Wall y p (cm) Centerline
Figure XI.14. Slip velocities and lift for neutrally buoyant particle at R = 20 . Steady-state relative values
for the lift force and the slip velocities. Dotted lines correspond to unstable equilibria. In the region
close to the wall, the lift force and the slip velocity have a similar nonlinear behavior. In the region
close to the centerline, the lift force appears to be proportional to the angular slip velocity.
§ Slip velocities, angular slip velocities and lift for non-neutrally buoyant
circular particles
The qualitative results, which were established in the previous section hold for heavier and
lighter than fluid particles. The slip velocity and angular slip velocity are positive and the angular

slip velocity discrepancy changes sign across y e , where y e is the place where the lift
discrepancy
ℒ = L − 14 (ρ p − ρ f )πd 2 g (XI.23)
in two dimensions vanishes. For heavier than liquid particles, the position of equilibrium moves
closer to the bottom wall, and for values of ρ p ρ f larger than a critical two equilibrium heights
exist (table XI.1).
1.03
1.02
1.01
yp (cm)
1
starting below equilibrium
0.99
starting above equilibrium
0.98
0 2 4 6 8 10
t (s)
Figure XI.15. Particle height about equilibrium for a heavier-than-fluid particle ( ρ p ρ f = 1.01 , R = 10 ).
The initial condition was the steady state solution from a constrained simulation at the initial height.
0.05
0.03
Ws - W se (s -1)
0.01
-0.01
starting below equilibrium
-0.03
starting above equilibrium
-0.05
0 2 4 6 8 10
t (s)
Figure XI.16. Angular slip velocity discrepancy about equilibrium for a heavier than fluid particle
( ρ p ρ f = 1.01 , R = 10 ). The evolution on the angular slip velocity discrepancy is consistent with the
evolution on the particle height.

5
4
y p (cm)
3
0
0 20 40 60 80 100
t (s)
Figure XI.17. Migration from steady flow of a heavy particle ρ p ρ f = 1.01 starting at rest near the wall
and centerline at R = 10 ( RG = 9.8 ). The particle starting at the centerline crosses the equilibrium
position and then moves upward.
Figures XI.15 and XI.16 are plots of the migration to equilibrium of particles starting above
and below but near to equilibrium when R = 10 and RG = 9.8 corresponding ρ p ρ f = 1.01 .
Figure XI.15 shows that the particle migrates to the same equilibrium whether it starts from
above or below the position of equilibrium. Figure XI.16 is a plot of the angular slip velocity
discrepancy versus time; it shows that the angular slip velocity discrepancy changes sign across
the position of the equilibrium.
5
4
3
W s (s-1 )
0
0 20 40 60 80 100
t (s)
Figure XI.18. Evolution of the particle angular slip velocity of a heavier than fluid particle ρ p ρ f = 1.01
starting at rest near the wall and centerline at R = 10 , RG = 9.8 . For the particle starting at the
centerline, the angular slip velocity function crosses the equilibrium value. When the angular slip
velocity is below the equilibrium value, the particle moves downward. When the angular slip velocity
is above the equilibrium value, the particle moves upward. A change in the sign of the angular slip
velocity discrepancy is evident at early times when the particle falling from the centerline crosses the
equilibrium; after this both particles are below the equilibrium and have essentially the same angular
slip velocity.

Figures XI.17 and XI.18 treat the same problem as in figures XI.15 and XI.16 but with a
different initial condition. In figure XI.17, the particle is started from rest near the wall and the
centerline
Figure XI.19 treats the problem of slip velocity and lift for the case R = 20 and RG = 9.8
( ρ p ρ f = 1.01 ) in which there are two stable equilibrium heights. In the dynamical simulation of
a particle started from steady flow near the centerline, the particle migrates downward to the
higher position of equilibrium and the angular velocity discrepancy is negative. When the
particle is started from the wall it migrates upward to the lower position of equilibrium and the
slip angular velocity discrepancy is positive, consistent with our hypothesis about the lift and the
slip angular velocity discrepancy.
In figure XI.20, we consider the case of heavier than fluid particles ρ p ρ f = 1.005 migrating
at R = 20 . For this relatively lightweight particle the equilibrium solutions are unique at R = 20
and the angular slip velocity discrepancy changes with the lift as the particle approaches y p = y e
from above or below.
0.4
starting close to the wall
0.2
W s - W se (s-1)
-0.2
-0.4
-0.6
-0.8
-0.5 0 0.5 1 1.5 2 2.5 3
y c -y ce (cm)
Figure XI.19. Analysis on the angular slip velocity discrepancy in the case of multiple equilibrium
( ρ p ρ f = 1.01 , R = 20 ). Case (a) particle released close to the wall: ( y p (t = 0) = 1.0 cm). Case (b)
particle released at the centerline: ( y p (t = 0) = 6.0 cm). For this data, the buoyant weight intersects
the lift force at three points, and two of them yield stable solutions (figure XI.8, table XI.1). For case
(a), the particle travels from a position close to the wall to the equilibrium height closer to the wall
y e = 1.323 cm, whereas for case (b), the particle travels from the centerline to the equilibrium height
far from the wall y e = 3.165 cm. The discrepancy changes sign at the equilibrium y p = y e where the
lift balances the buoyant weight.

1.2
1
starting close to the wall
0.8
W s - W se (s -1)
0.6
0.4
0.2
0
-0.2
-0.4
-0.6
-3 -2 -1 0 1 2 3
y p - ye (cm)
Figure XI.20. Analysis on the angular slip velocity discrepancy in the case of unique equilibrium
( ρ p ρ f = 1.005 , R = 20 ). Case (a), particle released close to the wall: ( y p (t = 0) = 1.0 cm). Case (b)
particle released at the centerline: ( y p (t = 0) = 6.0 cm). For this data the buoyant weight intersects the
lift force at only one point (figure XI.8, table XI.1). Therefore, no matter where the particle started it
will reach the same equilibrium height at y e = 3.834 cm. The discrepancy is positive if the local value
is greater than the equilibrium value, and it is negative for the opposite condition. The angular slip
velocity discrepancy changes its sign as the particle height discrepancy does. Note: the initial
condition for all the cases was the following: First, to get a fully developed velocity profile, a
simulation at the initial height ( y p (t = 0) = 1.0 cm or y p (t = 0) = 6.0 cm) was performed using a
constrained motion on the vertical direction. Secondly, the vertical motion constrain is released, and
therefore the particle travels to a preferential equilibrium height.
§ Summary
The lift and migration of neutrally buoyant and heavier-than-liquid circular particles in a
plane Poiseuille flow was studied using direct numerical simulation. The study looks at the
relation of slip velocity and angular slip velocity to lift and migration. No matter where the
neutrally bouyant particle is released, it will migrate to a unique equilibrium height and move
forward with a unique steady particle velocity and rotate with unique steady angular velocity.
Neutrally buoyant particles migrate to a radius which can be called the "Segré Silberberg" radius.
This radius is a reference; heavier-than-liquid particles also migrate to an equilibrium radius that
is close to the Segré-Silberberg radius if the particle density is close to the fluid density. The
particles migrate to an equilibrium position y e with shear rate γ e such that the local fluid
rotation − 12 γ e is slightly greater than the particle angular velocity Ω p . The angular slip velocity,
Ω s = Ω p + 12 γ e is always positive but at equilibrium it is very small; Ω p ≈ − 12 γ e can be proposed
as an approximation. The slip velocity at equilibrium U s = U fe − U p is always positive and
slowly varying.
Since the shear rate and slip velocities are one signed they do not explain why the lift
changes sign across the equilibrium radius. We found that the quantity, which does change sign

at y e , is the angular slip velocity discrepancy; the angular slip velocity minus the equilibrium
angular slip velocity Ω s − Ω se . Ω s − Ω se > 0 when y p < y e and Ω s − Ω se < 0 when y p > y e . The
adjustment of the angular velocity of a free particle is very critical to lift. One might think of the
angular velocity discrepancy as a shear flow analogue to the circulation in aerodynamic lift.
We derived a shear version of our long particle model. The long particle model arises when
the circular particle is replaced with a long rectangle of the same diameter as the circle, but so
long that we may neglect end effects. In the shear version we allow the rectangle to shear at the
rate − 12 γ of the local rotation. Using this model we can find explicit expressions for the fluid
rotation in which the velocity on either side of the long particle is matched by the fluid velocity;
then we satisfy the particles equation of motion in which the shear stress force balances the
pressure drop force. This leads to explicit expression for the velocity of the particle (XI.16) and
the slip velocity (XI.20) that is always positive. The shear version of the long particle model is in
good agreement with the results of numerical computation of the motion of a free circular
particle at points of stable equilibrium, both with respect to the particle velocity and the fluid
velocity on the cross-section containing the center of the circular particle.
The results given in this paper are for two dimensions. It remains to be seen how such results
carry over to three dimensions. We note that the celebrated lift formula (X.1) is a two-
dimensional result.

XI Slip velocity and lift at finite Reynolds numbers.....................................................................118

§ Equilibrium positions of neutrally buoyant and heavy particles................................................................. 118
§ Mechanism for lift ...................................................................................................................................... 119
§ Numerical simulation of migration and lift................................................................................................. 120
§ Long particle model.................................................................................................................................... 121
§ Slip velocities, angular slip velocities and lift for neutrally buoyant circular particles .............................. 127
§ Slip velocities, angular slip velocities and lift for non-neutrally buoyant circular particles ....................... 130
§ Summary..................................................................................................................................................... 134
(DISCARD THIS PAGE (only used to generate table of contents)

Stability and turning point bifurcations of a single particle in Poiseuille flow
XII Stability and turning point bifurcations of a single particle in Poiseuille

flow
Choi and Joseph 2001 performed simulations for single particle lift-off in Poiseuille flows at
much higher shear Reynolds numbers. They observed that the rise and other equilibrium
properties are not smooth functions of R. They found the existence of multiple steady states and
hysteresis. Figure XII.1 shows the plot of he/d vs. R at different values of angular velocity of the
particle. The particle density is 1.01 g/cc, W/d = 12, L/d = 22, η = 1 poise, d = 1 cm and ρf = 1
g/cc. The particle is initially placed close to the bottom wall. Simulations were performed in a
periodic channel with three different conditions on the angular motion of the particle: zero
hydrodynamic torque (free rotation), zero angular velocity (Ωp = 0) and zero slip angular
velocity (Ωs = 0). In each of these cases the equilibrium height shows a sharp rise after a critical
shear Reynolds number that is smallest for a non-rotating particle and is largest when the slip
angular velocity is suppressed. The sharp rise or ‘jump’ in the equilibrium height can be
explained in terms of turning point bifurcation to be discussed later. Choi & Joseph 2001
reported the freely rotating case shown in figure XII.1. The angular velocity of the particle is
seen to have little effect on the equilibrium height before the ‘jump.’ The greater the slip angular
velocity, the higher the particle rises after the ‘jump.’ Models for lift should account for this
effect of the slip angular velocity.
4.5
Ws = g /2 + Wp
4
3.5
3
he /d
2.5 Ws = g /2
0 < Ws < g /2
2 Ws = 0
1.5
0.5
5 10 15 20 25 30
R
Figure XII.1. Lift-off of a circular particle from a horizontal wall in a Poiseuille flow of a Newtonian
fluid (W/d = 12, L/d = 22, η = 1.0 poise, d = 1 cm, ρp = 1.01 g/cc).
Printed 03/11/02 136 • Interog-5.doc

1.4
W s = g /2 + Wp Ws = g /2
1.2 0 < Ws < g /2
Ws< 0
1
Us (cm/s)
0.8
0.6
0.4
0.2
0
5 10 15 20 25 30
R
Figure XII.2. Slip velocity vs. shear Reynolds number for the cases depicted in figure XII.1.
Figure XII.2 shows the plot of slip velocity vs. R for the case above. It is seen that the slip
velocity decreases before the ‘jump’ and increases sharply at the ‘jump’. The slip velocity does
not show a consistent trend with respect to the angular velocity of the particle. The slip angular
velocity also shows a sharp change at the ‘jump’ (figure XII.3). As expected, the slip angular
velocity is maximum for a non-rotating particle.

6
W s= g /2+ W p Ws = g /2
5 0 < Ws< g /2
Ws = 0
4
Ws (/s)
-1
5 10 15 20 25 30
R
Figure XII.3. Slip angular velocity vs. shear Reynolds number for the cases depicted in figure XII.1.
In figure XII.4 we plot the rise of a neutrally buoyant particle to the equilibrium height as a
function of time for W/d = 12, L/d = 22, d = 1 cm, η = 1 poise, ρf = 1 g/cc and R = 5.4. The
simulations are performed in a periodic channel. We compare the rise of freely rotating and non-
rotating particles. A neutrally buoyant freely rotating particle rises to a Segré-Silberberg radius;
the non-rotating one rises more. A smaller lift is obtained when the slip angular velocity is
entirely suppressed (Ωs = 0) but the particle does rise. The greater the slip angular velocity the
higher the particle rises.

5
Ws = °g/2 + Wp Ws = °g/2
4 Freely Rotating 0 < Wse <°g/2
Ws = 0
3
R = 5.4
Y = y/d
2
Neutrally Buoyant: rp/rf =1.00
0
0 50 100 150 200
Time (sec)
Figure XII.4. Rise vs. time for a neutrally buoyant particle (R = 5.4, W/d = 12, l/d = 22,
η = 1 poise, d = 1 cm).
Table XII.1 gives data for the results presented in figures XII.1-4. In the next section we
discuss the contribution to the hydrodynamic lift force from pressure and shear stress in a
Newtonian fluid.
To analyze the instability manifested in the jumps in rise heights and particle velocities found
by Choi and Joseph 2001 and shown in table XII.1 we may use another simulation method
introduced by Patankar, Huang, Ko and Joseph 2001. The motion is simulated in a periodic
channel in which the particle is free to rotate and translate in the axial (x-) direction. The height
of the particle center from the bottom wall of the channel is fixed so that it does not translate in
the transverse direction. There is no external body force in the axial direction and no external
torque is applied. Gravity acts in the negative y-direction. The particle is initially at rest and
eventually reaches a state of steady motion.

Table XII.1. Data structure for a freely translating circular particle levitated by Poiseuille flow (W/d =
12, L/d = 22, d = 1 cm, ρf = 1 g/cc and η = 1 poise). Bold numbers are for freely rotating particles.
All the dimensional variables are given in CGS units.
2.9725 9.81 5.40 0.90 0.6020 2.1250 3.0877 0.9627 0.0000 2.4291 2.4291
6.6881 9.81 8.10 1.35 0.8366 5.7950 6.3040 0.5090 0.0000 3.4853 3.4853
14.679 9.81 12.00 2.00 1.1130 11.960 12.117 0.1572 0.0000 4.8870 4.8870
18.578 9.81 13.50 2.25 1.2260 14.870 14.860 0.0100 0.0000 5.3708 5.3708
26.752 9.81 16.20 2.70 3.6200 40.260 40.953 0.6930 0.0000 3.2130 3.2130
39.963 9.81 19.80 3.30 4.0120 52.140 52.879 0.7389 0.0000 3.2802 3.2802
43.679 9.81 20.70 3.45 4.0830 55.070 55.761 0.6908 0.0000 3.3068 3.3068
74.312 9.81 27.00 4.50 4.4410 74.290 75.531 1.2414 0.0000 3.5077 3.5077
2.9725 9.81 5.40 0.90 0.6268 2.2960 3.2079 0.9119 1.5600 2.4180 0.8579
6.6881 9.81 8.10 1.35 0.8923 6.1420 6.6902 0.5482 2.8330 3.4477 0.6147
14.679 9.81 12.00 2.00 1.1100 11.740 12.088 0.3479 4.1220 4.8900 0.7680
18.578 9.81 13.50 2.25 1.1300 13.500 13.818 0.3185 4.6240 5.4787 0.8548
26.752 9.81 16.20 2.70 1.2110 17.410 17.638 0.2284 5.3200 6.4652 1.1452
36.413 9.81 18.90 3.15 1.2760 21.470 21.552 0.0820 6.0080 7.4403 1.4323
39.963 9.81 19.80 3.30 1.2900 22.720 22.796 0.0762 6.2140 7.7715 1.5575
43.679 9.81 20.70 3.45 3.2610 48.590 49.159 0.5688 3.9550 4.7248 0.7698
47.560 9.81 21.60 3.60 3.3800 51.790 52.444 0.6540 3.9490 4.7160 0.7670
74.312 9.81 27.00 4.50 3.8310 69.540 70.415 0.8747 4.1040 4.8803 0.7763
2.9725 9.81 5.40 0.90 0.6511 2.4530 3.3252 0.8722 2.4070 2.4070 0.0000
6.6881 9.81 8.10 1.35 0.9047 6.2120 6.7756 0.5636 3.4390 3.4393 0.0000
26.752 9.81 16.20 2.70 1.1990 17.190 17.483 0.2930 6.4810 6.4814 0.0000
43.679 9.81 20.70 3.45 1.2830 23.600 23.719 0.1186 8.1360 8.1368 0.0000
47.560 9.81 21.60 3.60 1.3080 25.010 25.173 0.1632 8.4460 8.4456 0.0000
52.991 9.81 22.80 3.80 3.2440 53.290 53.968 0.6784 5.2370 5.2364 0.0000
74.312 9.81 27.00 4.50 3.6520 67.790 68.595 0.8054 5.2830 5.2830 0.0000
∞ 0 5.4 0.90 4.9999 15.670 15.749 0.0800 0.0000 0.4500 0.4500
∞ 0 5.4 0.90 3.7530 13.780 13.928 0.1480 0.9580 1.0110 0.0530
∞ 0 5.4 0.90 3.6810 13.630 13.780 0.1500 1.0440 1.0440 0.0000
At steady state the particle translates in the axial direction at a constant velocity and rotates at
a constant angular velocity. At the prescribed height, these velocities are such that there is no net
hydrodynamic drag or torque. The flow field at steady state is independent of the particle density
since the particle acceleration is zero. Only the axial and angular motion equations of the particle
are solved in our simulations. The steady state translational and angular velocities as well as the
hydrodynamic lift force are independent of the particle densities used in our simulations. This
has been confirmed from our numerical results.
The hydrodynamic lift force L on the particle in the transverse direction depends on the
height of the particle and the shear Reynolds number for a Newtonian suspending fluid and given
channel and particle dimensions. We can select a particle of density ρp given by
L
ρp = ρ f + , (XII.1)
Vp g
such that the lift just balances the buoyant weight.

Figure XII.5 shows the plot L as a function of the height of its center at different values of
shear Reynolds number. The suspending fluid is Newtonian, L/d = 22, W/d = 12 and d = 1 cm.
The fluid density is 1 g/cc and its viscosity is 1 poise. This plot can be used to find the
equilibrium height of a particle of given density at different values of R.
200
150
(dyne/cm)
Reynolds number increases

from 10 to 80
100
50
0 1 2 3 4 5 6
h (cm)
Figure XII.5. The hydrodynamic lift force on the particle as a function of the height of its center from the
bottom wall at different shear Reynolds numbers. The bottom wall is h = 0 cm and the channel
centerline is h = 6 cm.
A particle of density ρp will be in equilibrium at a height where L = (ρp - ρf)gVp. As an

example we consider a particle of density 1.01 g/cc. This particle will be in equilibrium when
L = 7.705 dyne/cm. The equilibrium heights at a given shear Reynolds number are identified as
the points of intersection between the curve of L vs. h and L = 7.705 in figure XII.6. The
intersection points where the slope of the L vs. h curve is positive are unstable equilibrium points
whereas a negative slope represents a stable equilibrium point (figure XII.6). Figure XII.7 shows
the plot of equilibrium height of the particle of density 1.01 g/cc vs. R. We reproduce the
bifurcation diagram given by Choi & Joseph 2001. They obtained this diagram by performing
dynamic simulations where the particle was free to move in the transverse direction as well. Our
results are in good agreement with theirs. In fact, we are also able to plot the unstable branch for
the equilibrium height, which was not obtained from the dynamic simulations. From figure XII.7
we identify the nature of instability of the equilibrium height; it may be described as a double
turning point bifurcation. The change of stability at a turning point is not really a bifurcation
because a new branch of solutions does not arise at such a point (see, Iooss & Joseph 1990). The
two turning points give rise to a hysteresis loop depicted in figure XII.7. Similarly, we can plot
the equilibrium height diagrams for particles of different densities using figure XII.5.
Implications of multiple steady states for single particle lifting and on models of lift-off
in slurries should be a subject of future investigation.

20
(dyne/cm) 15
10
= 7.705
0
Stable equilibrium Unstable equilibrium
positions at R = 15 position at R = 15
›5
0 1 2 3 4 5 6
h (cm)
Figure XII.6. Finding the equilibrium height of a particle of a given density at different values of shear
Reynolds number.
4
he
2
Stable branch (our simulations)
Unstable branch (our simulations)
1 Dynamic simulation (Choi & Joseph 2000)
0
0 30 60 90
R
Figure XII.7. Equilibrium height as a function of shear Reynolds number for a particle of density
1.01 g/cm3.

§ Pressure lift and shear lift

Numerical simulation can be used to analyze the forces which enter into the lift balance
Lp + Ls = πd 2 (ρ p − ρ f )g 4 ,
ò
Lp = -pndΓ,
∂P
(XII.8)
ò
Ls = 2ηD[u] ⋅ n dΓ,
∂P
where the buoyant weight is balanced by the sum of the pressure lift Lp and the shear lift Ls. It is
well known that only the tangential (or shear) component of 2ηD[u]•n is non-zero on a rigid
surface. We define lift fractions
Lp
Φp =
Lp + Ls
Ls
Φs = (XII.9)
Lp + L
Φ p + Φs = 1
In figure XII.8 we plot the lift fraction vs. R for cases shown in figure IX.4a; the figure
shows that the pressure lift is greater than the shear lift and that the pressure lift fraction is
greater for heavy particles. Figure XII.9 shows the plot of lift fraction vs. R for cases in figure
XII.1. For a freely rotating particle the pressure lift is higher than the shear lift at lower shear
Reynolds numbers but after the ‘jump’ they are of the same order. A non-rotating particle always
has a greater contribution to lift from pressure.

Fp
0.8
Lift fraction
0.6 R G = 0.981
Lift 0ff
R G = 9.81
0.4 R G = 392.4
0.2
Fs
0
0 10 20 30 40 50 60
R
Figure XII.8. Lift fractions due to the pressure and the viscous shear stress (W/d = 12). The pressure lift
dominates.
1
Viscous shear lift fraction; free rotation
0.9 Viscous shear lift fraction; no rotation
Pressure lift fraction; free rotation
0.8 Pressure lift fraction; no rotation
0.7
Lift fraction
0.6
0.5
0.4
0.3
0.2
0.1
0
5 10 15 20 25
R
Figure XII.9. Lift fraction vs. shear Reynolds number for the cases shown in figure XII.1. Lift fractions
for a freely rotating and a non-rotating particle are shown.

Figure XII.10 shows the pressure and the viscous shear stress distributions around the
particle at different shear Reynolds numbers and particle rotations. The particle velocity lags the
undisturbed fluid velocity (figure XII.11). The curvature of the undisturbed velocity profile
creates a higher velocity of the fluid relative to the particle on the bottom half (figure XII.11).
This was recognized by Feng, Hu and Joseph 1994b. The stronger relative flow on the bottom
half results in a larger viscous shear stress at the bottom i.e. at θ = 180o (figures XII.10a, b, e).
50
q
40 Flow Pressure
Viscous shear stress
30
Pressure and viscous shear stress
Wall
20
10
›10
›20
›30
›40
0 100 200 300
q (deg)
Figure XII.10(a). Distributions of pressure and viscous shear stress on the surface of a freely rotating
circular particle in a Poiseuille flow of a Newtonian fluid. W/d = 12, L/d = 22, d = 1.0 cm, ρp/ρf =
1.01, R = 8.1 (before bifurcation).
50
q
40 Pressure
30 Flow
Pressure and viscous shear stress
20 Wall
10
›10
›20
›30
›40
0 100 200 300
q (deg)
Figure XII.10(b). Distributions of pressure and viscous shear stress on the surface of a freely rotating
1.01, R = 27 (after bifurcation).

50
q
40 Flow
Lift due to pressure
Lift due to viscous shear stress
30
Wall
20
Lift forces
10
›10
Lp = 5.681 dyne/cm
›20
Ls = 2.180 dyne/cm
›30
›40
0 100 200 300
q (deg)
Figure XII.10(c). The distribution of lift forces on the surface of a freely rotating circular particle in a
Poiseuille flow of a Newtonian fluid. W/d = 12, L/d = 22, d = 1.0 cm, ρp/ρf = 1.01, R = 8.1 (before
bifurcation).
50
q
40 Flow
30
Wall
20
Lift forces
10
›10 Lp = 3.649 dyne/cm

Ls = 3.620 dyne/cm
›20 Lift due to pressure
›30
›40
0 100 200 300
q (deg)
Figure XII.10(d). The distribution of lift forces on the surface of a freely rotating circular particle in a
Poiseuille flow of a Newtonian fluid. W/d = 12, L/d = 22, d = 1.0 cm, ρp/ρf = 1.01, R = 27 (after
bifurcation).

50
q
40 Flow Pressure
Pressure and viscous and elastic shear stress 30
Wall
20
10
›10
›20
›30
›40
0 100 200 300
q (deg)
Figure XII.10(e). Distributions of pressure and viscous shear stress on the surface of a non-rotating
1.01, R = 27 (after bifurcation).
50
q
40 Flow Lift due to pressure
30
Lp = 3.993dyne/cm
Wall Ls = 3.855 dyne/cm
20
Lift forces
10
›10
›20
›30
›40
0 100 200 300
q (deg)
Figure XII.10(f). The distribution of lift forces on the surface of a non-rotating circular particle in a
Poiseuille flow of a Newtonian fluid. W/d = 12, L/d = 22, d = 1.0 cm, ρp/ρf = 1.01, R = 27 (after
bifurcation).

Streamlines
Lower velocity is
incident on the top P2
S1
half of the particle.
Ωp
Flow relative
to the particle
P1
Higher velocity is S2
incident on the
bottom half of the
particle.
Undisturbed flow
profile
Figure XII.11. Cartoon depicting the fluid velocity and the streamlines relative to a particle in a plane
Poiseuille flow. The fluid approaches the particle with higher velocity in the bottom half of the
particle. Consequently, the pressure P1 (in bold) is greater than P2 and the viscous shear stress S1 (in
bold) is greater than S2.
The analysis of the action of pressure and shear in levitating heavier than liquid particles may
be carried out in the laboratory frame or in a frame fixed on the particle. In the laboratory frame
we note that the forward motion of the particle pushes the fluid nearer the wall forward, inducing
a return flow or vortex there. The high pressure on the bottom of the front face of the moving
sphere near the stagnation point induces a high lift. This mechanism can also be described in the
particle frame of reference.
Figure XII.11 shows the streamlines around the particle. The fluid velocity incident on the
bottom half gives rise to the high pressure P1 (in the third quadrant) that pushes the particle up.
The incident fluid moves up, as shown by the streamline in figure XII.11, giving rise to the
viscous shear stress S1 at θ = 270o in the upward direction. Similarly, pressure and shear forces,
P2 and S2 respectively, act on the particle due to the velocity incident on the top half as shown in
figure XII.11. Since the incident velocity on the bottom half is more, the lift due to P1 and S1
dominates giving rise to a net upward force on the particle. This is consistent with the
observations in figures XII.10c, d and f. The regions of low pressure on the particle surface are
seen to be less important in determining the lift on the particle as compared to the regions of high
pressure.
The viscous shear stresses near θ = 90o and θ = 270o are smaller for a non-rotating particle.
We see from figure XII.11 that the magnitudes of S1 and S2 would decrease for a non-rotating
particle due to smaller relative velocities between the fluid and the particle surface at θ = 90o
and θ = 270o. The plot of viscous shear stress distribution is therefore shifted in the positive
direction for a non-rotating particle (figure XII.10e) giving a greater lift as compared to a freely
rotating particle at the same equilibrium height; a non-rotating particle is seen to rise more.

XII Stability and turning point bifurcations of a single particle in Poiseuille flow 136
§ Pressure lift and shear lift ........................................................................................................................... 143

Fluidization by lift of 300 circular particles in plane Poiseuille flow by DNS
XIV Fluidization by lift of 300 circular particles in plane Poiseuille flow by

DNS
In fluidized beds and sedimentation columns in which particles are in a balance of buoyant
weight and drag, the cooperative effects of other nearby particles enter strongly into the
dynamics. These effects are described by theories of hindered motion. There is a very definite
relation between the fluidization by drag of one single isolated particle and the fluidization of a
particle in a swarm of other particles; for example, the Richardson-Zaki correlation 1954.
Analogous ideas must come into play in problems of slurries which are fluidized by lift rather
than by drag; hindered motion effects involving the effective viscosity and the effective density
of a suspension and other cooperative effects surely enter here but are not well understood.
Choi and Joseph 2001 carried out studies of fluidization by lift of 300 particles in the same
plane Poiseuille flow used to study the lift to equilibrium of a single particle. The computation is
carried out in a long periodic domain in which the volume fraction of solid circles ranges roughly
between 78 % and 31%. The study is framed as an initial value problem in which a closely
spaced cubic array of particles resting on the bottom of the channel are lifted into suspension.
The following picture emerges from this study. At early times the top of the array is only
slightly disturbed; since the cubic crystal array is not tightly packed the top layers move forward
relative to the bottom. The lifting of particles out of suspension is accomplished by a pressure
mechanism clearly revealed by the simulation; liquid is driven into the bed by high pressure at
the front and low pressure at the back of each circle in the top row. The particles are dislodged
by this pressure mechanism. At higher Reynolds numbers single particles are thrown out of the
bed in a manner resembling saltation. Typically isolated particles will fall back into the bed
because the drag on an isolated particle is less than when it is among many. The permanent
lifting of more particles of the bed takes shape in the formation of waves, which resemble water
waves. The wave amplitude grows as the pressure gradient and flow speed increase; particles are
levitated out of the bed and the levitated particles form a fluidized suspension over a basically
fixed bed. This can be described as bed erosion. It is possible to erode the whole bed and fluidize
all of the particles by lift if the pressure gradient is high enough and the bed depth small enough.
The wave amplitude decreases when the particles are fully fluidized.
The evolution to full fluidization is associated with a transition from a basically vertical
stratification of dynamic pressure to a basically horizontal stratification of dynamic pressure. The
final state of full fluidization is not steady. Internal pressure waves which propagate horizontally
are associated with the propagation of particle depleted regions which could be described as an
internal wave of the volume fraction.
A single heavier than liquid particle will not lift off at low pressure gradients; the particle
slides and rolls on the wall; lift-off to equilibrium occurs at critical values. In the fluidized
suspension we can track the evolution of the rise of the mass center of the particles. The final
state of full fluidization can be determined as the leveling off of the rise of the mass center curve.
The mass center did not rise even after a long computation when the pressure gradient was below
the one critical for the rise of a single particle.
When, as in a slurry, there are many particles the number N and places of N boundaries enter
into the problem description. In the present case, 300 is the number and the places of these 300

particles evolve as part of the solution. We have similarity then for all evolution problems for
which N, d/W, ρp/ρf, R, G are identical. In fact, the density ratio ρp/ρf enters only as the
coefficient of the acceleration terms, hence does not enter for steady flow. Though steady flows
of a single particle do occur, steady flow of many particles apparently do not to occur after lift
off; it is possible that the accelerations are small, however after the bed has fully expanded.
Here we study the problem of lift of 300 solid circles in plane Poiseuille flow (figure IX.2)
when ρp= 1.01 ρf, ρf = 1 gm/cc, W = 12d, d = 1 cm for viscosities η = 1, 0.2 and 0.01 poise
(from light oil to water). The calculations are carried out in a periodic domain L = 63d which is
long enough to contain four or more of the waves of pressure which characterize fully fluidized
beds of 300 particles. Initially the circles are arranged in a cubic crystal array in which the
particles do not touch; they are separated by 0.05 cm when η = 1 and η = 0.2 poise, and by 0.1
cm when η = 0.01 poise. The height of the array is 5.25 cm and 5.35 cm.
We recall that the results of calculations may be generalized by computing the values of the
shear Reynolds number R = γ w d 2 / v and the gravity number G = d (ρ p − ρ f )g / γ wη defined in
(IX.13). The value RG = RG = ρ f d 3 (ρ p − ρ f )g / η 2 = 9.81 / η 2 is independent of γ w . The
running index in our calculation is the pressure gradient p ; given η this determines
R = ρd 2W p / 2η 2 = 6 p / η 2 .
§ Case 1: η = 1 poise, RG = 9.81

Figure XIV.1 shows the height of the center of gravity of the 300 particles as a function of
R = 6 p . We say that the bed height has attained its final fully fluidized value when the rise
curve levels off. Average values of the bed heightH, the average velocityU cm/sec and angular
velocityΩsec-1 at full inflation are given in Table XIV.1. The time taken for the center of gravity
to reach its fully fluidized value increases as the Reynolds number R decreases. Figure XIV.2,
XIV.3, XIV.4 and XIV.5 show snapshots of the evolution of the bed to full fluidization at values
of R, G, given in the caption to figure XIV.1.
Animations for these snapshots can be found at our URL http://www.aem.umn.edu/Solid-
Liquid_Flows. The snapshots are decorated by shades of gray coded to reveal the distribution of
dynamic pressure ( p in equation (IX.6)); dark means low pressure. Figures XIV.6 and XIV.7
give graphs of the pressure at different cross-sections of the channel for R = 120 at an early time
t = 0.9 sec and when the suspension is fully fluidized at t = 27 sec. At early times the pressure is
stratified vertically, but not horizontally; at the later time the pressure is stratified horizontally
and not vertically.

5.5 (a)
5 (b)
4.5
(c)
4
(d)
3.5
(e)
H(t) 3
2.5
2 (a) (b) (c) (d) (e)

1.5 R 120 60 30 16.2 5.4
G 0.08 0.16 0.33 0.61 1.82
1 R/G 1500 375 90.9 26.56 2.97
p 20.0 10.0 5.0 2.7 0.9
0.5
0
0.01 50.01 100.01
Time (sec)
Figure XIV.1. Rise curves for the center of gravity of 300 circular particles fluidized by lift (fluid
viscosity = 1.0 poise, G = 9.81/R). The time scale for the slow rise at p = 0.9 dynes/cm2 has been
compressed by 2; the real time corresponding say, to 50.01 is 100.02 sec. The bed is said to be fully
fluidized when the rise curves level off. The time to full fluidization is longer when the Reynolds
number is smaller.
Table XIV.1 Data for the forward motion of a fluidized suspension of 300 particles after the bed has
fully inflated and the average height H of all particles has stopped increasing ( η = 1.0) . H = H o =
2.65d at t = 0. U and Ω are the average velocity and angular velocity of the particles.
R G p H U Ω
5.4 1.82 0.9 3.01 2.09 0.28
16.2 0.61 2.7 3.54 9.55 1.026
24 0.41 4.0 4.00 18.62 1.75
30 0.33 5.0 4.17 25.63 2.25
60 0.16 10.0 4.69 55.60 3.85
120 0.08 20.0 5.24 119.76 4.90

Figure XIV.2. Snapshots of the fluidization of lift of 300 circular particles ρp = 1.01 g/cm3 when η = 1
poise (R = 5.4, G = 1.82). The flow is from left to right. The gray scale gives the pressure intensity
and dark is for low pressure. At early times particles are wedged out of the top layer by high pressure
at the front and low pressure at the back of each and every circle in the top row. The vertical
stratification of pressure at early times develops into a "periodic" horizontal stratification, a
propagating pressure wave. The final inflated bed has eroded, rather tightly packed at the bottom
with fluidized particles at the top.

Figure XIV.3. Fluidization of 300 particles (R = 16.2, G = 0.61). The conditions are the same as in
figure XIV.2 except that the lift forces are greater leading to a more complete and faster fluidization.

Figure XIV.4. Fluidization of 300 particles (R = 60, G = 0.16). The conditions are as in figure XIV.2.
The ratio R/G = γ w2 d ( g∆ρ / ρ ) measures the ratio of lift to buoyant weight. Here the ratio is very
large leading to fast and complete fluidization; the entire bed has eroded.

Figure XIV.5. Fluidization of 300 particles (R = 120, G = 0.08). The conditions are the same as in
figure XIV.2 and figure XIV.4 but the ratio of lift to buoyant weight is greater and the fluidization is
faster and the particle mass center rises higher than in the previous figures.

In figures XIV.6 and XIV.7 we plotted the distribution of pressures at an early and late time
when R = 120, G = 0.08. The figures show that the vertical stratification of pressure at an early
time evolves to waves of pressure which are associated with propagating number density or
voidage waves.
12
12
11
11
10
(a) 10
9 (a)
(b) 9
8 (b)
(c) 8
(c)
y (cm)
7 7
y (cm)
6 6
5 5
4 4
3 3
2 2
1 1
0 0
-100 -50 0 50 100 150 -100 -50 0 50 100 150
p (dyne/cm2) u (cm/sec)
Figure XIV.6. Distribution of dynamic pressure p and streamwise velocity u at an early time (t = 0.9
sec) when R = 120, G = 0.08 (cf. Figure XIV.5) at different cross sections of the channel. The
dynamic pressure is stratified vertically but not horizontally. The rows of particles slide relative to
one another moving like rigid bands separated by liquid.

12 11
(b) (c) (a) (e) (d)
11
10
10
9
9
8
8
y (cm)
7
7 (a)
y (cm)
6 6
(b)
5 5 (c)
4 4
3 3
2 2
1 1
0
-5000 -4000 -3000 -2000 50 100
Figure XIV.7. Distribution of dynamic pressure p and streamwise velocity u after full fluidization (t =
27 sec) when R = 120, G= 0.08. The pressure p is stratified horizontally and not vertically; pressure
pulses propagate horizontally.
§ Case 2: η= 0.2 poise, RG = 245

Figure XIV.8 shows the height of the center of the gravity of the 300 particles as a function
of R = 6 p / η 2 . The interpretation of figure XIV.8 is basically the same as figure XIV.1. Average
values of the bed heightH cm, velocityU cm/sec, and angular velocityΩ sec-1 at full
fluidization are given in table XIV.2. Snapshots of the evolution to full fluidization are shown in
figures XIV.9 and XIV.10.

Table. XIV.2 Data for the forward motion of a fluidized suspension of 300 particles after the bed has fully
inflated and the average height H of all particles has stopped increasing ( η = 0.2) . H = H o =
2.65d at t = 0. U and Ω are the average velocity and angular velocity of the particles.
R G p H U Ω
45 5.44 0.3 2.64 3.17 0.42
150 1.63 1.0 3.30 10.75 1.25
300 0.82 2.0 3.82 22.98 2.43
450 0.54 3.0 4.75 34.15 2.02
5.5
5 (a)
4.5
(b)
4
3.5
H(t) 3 (c)
(d)
2.5
2 (a) (b) (c) (d)
1.5 R 450 300 150 45

G 0.54 0.82 1.63 5.44
1 R/G 833 366 92 8.27
p 3.0 2.0 1.0 0.3
0.5
0
0.0015937 10.0016 20.0016
Time (sec)
viscosity = 0.2, RG = 9.81/η2 = 245.) p is in dyne/cm2. The bed is fully fluidized when the rise curves
level off. The time to full fluidization is longer when the Reynolds number is smaller. The time to full
fluidization is faster when η = 0.2 than when η = 1 (figure XIV.1). The time is scaled down by 5 and
the center of gravity of the particles will eventually rise when ( R, p ) = (45,0.3) .
The description in the caption of figure XIV.2 applies also to figures XIV.9 and XIV.10
except that the particles are more mobile when the viscosity of the fluid is smaller. The particle
laden region at t = 25 sec in figure XIV.9, and t = 1.98 in figure XIV.10 is separated from the
particle free region by an "interface" which propagates like an interfacial wave. This interface
disappears at a higher R = 450, figure XIV.10 for t > 2.7 sec, because the stronger lift forces push

wave crests into the top of the channel; however, the pressure and associated void fraction wave
persists.
Figure XIV.9. Fluidization of 300 particles (η = 0.2 poise, R = 150, G = 1.63). The final state of the
fluidization at t = 25 sec has not fully eroded. The particles that lift out of the bed can be described as
saltating. A propagating "interfacial" wave is associated with the propagating pressure wave at t =
250.

Figure XIV.10. Fluidization of 300 particles (η = 0.2 poise, R = 450, G =0.54). The flow is from left to
right. The particles can be lifted to the top of the channel.
The pressure wave at t0 = 4.204 sec for the case (η = 0.2, R = 450) in figure XIV.10 is
analyzed in figure XIV.11. The period of this wave is T = 0.56 sec and its wavelength is 16 cm.
The pressure wave is associated with a wave of solids fraction which could also be described as
the passage of internal wave crests and troughs.

Figure XIV.11. Propagation of the pressure wave centered at t0 = 4.204 sec in case η = 0.2 poise, R =
450 shown in figure XIV.10. The period T of this wave is T = 0.56 sec and its wavelength is 16 cm.

12 (c)
11 11
(a) (b) (a) (b) (c)
10 10
9 9
8 8
y (cm)
7 7
y (cm) 6
6
5 5
4 4
3 3
2 2
1 1
0 0
-30 -20 -10 0 0 10 20 30 40 50 60 70
Figure XIV.12. Distribution of dynamic pressure p and streamwise velocity u at t = 0.1 sec when R =
350, η = 0.2 poise. The distribution is basically vertical. This kind of distribution is typical of the
earliest times. The vertical steps show how the fluid supports rows of particles. The particles move
forward together, with only a small velocity.
Figures XIV.12 and XIV.13, like figures XIV.6 and XIV.7, show the evolution of dynamic
pressure from an essentially vertical stratification at early time (t = 0.1 sec) to propagating
horizontal waves in the fully developed suspension at t = 4.95 sec.

11 11
(A) (B) (C)
10 (c) (a) 10
9 9
8 8
(b)
y (cm)
7 7
y (cm)
(a) (b)
6 6
5 5
4 4
3 3
2 2
(A)
1 1 (B)
-500 0 10 20 30 40 50
p (dyne/sec2) u (cm/sec)
Figure XIV.13. Distribution of dynamic pressure p and streamwise velocity u at t = 4.95 when R = 350,
η = 0.2 poise (see figure XIV.10). The pressure and velocity distribution can be associated with
"crest" and "trough" propagating void fractions.
§ Case 3: η = 0.1 poise, RG = 9.81/η

η2 = 9.8×
×103
Figure XIV.4 gives the height of the center of gravity of 300 particles in a liquid with η = 0.1
poise as a function of R = 6 p / η 2 = 600 p . The interpretation of figure XIV.4 is basically the
same as figure XIV.1. Snapshots of the evolution to full fluidization are shown in figures XIV.15
to XIV.21. Notice the overshoot in the bed height at early times. The final bed height is an
increasing function of p (or R). Data for the simulation after the bed settles down to its final
height is shown in table XIV.3.

Table XIV.3. Data for the forward motion of a fluidized suspension of 300 particles after the bed has
fully inflated and the average height H of all particles has stopped increasing (η = 0.1).
H = H 0 = 2.65d at t = 0. U and Ω are the average velocity and angular velocity of the particles.
R G p H U Ω
120 8.18 0.2 2.93 2.82 0.32
180 5.45 0.3 3.07 5.21 0.58
300 3.27 0.5 3.2 8.01 0.84
420 2.34 0.7 3.28 12.43 1.31
600 1.64 1.0 3.53 18.92 1.94
1200 0.82 2.0 3.96 43.94 3.67
1800 0.55 3.0 4.34 72.94 4.17
5.5
4.5 (a)
(b)
4
(c)
3.5
(d) (f)
3 (e) (g)
H(t)
2.5
2 (a) (b) (c) (d) (e) (f) (g)
1.5 R 1800 1200 600 420 300 180 120

G 0.55 0.82 1.64 2.34 3.27 5.45 8.18
1 R/G 3303 1467 366 179 92 33.0 14.7
p 3.0 2.0 1.0 0.7 0.5 0.3 0.2
0.5
0
0 25 50 75 100
Time (sec)
viscosity = 0.1 poise, R = 600 p RG = 9.81/η2 = 9.8×103, p is in dynes/cm2). Notice the overshoot at
early times. The final bed height is an increasing function of p .

Figure XIV.15. Fluidization of 300 particles (η = 0.1 poise, R = 120, G = 8.18). The flow is from left to
right. The bed height rises slightly. The final condition at t = 73 can be described as a partially
eroded bed.

Figure XIV.16. Fluidization of 300 particles (η = 0.1 poise, R = 180, G = 5.45). The bed has eroded
more. At t = 106 sec. The particles at the top are moved by saltation waves.

Figure XIV.17. Fluidization of 300 particles (η = 0.1 poise, R = 300, G = 3.27).

right. The bed height increases with R. At R = 420 the bed has nearly eroded.

Figure XIV.19. Fluidization of 300 particles (η = 0.1 poise, R = 600, G = 1.64). The fluidization of the
particle is accomplished by pressure waves. At R = 600 even the bottom row of particles has eroded.

right. The bed rises at t = 3.5 to the top wall where γ for the flow without particles is negative. The
environment at the top wall is unfavorable to the maintenance of positive fluid or particle rotations
required for lift. The lift then at the top wall must decrease and the bed settle into its final fully
eroded condition.


XIV Fluidization by lift of 300 circular particles in plane Poiseuille flow by DNS ..................163
§ Case 1: η = 1 poise, RG = 9.81................................................................................................................... 164
§ Case 2: η= 0.2 poise, RG = 245................................................................................................................... 171
§ Case 3: η = 0.1 poise, RG = 9.81/η2 = 9.8×103 ........................................................................................... 177

Lift off of a single particle in plane-Poiseuille flow of an Oldroyd-B fluid
XIII Lift off of a single particle in plane-Poiseuille flow of an Oldroyd-B fluid

In this section we study the effect of fluid elasticity on the lift-off of circular particles in
plane Poiseuille flow. There is limited literature on this problem.
Krishnan & Leighton 1995 calculated the lift force on a smooth sphere rotating and
translating in a simple shear flow in contact with a rigid wall. Hu and Joseph 1999 extended their
analysis to second-order fluids. Their results were valid at low Reynolds numbers. The non-
dimensional lift RG for an Oldroyd-B fluid is given by Hu and Joseph 1999
RG =
3
2π
(1.755 RU2 + 0.1365RΩ2 − 0.019 RU RΩ − 4.522 RU R + 0.303RΩ R + 2.314 R 2 )
24 æ λ2 ö 2
+ (
E çç1- ÷÷ RU + 0.25RΩ2 − 0.25RU RΩ ,
5ε è λ1 ø
) (XIII.1)
ρ f U pd ρ f Ωpd 2
RU = and RΩ = .
η η
The above expression is valid in the limit of slow and slowly varying flows so that the second-
order fluid expansion is valid. For a freely translating and rotating sphere, RU and RΩ are
functions of R and the gap size. The above calculations were performed for semi-infinite
domains. Hence W/d is not a parameter of the problem.
The expression for the Newtonian case is obtained by substituting E = 0 in equation (XIII.1).
The resulting expression is valid in the limit of zero gap size. A heavy particle freely translating
and rotating in contact with a plane wall in simple shear flow of a Newtonian fluid is lifted from
the wall and suspended in the fluid if the shear Reynolds number R is greater than a critical
value. After the critical shear Reynolds number the particle rises from the wall to an equilibrium
height at which the buoyant weight just balances the upward thrust from the hydrodynamic force.
In a Newtonian fluid the case of zero separation distance corresponds to an infinite drag force
due to the logarithmic singularities in the lubrication equations for drag and torque. This results
in zero translational and rotational velocities of the particle Krishnan & Leighton 1995. For a
particle in a viscoelastic fluid the elastic component of the lift force is also singular when the gap
between the sphere and the wall approaches zero. This is an important qualitative feature that
differentiates the lift force in a Newtonian and a viscoelastic fluid.
N. Patankar, Huang, Ko and Joseph 2001 studied the lift-off of a single particle in Newtonian
and viscoelastic fluids by direct numerical simulation. They considered a particle heavier than
the fluid driven forward on the bottom of a channel by a plane Poiseuille flow. After a certain
critical Reynolds number the particle rises from the wall to an equilibrium height at which the
buoyant weight just balances the upward thrust from the hydrodynamic force. A correlation for
the critical shear Reynolds number for lift-off was obtained for the particle in Newtonian fluid.
Ko, Patankar and Joseph 2001 obtained lift-off correlations for a circular particle in a
Poiseuille flow.
The set up is the same as that studied in Chapter IX and depicted by the cartoon in figure
IX.2. The only difference is the dimensionless stress T is governed by Oldroyd-B rate equation

∇ é λ ∇ù
T + De T = ê A + 2 De A ú , A = 2D[u] (XIII.2)
ë λ1 û
where λ2 is the retardation time, λ1 is the relaxation time and

o W
De = λ1 γ w = λ p (XIII.3)
12η
is the Deborah number. The dimensionless scales are those introduced just below (XI.9) and the
original work is by Patankar, Huang, Ko and Joseph 2001. The fluid and solid equations of
motions are given by (IX.10) with ∇2u replaced by ∇•T and {-p1 + 2ηD[u]} replaced by
{-p1 + T}. Now there are six dimensionless groups
ρ p 2d λ
, , R, G, De, 1 (XIII.4)
ρf W λ2
replacing the four given by (IX.13). For steady flow ρp/ρf is not an independent parameter. It is
convenient to replace De and G with
RG = RG, E = De/R = λ1η/ρfd2 (XIII.5)
The gravity Reynolds number RG and elasticity number E are both independent of γ w .
Shear thinning effects on lift-off are not reported here but they were considered in the cross-
stream migration studies summarized in figure V.5. The effects of shear thinning may be
addressed by replacing the constant viscosity in the dimensional form of the Oldroyd-B rate law
with η (γ ) where
o
η æç γ ö÷ − η ∞ n −1
è ø é æ o 2ù 2
ö
= ê1 + ç λ3 γ ÷ ú (XIII.6)
η0 − η∞ êë è ø ûú
is expressed by the Bird-Carreau model, η0 = η (0) is the zero shear value, η∞ = η(∞) is the
plateau value and
o 1
γ = tri A 2 , A = 2D[u]
2
is an invariant form which reduces to γ = du / dy pure shear. In forming dimensionless

equations, η0 is used as the scale for viscosity.
§ Neutrally buoyant particle

Figure XIII.1 shows the equilibrium height vs. Deborah number for a neutrally buoyant
particle in an Oldroyd-B fluid. The parameters are as specified in the figure. A freely rotating
neutrally buoyant particle migrates to an equilibrium radius between the channel centerline and

the wall as in the celebrated experiments of Segré & Silberberg 1961, 1962 with the caveat that
the equilibrium radius here depends on the elasticity parameter. The particle rises more as the
fluid elasticity is increased and non-rotating particles rise even more. In fact, at high enough
Deborah numbers a non-rotating particle migrates all the way to the center of the channel; the
centerline is then a stable position of equilibrium and the Segré-Silberberg effect does not occur.
5.5
Wp= 0
5
h e /d
4.5
4
Wp = 0
3.5
3
0 0.2 0.4 0.6 0.8 1
E = De /R
Figure XIII.1 Lift-off of a circular particle from a horizontal wall in a Poiseuille flow of an Oldroyd-B
fluid (W/d = 12, l/d = 22, η = 1.0 poise, d = 1.0 cm, R = 0.60).

Particle position perturbed

from the channel centerline
V1
Channel centerline
Flow relative
to the particle
V2
V2 > V1 Undisturbed flow

profile
Figure XIII.2 Cartoon showing the fluid velocity relative to a non-rotating particle perturbed from the
channel centerline in a plane Poiseuille flow.
Figure XIII.2 shows the flow relative to a non-rotating particle whose position is displaced
from the channel centerline. Relative to the particle, the flow comes from the left. In a
Newtonian fluid the effect of inertia is to reduce the average pressure on the bottom part of the
channel since V2 > V1. As a result the particle is sucked further away from the channel centerline
making it an unstable equilibrium position Feng et al. 1994. Joseph 1996 and Joseph & Feng
1996 argued that the elasticity of the fluid gives rise to a compressive stress that is proportional
to the square of the shear-rate on the particle surface. This compressive stress acts through the
pressure and would push the particle towards the channel centerline. At high enough Deborah
numbers this effect would dominate making the channel centerline the stable equilibrium
position. Our simulation results agree with this prediction.
Figure XIII.3 compares the effect of shear Reynolds number on the equilibrium height of a
particle in a Newtonian and an Oldroyd-B fluid. The critical shear Reynolds number for lift-off
in an Oldroyd-B fluid is smaller than that in a Newtonian fluid. The particle rises more at higher
shear Reynolds numbers. At a given shear Reynolds number the particle rises more in an
Oldroyd-B fluid. The fluid elasticity is seen to enhance the lift on a particle. Figure XIII.4 shows
the equilibrium height vs. Deborah number at a fixed R for a heavy particle. A non-rotating
particle rises more. In general the lift is seen to be greater at higher Deborah numbers.

0.8
W/d = 12, rp /rf = 1.01

0.75
De = 0
0.7 De = 0.25
he (cm)
0.65
Lift off
0.6
0.55
0.5
0 1 2 3 4 5 6 7
R
Figure XIII.3 Equilibrium height vs. shear Reynolds number for a particle in a Poiseuille flow of an
Oldroyd-B fluid.
0.59
0.58 Wp= 0
0.57
0.56
he /d
0.55
0.54 Wp = 0
0.53
0.52
0.51
0 0.2 0.4 0.6 0.8 1
E = De /Re
Figure XIII.4 Lift-ff of a circular particle from a horizontal wall in a Poiseuille flow of an Oldroyd-B
fluid (W/d = 12, l/d = 22, η = 1.0 poise, d = 1.0 cm, ρp = 1.001 g/cc).

§ Pressure shear and viscoelastic lift forces

The stress at any point in an Oldroyd-B fluid can be decomposed as T = -pI+ηA+ττe, where
τe is the elastic stress. The elastic component of lift Le on a particle is given by
Le = ò τ e ⋅ ndΓ . (XIII.7)
∂P
Only the tangential (or shear) component of τe⋅n on a rigid surface is non-zero for an Oldroyd-B
fluid Huang, Hu & Joseph 1998 and Patankar 1997. The lift fractions are defined as
Lp
Φp = ,
L p + Ls + Le
Ls
Φs = ,
L p + Ls + Le (XIII.8)
Le
Φe = ,
L p + Ls + Le
Φ p + Φs + Φe = 1.
Figure XIII.5 shows the lift fractions vs. shear Reynolds number for the cases shown in figure
XIII.3. Maximum contribution to the lift force on a particle in an Oldroyd-B fluid comes from
the pressure whereas the elastic stress makes the least contribution. The pressure lift fraction in
an Oldroyd-B fluid is typically larger than that in a Newtonian fluid.

Fp , De = 0.25
0.8
Fp , De = 0
0.6
Lift fraction
Lift off
0.4
W/d = 12, rp /rf = 1.01 Fs , De = 0
0.2
Fs , De = 0.25
Fe , De = 0.25
0
-0.2
0 1 2 3 4 5 6
R
Figure XIII.5. Lift fractions vs. shear Reynolds number for a particle in a Poiseuille flow of an Oldroyd-
B fluid.
Figures XIII.6a-6d show that a freely moving particle in a Newtonian fluid does not lift-off at
the given parameters whereas it lifts off in an Oldroyd-B fluid under similar conditions. The
dominant contribution to the lift force comes from the high pressure in the third quadrant (figure
XIII.6d). The additional upward thrust on the particle in the Oldroyd-B fluid comes from the
pressure in the bottom half of the particle, where the shear-rate is larger (figure XIII.6d) – in
agreement with the argument of Joseph 1996 and Joseph & Feng 1996. Figures XIII.6e and 6f
show that the contribution to lift from the viscous shear stress is more for a non-rotating particle
than for a freely rotating one. The contribution from the pressure is still dominant.

30
q
20 Flow
Press ure
Press ure and viscous shear stress Viscous shear stress
10 Wa ll
›10
›20
›30
0 100 200 300

q (deg)
Figure XIII.6(a). Distributions of pressure and viscous shear stress on the surface of a freely rotating
circular particle in a Poiseuille flow of a Newtonian fluid. W/d = 12, L/d = 22, d = 1.0 cm,
ρp/ρf = 1.001, R = 0.6. The particle does not lift off.
30
q
Press ure
Press ure and viscous and elastic shear stresses
20 Flow Viscous shear stress

Elastic shear stress
10 Wa ll
›10
›20
›30
0 100 200 300

q (deg)
Figure XIII.6(b). Distributions of pressure and viscous and elastic shear stresses on the surface of a
freely rotating circular particle in a Poiseuille flow of a Newtonian fluid. W/d = 12, L/d = 22,
d = 1.0 cm, ρp /ρf =1.001, R = 0.6, E = 0.5. The particle lifts off.

30
q
20 Flow
10 Wall
Lift forces
-10
-20
-30
0 100 200 300
q (deg)
Figure XIII.6(c). The distribution of lift forces on the surface of a freely rotating particle. W/d = 12,
L/d = 22, d = 1.0 cm, ρp /ρf =1.001, R = 0.6, Newtonian fluid. The particle does not lift off.
30
q
20 Flow Lp = 0.7425 dyne/cm
Ls = 0.0954 dyne/cm
Le =›0.1020 dyne/cm
10 Wall
Lift forces
›10
›20 Lift due to elastic shear stress
›30
0 100 200 300
q (deg)
Figure XIII.6(d). The distribution of lift forces on the surface of a freely rotating particle. W/d = 12,
L/d = 22, d = 1.0 cm, ρp /ρf =1.001, R = 0.6, E = 0.5. The particle lifts off.

8
Press ure
Press ure and viscous and elastic shear stresses

Elastic shear stress
4
›4
›8
0 100 200 300

q (deg)
Figure XIII.6(e). Distributions of pressure and viscous and elastic shear stresses on the surface of a
lifted circular particle in a Poiseuille flow of an Oldroyd-B fluid. W/d = 12, L/d = 22, d = 1.0 cm
ρp/ρf = 1.001, R = 0.6, E = 0.5, Ωp = 0.
8
Lift due to viscous shear stress L p = 0.6402 dyne/ cm
Lift due to elastic shear stress L s = 0.1535 dyne/ cm
L e = 0.0020 dyne/ cm
4
Lift forces
›4
›8
0 100 200 300

q (deg)
Figure XIII.6(f). The distribution of lift forces on the surface of a lifted circular particle in an
Oldroyd-B fluid. W/d = 12, L/d = 22, d = 1.0 cm, ρp /ρf =1.001, R = 0.6, E = 0.5, Ωp = 0.
§ Power law correlations for lift-off in an Oldroyd B fluid.

Correlations like those shown in figure IX.5 were obtained for Oldroyd B fluids by Ko,
Patankar and Joseph 2001. The critical lift-off Reynolds number is the minimum shear Reynolds

number required to lift a particle to an equilibrium height greater than 0.501d; 0.001d can be
viewed as surface roughness with a dimensionless gap size ε = δ/d. Figure IX.5 shows that
channel width-particle diameter ratio W/d is close to asymptotic large W/d limit when W/d > 12.
In the simulations reported below, W/d = 12.
Figures XIII.7a and XIII.7b show the plot of RG vs. the critical shear Reynolds number for
lift-off at different values of the elasticity number and λ2/λ1, respectively. It is seen that larger R
is required to lift a heavier particle. The fluid elasticity enhances the lift on the circular particle.
The data from the simulations can be represented by a power law equation given by RG = aRn,
where the values of a and n are given in the figures. We observe that the slopes, n, for a
Newtonian and an Oldroyd-B fluid are different. The slope does not vary significantly for
Oldroyd-B fluids with non-zero elasticity numbers and when λ2/λ1 ≠ 1. The value of a changes
as E and λ2/λ1 changes. This is in agreement with equation (XIII.1).
At low Reynolds numbers (XIII.1) predicts that a depends on E(1-λ2/λ1) in a three-

dimensional case. Figure XIII.8 shows the plot of a vs. E(1-λ2/λ1). We observe an almost linear
variation of a with respect to E(1-λ2/λ1); in qualitative agreement with the prediction for a three-
dimensional case in equation (XIII.1).
In the above simulations the critical shear Reynolds number for lift-off is defined for a given
equilibrium height corresponding to ε = 0.001. Equation (XIII.1) predicts that in general RG is
also a function of the gap size. Figures XIII.9a and XIII.9b show the plot of RG vs. the shear
Reynolds number for a Newtonian and an Oldroyd-B fluid, respectively. The parameters are as
defined in the figure. The slope n does not significantly change with the gap size whereas the
value of a does depend on ε. Figure XIII.10 compares the variation of a vs. ε for the given
parameters for a Newtonian and an Oldroyd-B fluid. It is seen that the value of a increases
rapidly as the gap size tends to zero in an Oldroyd-B fluid suggesting that the lift force is
singular in this limit, in qualitative agreement with the theoretical predictions of equation
(XIII.1). Such behavior is not observed for the Newtonian case.

10000
Newtonian fluid : E =0.00
Oldroyd-B fluid : E = 0.05
1000 Oldroyd-B fluid : E = 0.1
Oldroyd-B fluid : E = 0.2
100
RG
10
R G = aR n
E = 0.00 : a = 2.3648, n = 1.3904
1 E = 0.05 : a = 5.8040, n = 2.1356
E = 0.1 : a = 9.9508, n = 2.1334
E = 0.2 : a = 17.221, n = 2.1064
0.1
0.1 1 10 100 1000
R
Figure XIII.7(a). The plot of RG vs. the shear Reynolds number R for lift-off on a logarithmic scale for
a Newtonian and an Oldroyd-B fluid at different elasticity numbers. (W/d = 12, ε = 0.001, λ2/ λ1 =
0.125)
10000
Newtonian fluid : λ 2 / λ 1 = 1.0
Oldroyd-B fluid : λ 2 / λ 1 = 0.5
Oldroyd-B fluid : λ 2 / λ 1 = 0.125
1000
Oldroyd-B fluid : λ / λ = 0.0
2 1
100
RG
10
R G = aR n
λ 2 / λ 1 = 1.0 : a = 2.3648, n = 1.3904
1 λ 2 / λ 1 = 0.5 : a = 3.7137, n = 2.0805
λ 2 / λ 1 = 0.125 : a = 5.8040, n = 2.1356
λ 2 / λ 1 = 0.0 : a = 6.4876, n = 2.1432
0.1
0.1 1 10 100 1000
R
Figure XIII.7(b). The plot of RG vs. the shear Reynolds number R for lift-off on a logarithmic scale for
a Newtonian and an Oldroyd-B fluid at different values of l2/l1. (W/d = 12, ε = 0.001, E = 0.05)

20
15
10
a
a = mE(1- λ 2 / λ 1 ) + c
5 m = 87.049, c = 2.0554
0
0 0.05 0.1 0.15 0.2
E(1- λ 2 / λ 1 )
Figure XIII.8. The plot of a vs. E(1-λ2 /λ 1). (W/d = 12, ε = 0.001)
1000
ε = 0.001
ε = 0.005
ε = 0.010
ε = 0.015
ε = 0.022
100 R G = aR n
ε = 0.001 : a = 2.3648, n = 1.3904
RG
ε = 0.005 : a = 2.1705, n = 1.3742

ε = 0.010 : a = 1.8668, n = 1.3862
ε = 0.015 : a = 1.7141, n = 1.3835
ε = 0.022 : a = 1.6207, n = 1.3752
10
10 R 100
Figure XIII.9(a). The plot of RG vs. the critical shear Reynolds number R for lift-off on a logarithmic
scale for a Newtonian fluid for various gap sizes (W/d = 12).

1000
n
R G = aR
ε = 0.001 : a = 5.8040, n = 2.1356
ε = 0.005 : a = 2.9164, n = 2.0674
100
ε = 0.010 : a = 2.0402, n = 2.0306
ε = 0.015 : a = 1.6724, n = 2.0066
ε = 0.020 : a = 1.4654, n = 1.9918
RG
10
ε = 0.001
ε = 0.005
1 ε = 0.010
ε = 0.015
ε = 0.020
0.1
0.1 1 R 10
Figure XIII.9(b). The plot of RG vs. the critical shear Reynolds number R for lift-off on a logarithmic
scale at Oldroyd-B fluid for various gap sizes. (W/d = 12, E = 0.05, λ2/ λ1 = 0.125)
8
Oldroyd-B fluid : E = 0.05, λ 2 / λ 1 = 0.125
Newtonian fluid : E = 0.0 and/or λ 2 / λ 1 = 1.0
6
-0.4615
a = 0.2435 ε
4
a
2
a = 1555.4 ε - 72.31 ε + 2.4541
0
0 0.005 0.01 0.015 0.02 0.025
ε
Figure XIII.10. The comparison of the variation of a vs. ε for a Newtonian and an Oldroyd-B fluid with
λ2/ λ1 = 0.125 and E = 0.05 (W/d = 12).

XIII Lift off of a single particle in plane-Poiseuille flow of an Oldroyd-B fluid.......................149

§ Neutrally buoyant particle .......................................................................................................................... 150
§ Pressure shear and viscoelastic lift forces................................................................................................... 154
§ Power law correlations for lift-off in an Oldroyd B fluid. .......................................................................... 158

Fluidization by lift of 300 circular particles in
plane Poiseuille flow by DNS
XIV Fluidization by lift of 300 circular particles in

DISCARD THIS PAGE (only used to generate table of contents)
XIV Fluidization by lift of 300 circular particles in plane Poiseuille flow by DNS .................185
§ Case 4: : η = 0.05 poise, RG = 9.81/η2 = 3.92×104 ..................................................................................... 186
§ Case 5: η = 0.01 poise, RG = 9.81/η2 = 9.8×104 ......................................................................................... 194
§ An engineering correlation for the lift-off of circular particle in
a plane Poiseuille flow of a Newtonian fluid.............................................................................................. 199
§ Inertial mechanism of fluidization .............................................................................................................. 202
à
page breakà

§ Case 4: : η = 0.05 poise, RG = 9.81/η

η2 = 3.92×
×104
Figure XIV.22 gives the height of the center of gravity of 300 particles in a liquid with
η = 0.05 poise as a function of R = 6 p / η 2 = 2400 p . Snapshots of the evolution to full
fluidization are shown in figures XIV.23 to XIV.28. Notice the overshoot in the bed height at
early times. The final bed height is an increasing function of p (or R). Data for the simulation
after the bed settles down to its final height is shown in table XIV.4.
fully inflated and the average height H of all particles has stopped increasing (η = 0.05).
H = H 0 = 2.65d at t = 0. U and Ω are the average velocity and angular velocity of the particles.
R G p H U Ω
240 16.35 0.1 2.83 1.81 0.18
480 8.18 0.2 3.03 4.51 0.46
960 4.09 0.4 3.19 8.67 0.83
1200 3.27 0.5 3.25 10.84 1.07
1680 2.34 0.7 3.41 18.63 1.78
2400 1.64 1.0 3.60 27.13 2.56

5.5
4.5
4
(a) (b)
3.5 (c)
(e) (f)
H(t) 3
(d)
2.5
2 (a) (b) (c) (d) (e) (f)
1.5 R 2400 1680 1200 960 480 240

G 1.64 2.34 3.27 4.09 8.18 16.35
1 R/G 1467 719 366 235 59 14.7
p 1.0 0.7 0.5 0.4 0.2 0.1
0.5
0
0 50 100
Time (sec)
viscosity = 0.05 poise, RG = 9.81/η2 =3924, p is in dynes/cm2). Notice the overshoot at early times.
The final bed height is an increasing function of p . The bed is fully fluidized when the rise curves
level off. The time to full fluidization is longer when the Reynolds number is smaller.

Figure XIV.23. Fluidization of 300 particles (η = 0.05 poise, R = 240, G = 16.35). Flow is from left to
right. The qualitative features of fluidization by lift are as in the case η = 0.1 (see figure XIV.39).





Figure XIV.28. Fluidization of 300 particles (η = 0.05 poise, R = 2400, G = 1.64). The flow is from left
to right. The bed rises to the top of the channel where the environment for positive circulation
required for lift is unfavorable. The terminal flow is fully eroded and wavy.

§ Case 5: η = 0.01 poise, RG = 9.81/η

η2 = 9.8×
×104
Figure XIV.29 gives the height of center of gravity of 300 particles in water as a function of
R = 6 p / η 2 = 6 × 10 4 p . The interpretation of figure XIV.29 is basically the same as figure
XIV.1. Snapshots of the evolution to full fluidization are shown in figures XIV.30 and XIV.31.
The evolution to full fluidization is accomplished by pressure waves. For low Reynolds
numbers the bed expansion is small; the snapshot at t = 116 sec in figure XIV.30 is an example
of bed expansion in water at low R. The bed has expanded by the development of voids and
dislocations near the top of the bed; only a few particles are fluidized. The bed is not severely
eroded.
Fluidization in water at high Reynolds numbers is greatly different. The fully fluidized bed
shown at t = 0.46 sec in figure XIV.31 is completely eroded. The evolution of the fluidized
suspension is driven by a propagating pressure wave, which is in one-to-one correspondence
with the propagation of voids. At t = 0.27 sec, before the bed has fully fluidized, these voids
coincide with wave troughs.
7
6 .5
6
5 .5
5 (a)
4 .5
H(t) 4
3 .5
3 (b)
2 .5
2 (a) (b)
R 60,000 122
1 .5
G 1.63 817
1 R/G 36,810 1.47
0 .5 p 1 0.02
0
0 50 100
Time (sec)
Figure XIV.29. Rise curves for the center of gravity of 300 circular particles fluidized by lift (η = 0.01,
RG = 9.81/ η 2 = 9.81 × 104). p is in dyne/cm2. The time scale for the fast rise at p =1.0 has been
expanded by 100; the real time corresponding, say, to 50 is 0.5sec. The rise to full fluidization is very
rapid and at full fluidization the mass center of the particles is closer to the top than to the bottom
wall. The bed inflation at R = 1200 is modest; at early times the position of the mass center actually
decreases because the circles are more efficiently packed. (hexagonally, rather than cubically
packed at t = 27.)

fully inflated and the average heightH of all particles has stopped increasing (η = 0.01).H =H0 =
2.65d at t = 0.U and Ω are the average velocity and angular velocity of particles.
R G p H U Ω
1200 81.75 0.02 2.77 0.63 0.04
6×104 1.64 1.0 6.02 116.20 4.50
Figure XIV.30. Fluidization of 300 particles (η = 0.01 poise, R = 1200, G = 81.75). The flow is from left
to right. This is a "relatively" heavy suspension with a smaller value of R/G. At t = 27 the flow packs
the initial cubic array more closely into a hexagonal array and the mean bed height drops. The final
fluid condition for t = 116 is mildly inflated, more closely packed at the bottom than the top.

Figure XIV.31. Fluidization of 300 particles in water (η = 0.01 poise, R = 6 × 104, G = 1.64), This is a
"relatively" light suspension with a much larger value of R/G. Particles fluidize easily; the mean bed
height is higher than mid-channel.

12
11 (d) 11
10 10 (a)
9 (b)
9
(c)
8 8
(f) (e)
y 7 (a) (b) y 7
6 (c) 6
5 5
4 4
3 3
2 2
1 1
0
-5000 0 5000 0 100 200 300 400 500
Figure XIV.32. Distribution of dynamic pressure p and streamwise velocity u at t = 0.1 sec for the
fluidization of 300 particles in water when R = 6 × 104 (see figure XIV.31). In this case t = 0.1 sec is
not an early time; the propagating pressure wave has already developed, but it does not yet have a
big effect on the velocity distribution.
Figure XIV.32 and XIV.33 show how the dynamic pressure p develops as the bed evolves to
full fluidization. For this case t = 0.1 sec is not an early time; the periodic pressure pulses which
drive particles into suspension have already developed.
Figure XIV.34 focuses on the wave properties of the evolving fluidization of 300 particles in
water when R = 1200. The t = 40 sec panel of figure XIV.30 shows a propagating and nearly
spatially periodic wave of particles. Wave forms for the dynamic pressure p, the vertical velocity
V and the horizontal velocity U are shown.

12 12
11 11
(b)
10 10 (d)
(c) (c)
9 (a) (b) 9 (a)
(d) y
8 (e) 8
y (e)
7 7
6 6
5 5
4 4
3 3
2 2
1 1
0 50000 -200 -100 0 100 200 300 400

Figure XIV.33. Distribution of dynamic pressure p and streamwise velocity u at t = 0.46 sec (see figure
XIV.31). The pressure and velocity wave coincides with the propagation of internal waves of void
fractions.

-10
p(dyne/cm2)
-20
-30
-40
-50
10 20 30 40 50 60
(a) x(cm)
5
4
Velocity(cm/sec)
u
3
0
v
-1
-2
10 20 30 40 50 60
(b) x (cm)
Figure XIV.34. (a) Distribution of dynamic pressure p, (b) vertical velocity v and horizontal velocity u in
water near the upper wall (y = 11 cm) at t = 40 when R = 1200, G = 81.75, see figure XIV.30). These
spatially periodic waves propagate.
§ An engineering correlation for the lift-off of circular particle in a plane

Poiseuille flow of a Newtonian fluid
Choi and Joseph 2001 simulated the fluidization by lift of 300 circular particles in a plane
Poiseuille flow of a Newtonian fluid by direct numerical simulation. N. Patankar, Ko, Choi and
Joseph 2001 performed similar simulations and reported an engineering correlation for lift-off of
many particles in the Poiseuille flow of a Newtonian fluid.
The parameters of the problem are R, RG, W/d and the average equilibrium height H of the
bed. We take W/d = 12 and l/d = 63 for all the simulations. The average fluid fraction ε of the
fluid-particle mixture depends on the average height of the bed. The effective weight of a particle

in a suspension is equal to (πd2/4)(ρs - ρc)g = ε (πd2/4)( ρs - ρf)g, Foscolo and Gibilaro 1984,
Joseph 1990, where ρc is the effective or composite density of the fluid-particle mixture.
Consequently, the net buoyant weight (or lift) Ls on a particle in a suspension is given by
Ls = εL = ε (πd2/4)( ρs - ρf)g (XIV.1)
where L is the buoyant weight (or lift) on a particle in the absence of any other particles. We
therefore plot εRG vs. R in figure XIV.35 at different values of fluid viscosities. At a given
viscosity the value of RG is constant. The fluid fraction or the average equilibrium height of the
particle bed increased as the Reynolds number is increased. Heavier particles are lifted to a
smaller average equilibrium height at the same Reynolds number. We observe that the data at
each viscosity can be represented by a power law equation of the form RG = cRm, where the
values of c and m are given in the figure. The value of c varies significantly with respect to RG
whereas the value of m does not show large variation. Figure XIV.36 shows the plot of c vs. RG
on a logarithmic scale. The functional dependence of c on RG is represented in terms of a power
law equation. Combining this result with that in figure XIV.35 we can arrive at a correlation for
RG as a function of R and ε. In figure XIV.37 we reduce all the data points to a single curve. The
prefactor in the expression for c is changed from 0.368 in figure XIV.36 to 0.4119 in figure
XIV.37 for better agreement between the data and the correlation. The average value of the
exponent m is also obtained from the curve fit for all the data points. We get the following
correlation for RG as a function of R and ε
RG = 3.27 × 10-4ε-9.05R1.249 (XIV.2)

10000
εRG = cRm
η = 0.05; RG = 3924 : c = 602.86, m = 0.1453
η = 0.1 ; RG = 981 : c = 159.75, m = 0.1655
1000 η = 0.2 ; RG = 245.25 : c = 32.517, m = 0.2364
η = 1.0 ; RG = 9.81 : c = 2.8397, m = 0.175
100
η = 0.05
εRG
poise (our simulations)

η = 0.1 poise (our simulations)
η = 0.2 poise (from Choi & Joseph 2000)
10 η = 1.0 poise (from Choi & Joseph 2000)
1
1 10 100 1000 10000
R
Figure XIV.35. The plot of εRG vs. the shear Reynolds number R on a logarithmic scale for 300 particles
in a plane Poiseuille flow of a Newtonian fluid at different values of fluid viscosities.
1000
100
c
0.8895
10 c = 0.368RG
1
1 10 100 1000 10000
RG
Figure XIV.36. The plot of c vs. RG on a logarithmic scale.

3.5 10
3
(b)
2.5 (a)
εRG
K=
0.4119R G 0.8895
2
K = R 0.138
K
K
K=
εRG
1.5
0.4119R G 0.8895
1
K = R 0.138
0.5
0 1
0 500 1000 1500 2000 2500 3000 1 10 100 1000 10000
R R
R G = 3.27*10 -4 ε -9.05 R 1.249
Figure XIV.37. An engineering correlation for lift-off from numerical simulations of 300 circular
particles in a plane Poiseuille flow of a Newtonian fluid (W/d = 12). (a) Regular scale (b)
Logarithmic scale.
§ Inertial mechanism of fluidization

We studied the fluidization of 300 circular particles in a Poiseuille flow. Initially the particles
are arranged in a cubic array filling nearly half the channel. The flow breaks the cubic array and
inflates the bed by pumping liquid into the bed. The pressures that develop in the bed can levitate
the particles. Bed inflation may be divided into two regimes; an eroded bed in which only the top
rows of the bed have been inflated, and a fully fluidized bed in which all of the particles are
supported by lift forces from the fluid flow. The pumping of liquid into the bed at the earliest
times appears to be a universal inertial effect associated with potential flow around spheres and
circles. This inertial effect produces high pressure at the front and low pressure at the back side
of each circle in the top row of the array. This produces a pressure differential front to back
creating a flow into and out the bed, which dislodges particles from the top row. Further
fluidization is driven by the development of a periodic wave of pressure and number density
which are clearly evident in the snapshots of the bed evolution.
Apart from aspect and density ratios, the dimensionless equations are fully specified by the
values of a shear Reynolds number R = γwd 2 / v and a gravity number G = d ρ p − ρ f g / γ wη ( )
and the number of particles in the cell. The ratio R/G is independent of η and can be viewed as
the ratio of lift to buoyant weight whereas the product RG is independent of γ w and can be
regarded as the ratio of buoyant to viscous damping; you get rapid bouncing around when RG is
large and high lifts, high average height and very inflated beds when R/G is large. A summary of
the average height, velocity and angular velocity of the particles in the fully inflated beds for η =
1, 0.2 and 0.01 poise are presented in tables XIV.1~4.

We have shown (see figures X.3, 4) that even one particle in a Poiseuille flow can reduce the
velocity of the fluid globally; this change is apparently due to the drag of a lagging particle
producing liquid holdup. The reduction of the velocity is all the more greater when there are
many particles. In figure XIV.38 we have used tecplot to plot profiles at 6 different sections of
the flow of 300 particles at p = 20 dyne/cm3, t = 28 sec identified by vertical lines in figure
XIV.5. The open circles give the velocity at fluid points. The straight line segments pass through
particles which are rotating. No slip velocity is seen in such a plot; the velocity is continuous
through the particle.
11 11 11
y 10 10 10
9 9 9
8 8 8
7 7 7
6 6 6
5
4
x=2.2 5
4 x=6.5
5
x=18
4
3 3
3
2 2
2
1 1
1
50 100 0
50 100
streamwise velocity 50 100
11 11 11
10 10 10
9 9 9
8 8 8
7 7 7
6 6 6
5 5 5
4 x=30 4
x=40 4 x=50
3 3 3
2 2 2
1 1 1
50 100 50 100 50 100
Figure XIV.38. Fluid velocity profiles at 6 different sections of the flow of 300 particles at p = 20
dyne/cm3, η = 1.0 and when it is fully fluidized (t = 28).
In figure XIV.39 we plotted the 10 average fluid velocities as a function of y. They are
obtained by averaging fluid velocities at about 1000 points of x coordinate at a fixed y. These
averages are shown as 10 black circles. A scatter plot of particle velocities is shown and a
polynomial fit to this scatter plot is given as a light solid. There does not seem to be a large
difference between the average fluid velocity and the average solid velocity defined in this way,
but a small positive difference between the fluid and particle velocities is vaguely evident. On
the other hand, the difference between the composite fluid-solid velocity and the particle free
Poiseuille flow profile is dramatic.

The data shown in figure XIV.39 suggests that one of the main effects of particles in a fluid
is to radically reduce the velocity of the composite. Each particle produces a drag on the fluid in
a freely moving suspension. The effect of such a distributed drag is equivalent to some form of
effective viscosity. Algorithms for the construction of such effective viscosities would find many
important applications. Unfortunately the empirical forms of effective viscosity functions which
work well for uniform fluidized and sedimenting suspensions work much less well for sheared
suspensions.
12
(a)
11
+
10 ++
+++
+++
+
+ ++
9 +++
++
++++
+
+
++++
8 ++++ ( (b)
+++
+++++
+
++++
+++ ++
y 7 +++++ +
+
+ ++++
+++
++
+
+++++
6 +++ ++++
+
++++++
++++++
+
+++
++++
5 (d)++++++++ (c) (e)
+++++
++++++
++
4 + ++++
+++
++
++++++
+ +
+
+++
3 +++++
+ +++++
++++++
++++++
+ + +
+
+++
2 ++
++++++++
+++ +
++
++++
1 ++++
+++
++++
++++
0
0 50 100 150 200 250 300 350 400
Velocity
Figure XIV.39. Scatter plot of particle velocities (+) of 300 particles taken from figure XIV.38 (η = 1
poise, p = 20 dyne/cm3); (a) average particle velocity for the scatter (  ); (b) average fluid velocity
taken from the tecplots in figure XIV.38 at 10 values of y over 1000 points; (c) two-fluid Poiseuille
flow model based on (XIV.5) and (XIV.6) using the effective viscosity ηm = 3.06 from (XIV.3) with φ =
0.36 and (d) ηm = 4.19 from (XIV.4); (e) undisturbed Poiseuille flow. The particles “hold up” the
fluid. The increased drag on the fluid due to the free particles can be modeled as an effect of an
increased viscosity of the fluid-solid mixture.
For comparison, we calculated velocity profiles using a two-fluid effective viscosity theory.
We suppose that the particle-laden region is an effective fluid with an effective viscosity and an
effective density, though the effective density is not needed for the calculation to follow. The
selection of the region occupied by the fluid-solid mixture is somewhat arbitrary. We suppose
that the mixture has a uniform volume fraction under a flat interface of height 2 H where H is
the mean height. The volume fraction is obtained as 300πd 2 / 8HL where d = 1cm, L = 63cm.
Two highly regarded expressions for the effective viscosity of a uniform suspension of spheres
of volume fraction φ are:
η m = η f /(1 − φ / A) 2 , A=0.638 (XIV.3)

which is due to Kataoka et al. 1978 into two dimension and
η m = η f (1 + 2.5φ + 10.05φ 2 + 0.00273e16.6φ ) (XIV.4)
which is due to Thomas 1965. These expressions were obtained from experiments and may not
apply in two dimensions. We used A = A2 D = 0.8328 which scales A into two dimension in the
0.907
ratio of close packed hexagonal packings in 2 and 3 dimensions A2 D = A3 D . The effective
0.740
theory is a two-fluid stratified Poiseuille flow satisfying the following equations
∂ 2u f ∂ 2um
p =ηf for y ≥ h, p =η for y ≤ h (XIV.5)
∂y 2 ∂y 2
where h = 2 H is the interface between the pure fluid and the mixture. At the boundary
u f (W ) = 0.0, u m (0) = 0.0 and the velocity and shear stress are continuous on y = h;
u f ( h) = u m ( h )
∂u f ∂u m
ηf ( h) = η ( h) . (XIV.6)
∂y ∂y
The comparison of the effective theory corresponding to (XIV.3) and (XIV.4) with the
numerical simulation is exhibited in figure XIV.39. The agreement between the effective theory
and the simulation is far from perfect but there is agreement within a large tolerance. The
effective theory however has several big defects which must be overcome before it can be used
to model the slurry; first and foremost the theory requires that h be specified; here from the
simulation a second problem is that the effective viscosity for uniform suspension need not be a
good representation of sheared suspension. It is certain the effective density of the slurry must
enter into the height of the fluidized slurry in ways we don’t yet understand.
The small difference between the spatially averaged fluid and solid velocities cannot be said
to be clearly evident at all positions on a cross section. It is our position that a positive slip
velocity is required to support the buoyant weight of particles. In fact, we would expect that the
difference between the particle weight and the composite densityρ(φ) = ρpφ + ρf (1- φ) is a
factor in the unknown formula for the lift on a particle in a swarm of volume fraction φ. In the
present case ρp -ρ(φ) = (ρp - ρf )(1- φ) = 0.01(1- φ) is very small and a large slip velocity is not
required to levitate a particle. The calculation of the slip velocity ought to be defined in terms of
time or ensemble averages at a fixed point, which are not readily calculated with the ALE
method used here.

12 (a)
++ + ++ + + + +
++
+ +
+ Scatter plot for particle velocity
+ +
+ + +
11 +
+
+
+
+
+ + ++
++ + + +
+ + + +++ +
+ +
+ ++ + +
10 + + + ++ +
+ + +
++ +
+ +
++
+ ++ + + ++
++ ++
9 + +
++ +
++ +
+ +
++ +
+ +
+ ++ ++
8 + + + + ++
+ +
Average of particle velocity
+ +
+ + + ++
+ ++ + ++ +
+
7 + ++
+ +
++ + ++
y (cm) +
+
+
+ ++ + + +
6
+
+
+ + ++
++ (b)
+
++ +
+ + + +
++
+ + ++ + +++
5 ++
+ + + ++
+
Undisturbed velocity (c)
+
+ + +
+ + ++ +
++ + + +
4 + ++ + +
++ + + + + +
+ +
+ + + ++ ++ + +
+ + +
3 + ++ +
++ + + ++ + + + +
+
++
+ ++
+ + + +
+
+ + +
++
2 + +
+
+ ++ + +++ +
+ ++
+ + + ++
++ + + + + +
++ +
1 + ++
+ + + +
++ ++ +
+ +++++++ + +++ + +
0
0 500 1000 1500 2000
Velocity (cm/sec)
Figure XIV.40. Scatter plot for the fluidization of 300 particles in water (η = 0.01 poise) shown in figure
XIV.31. In this case, particles fill the whole channel and the scatter is caused by large fluctuations.
(a) The average velocity profile (  ) is rather flat as would be expected from turbulent flow. The
black dots • are average fluid velocities from the tecplots in figure XIV.31. (b) Velocity profile for
Poiseuille flow of a fluid with effective viscosity ηm = 0.032 given by (XIV.4) for φ = 0.31. This does
not agree with the simulation. Perhaps it would be better to create an effective "eddy viscosity"
theory for this weakly turbulent flow. (c) Undisturbed Poiseuille flow without particles.
Figure XIV.40 gives a scatter plot for the water flow in figure XIV.31 ( η = 0.01 poise,
R = 6 × 10 4 , G = 164). The results here are what might be expected of turbulent flow and appear
to represent the natural extrapolation of results given in this paper at lower Reynolds numbers.
The relevant Reynolds number, based on the average particle velocity U = 200 cm/ sec, the
effective viscosity ηm = 0.032 poise and particle diameter d, is about 6000, a value at which one
might expect weak turbulence. Informed readers will question the validity of our computation
using Choi’s 2000 split method in direct numerical simulation with no artificial viscosity or
turbulence model. The natural way to test this result is to do mesh refinement. Our unstructured
mesh is generated automatically from nodes on the surface of the circular particle. Our
calculation converged, and converged solutions have nearly the same height history for 15 to 30
nodes on the circular particle when the time step size ( ∆t ) used is about 10 −5 (the corresponding
Courant number is about 0.1); the converged solution cannot be obtained when the number of
nodes is less than 12 or sometimes greater than 30. When the number of nodes is greater than 30,
the calculation may stop because the mesh generator is unable to generate mesh due to the
distortion of very short elements during motion. Certainly we have not established the validity of
such a high Reynolds number computation, but the results do converge to something which
appears reasonable and survives tests of mesh refinement.

XV Bi-power law correlations for sediment transport in pressure driven
channel flows
Direct numerical simulation (DNS) can be used to extract information implicit in the
equations of fluid-particle motion. We have investigated the lift-off of a single particle and many
particles in pressure driven flows by 2D DNS (N.A. Patankar, Huang, Ko & Joseph 2001a, Ko,
N.A. Patankar & Joseph 2001, Joseph, & Ocando 2001, Choi & Joseph 2001, N.A. Patankar, Ko,
Choi & Joseph 2001b). We show that the lift-off of single particles and many particles in
horizontal flows follow laws of similarity, power laws, which may be obtained by plotting
simulation data on log-log plots. Power laws emerge as in the case of Richardson-Zaki
correlations for fluidization by drag. Power laws also emerge from the experimental data from
STIM-LAB (Patankar, Joseph, Wang, Barree, Conway & Asadi 2002 and Wang, Joseph,
Patankar, Conway & Barree 2002). These engineering correlations for lift-off can be used to
predict proppant placement in hydraulic fracturing.
The fracturing industry makes extensive use of numerical simulation schemes based on
models and programmed to run on PC’s to guide field operations. These simulations are used to
predict how the fracture crack opens and closes and how proppant is transported in the crack.
Commercial packages dealing with these problems and propriety packages developed by oil
service companies are used extensively. These numerical schemes solve the average equations
for the fluid and the proppant phases. The solid and the fluid are considered as inter-penetrating
mixtures, which are governed by conservation laws. Interaction between the inter-penetrating
phases is modeled. Models for drag and lift forces on the particles must be used for fluid-
proppant interaction. Models for the drag force on particles in solid-liquid mixtures is a
complicated issue and usually rely on the well-known Richardson-Zaki 1954 correlation. Models
for lift forces in mixtures are much less well developed than models for drag. Therefore, none of
the packages model the all important levitation of proppants by hydrodynamic lift. The power
law models we are developing from DNS and experiments may be incorporated in the model-
based simulation techniques similar to the model for drag.
§ Analogy between fluidization by drag and lift

(a)
(b)
g
g
Uniform fluid flow Shear flow of the fluid

Figure XV.1. (a) Heavy particles fluidized by uniform fluid flow from the bottom of a vertical column. (b) Heavy
particles fluidized by lift due to shear flow of the fluid in a horizontal channel. Gravity acts vertically
downwards.
Fluidization by drag and shear is depicted in the cartoons in figure XV.1. In figure XV.1a the
fluid enters at the bottom of a vertical column at a uniform fluidization velocity. At equilibrium,
the drag exerted by the fluid balances the net buoyant weight of the particles. The particle bed
Last printed 03/11/02 11:50 AM 207 ♦ Interog-9.doc

acquires a height corresponding to the average particle fraction φ. When the fluidizing velocity is
increased the particle bed expands. Richardson & Zaki 1954 did experiments with different
fluids, particles and fluidization velocities. They plotted their data in log-log plots; miraculously
this data fell on straight lines whose slope and intercept could be determined. This showed that
the variables follow power laws; a theoretical explanation for this outstanding result has not been
proposed. After processing the data Richardson & Zaki (RZ) found that
Vφ = V0 [1-φ]n(R0) (XV.1)
where Vφ is the fluidization velocity at the column entrance or the composite velocity. V0 is the
"blow out" velocity, when φ = 0; when Vφ > V0 all the particles are blown out of the bed. Clearly
Vφ < V0 for a fluidized bed. For fluidization columns with large cross section in comparison to
the particle size, the RZ exponent n(R0) depends on the Reynolds number R0 = V0d/ν only, where
d denotes the particle size e.g. diameter of a spherical particle and ν is the kinematic viscosity of
the fluid. The power law in the RZ case is an example of what Barenblatt 1996 calls "incomplete
self similarity" because the power itself depends on the Reynolds number, a third parameter. Pan,
Joseph, Bai, Glowinski & Sarin 2001 carried out 3D DNS of the fluidization of 1204 spheres and
obtained a correlation in agreement with (XV.1). The Richardson-Zaki correlation gives different
expressions for n for different values of R0. In the appendix R.D. Barree presents a way of
representing the various expressions for n by a single continuous function.
(XV.1) describes the complicated dynamics of fluidization by drag. The single particle
fluidization velocity plays a key role in obtaining the fluidization velocity of concentrated
suspensions. An expression for the drag force Fd(1) on a single isolated particle in an infinite
ambient of the fluid is given by a drag law, e.g.
ì 3πηdV0 , laminar
Fd (1) = í (XV.2)
î0.055πρ f d V0 , turbulent
2 2
where η is the fluid dynamic viscosity, ρf is the fluid density and spherical particles are
considered. In a fluidized bed the total force F acting on a particle is (Foscolo & Gibilaro 1984,
Joseph 1990)
F (ε , R0 ) = Fd (ε ) − FB (ε ) , (XV.3)
where ε is the fluid fraction, Fd(ε) is the drag on a single particle in the fluid-particle mixture and
FB(ε) is the effective buoyant weight of a particle in the suspension. We have, FB(ε) = Vp(ρp –
ρc)g = εVp(ρp - ρf)g = εFB(1), where ρp is the particle density, Vp is the volume of the particle, g
is the gravitational acceleration, FB(1) is the buoyant weight of an isolated particle and ρc =
ερf + φρp is the effective or composite density of the fluid-particle mixture. At steady conditions
F (ε , R0 ) = 0,
i.e. Fd (ε ) = FB (ε ) = εFB (1), (XV.4)
i.e. Fd (ε ) = εFd (1).
For spherical particles, (XV.1), (XV.2) and (XV.4) give

[πd 6 ][ρ − ρ ]g = ìïíï0.0553πηπρdVd εV ε
− n ( R0 )
3 φ , laminar
p f 2 2 − 2 n ( R0 )
î f φ , turbulent
or , (XV.5)
ìï18ε −n ( R0 ) Rφ , laminar
RG = í −n ( R ) 2
[ ]
ïî ε 0 Rφ , turbulent
where RG = ρ f [ ρ p − ρ f ]gd 3 η 2 represents the Reynolds number based on the sedimentation

velocity scale VG = [ ρ p − ρ f ]gd 2 η and Rφ = ρ f Vφ d η . (XV.5) is another form of the
correlation for fluidization by drag and can be written as
RG = a(R0 )Rφp ( R0 )ε q ( R0 ) . (XV.6)
Figure XV.1b shows the fluidization of particles by shear flow observed in experiments and
numerical simulations. At equilibrium the average lift exerted by the fluid should balance the net
buoyant weight of the particles. When the applied shear rate is increased the particle bed
expands. This is similar to the fluidization by drag where the mechanism for bed expansion is
different. Correlations analogous to (XV.6) may be expected for fluidization by shear. In that
case a Reynolds number based on the applied shear rate should be defined instead of Rφ. The
prefactor and the exponents may be determined from experimental or numerical data.
§ Direct numerical simulation (DNS) of solid-liquid flows

Choi & Joseph 2001 and N.A. Patankar et. al 2001b used the ALE scheme to study the
fluidization by lift of 300 circular particles in a plane Poiseuille flow by direct numerical
simulation.
Particles are initially placed at the bottom of a periodic channel of height H1 in a close
packed ordered configuration (figure XV.2). The flow is driven by an external pressure gradient.
At steady condition, the particle bed reaches a constant height (figure XV.2). The height of the
clear fluid region above the particle bed is H2. From non-dimensional analysis we get
æ H ö
RG = f ç R,ε , ε , 1 ÷,
è max d ø
where (XV.7)
Nπd 2 Nπd 2
ε = 1- , ε max = 1 - ,
4[H 1 − H 2 ] l 4 H 1l
ρ f γ w d 2
where the shear Reynolds number is R = , the gravity Reynolds number or non-
η
ρ f (ρ p − ρ f )gd 3
dimensional lift is RG = , N is the number of particles, l is the channel length, ε
η2
is the average fluid fraction in the particle bed whose height is (H1 – H2) and εmax is the fluid
fraction in the particle bed if the particles occupy the entire height of the channel i.e. if H2 = 0.

Figure XV.2. (N.A. Patankar, Ko, Choi & Joseph 2001b) Lift-off of 300 heavy particles in a plane pressure driven
flow of a Newtonian fluid, Re = 1800. Contour plot of the horizontal velocity component is shown.
During the simulations εmax and H1/d were constant (N.A. Patankar et al. 2001b, Choi &
Joseph 2001). In that case, RG is a function of R and ε only. N.A. Patankar et al. 2001b obtained
the following correlation
RG = 3.27 ×10 −4 ε −9.05 R1.249

or (XV.8)
−9.05
éε − H 2 ù
ê max H 1ú
RG = 3.27 ×10 −4 ê ú R1.249 .
H
ê 1− 2 H ú
ë 1 û
The correlation above is of the same form as that expected from (XV.6). This shows that
fluidization of slurries by lift also falls into enabling correlations of the RZ type and the above
correlation by N.A. Patankar et al. 2001b could be called a Richardson-Zaki type of correlation
for fluidization by lift. Lift results for fluidized slurries are power laws in appropriate

dimensionless parameters. These power laws are in the form of engineering type correlations; to
use them in applications we need rules for converting two- to three-dimensional results. The goal
of our future work is to generate power laws for engineering applications by processing results of
simulations in 3D just as we have done in 2D.
The DNS results are in agreement with the expected power law form in (XV.6) from the
analogy between fluidization by drag and shear. In the next sections we present the experimental
results for proppant transport to verify the prediction of power laws from DNS (N.A. Patankar, et
al. 2002, Wang, et al. 2002).
§ Experimental setup
Kerns, Perkins, and Wyant 1959 reported the earliest experimental investigation of proppant
transport in narrow slots. STIM-LAB did more experiments to better understand the processes
involved in proppant transport by water and other thin fluids. We have analyzed the data
obtained from their experiments. The apparatus used by STIM-LAB was constructed so that the
transport of proppant in a horizontally oriented slot could be observed. A schematic of the
apparatus is shown in figure XV.3.
Proppant Metering Feeder
Open Standpipe
Pressure Regulator
Fluid Supply Hose Hose and Fluid Exit
See Figure XV.7
Channel Above Proppant

Initial
Proppant Perforations
Emplacement
1 by 8 foot Plexiglas Slot
Proppant Trap
Figure XV. 3. (N.A. Patankar et al. 2002) The experimental setup for proppant transport. Proppant and fluid are
added at the left where they enter over the full height of the slot. Materials exit at the right through
perforations.
Proppant can be added at a constant rate and water flow rate is also constant. Proppant and
water enter the 7.94 mm wide slot through an open end that is 30.5 cm tall. The proppant and
water then move through the 2.44 m length of the slot where they exit via three 8 mm
perforations spaced 7.62 cm apart on the 30.5 cm tall end of the slot. The proppant and water
flow rates were varied, proppants of varying size and density were added and water at different

temperatures was used. Observations were recorded and portions of the experiments were video
taped.
The evolution of the proppant bed in the experiments is well described in figure XV.4. The
portion shown in figure XV.4 is marked in figure XV.3. In the steady state there is an initial
development length (see figure XV.3) followed by a flat bed region shown in figure XV.4 and
marked in figure XV.3. There are three distinct zones in the flat bed region. The bottom part of
the bed is immobile; it is a stationary porous medium that supports liquid throughput that might
be modeled by Darcy's law. Above the immobile bed is a mobile bed in which particles move by
sliding and rolling or advection after suspension or a combination of these modes. Above the
mobile bed is the clear fluid zone. At steady state the volumetric fluid flow rate Qf and the
volumetric proppant flow rate Qp in and out of this region are constant. At steady state, these are
equal to the rate at which the fluid and proppant are injected in the slot.
Clean fluid
H1 H2
Qf Qf
Mobile bed
Qp Qp
Immobile bed
Figure XV.4. (N.A. Patankar et al. 2002) Proppant transport in thin fluids at steady state conditions. In Case 1 only
fluid is pumped, QP = 0, H1 = H2; the particles are immobile. In Case 2 proppants are also injected, QP ≠ 0,
H1 ≠ H2; there is a mobile bed of height H1 - H2. The channel width W = 7.94 mm.
STIM-LAB carried out two types of experiments looking at the transport of proppants in thin
fluids. In Case 1 only fluid is pumped, QP = 0, H1 = H2; the particles are immobile. We call case
1 erosion case. In Case 2 proppants are also injected, QP ≠ 0, H1 ≠ H2; there is a mobile bed of
height H1 - H2. We call case 2 bed load transport case. The channel width W = 7.94 mm. A
simplified description of the experiment is that a bed of proppant is eroded by the flow of water.
When proppant is not injected as in Case 1, the faster the flow of water the deeper is the channel
above the proppants. We are seeking to predict the height above the channel for the given fluid
flow rate. In Case 2, we seek to predict both the clear fluid height as well as the mobile bed
height as functions of Qf and Qp. In the experiments the fluid and the proppant flow rates are
controlled and the heights H1, H2 are measured.
In the DNS of 300 particles reported by Choi & Joseph 2001 and N.A. Patankar et al. 2001b
(figure XV.2), we have a set up similar to that in figure XV.4. The value of H1 in figure XV.4 is
equivalent to the height of the channel in the simulations. In the simulations, data is obtained for
a fixed value of H1/d. This is not the case with the experimental data.

§ Experimental correlations for sediment transport
Dimensionless parameters
The dimensionless parameters in this problem are listed below:
Gravity Reynolds number
ρ f [ ρ p − ρ f ]gd 3
RG = . (XV.9)
η2
Gravity Reynolds number for the fluid
η / ρf 1 ρ f gW
2 3
λ= . ( 2 = ) (XV.10)
W 3/ 2 g λ η2
Note that 1/λ can be viewed as the Reynolds number based on a velocity scale V ~ gW and
length scale W.
Fluid Reynolds number based on channel width
~
ρ f VW ρ f Qf ~ Qf
Rf = = , where V = 2 . (XV.11)
η Wη W
Proppant Reynolds number based on channel width
ρ pV W ρ p Q p Qp
Rp = = , where V = 2 . (XV.12)
η Wη W
Particle diameter/channel width: d/W.

Height of bed/channel width: H/W.
Power law correlations for the erosion case (Patankar et al. 2002)
The erosion case: H1 = H2 = H finds the critical condition of the initial motion of the
proppant. Only fluid is injected in the channel. The particle bed is immobile. There is an
equilibrium value of H corresponding to a given fluid flow rate. When the fluid flow rate is
increased beyond the critical value for a given initial height H, the proppants are eroded from the
bed and washed out until a new equilibrium height H of the clear fluid region above an immobile
bed is achieved for the new flow rate.
Table XV.1 gives the data from these experiments. Figure XV.5 shows a plot of H/W vs. Rf at
different values of RG.

100
RG = 86.838
0.614
H/W = 0.0304R f
H/W 10 RG = 100415.8
RG = 521.37-20342.9
1 H/W = 0.0006R f 0.8672
H/W = 0.001R f 0.9042
0.1
100 1000 10000 100000
Rf
Figure XV.5. Plot of H/W vs. Rf at different values of RG on a logarithmic scale.
Proppants d H η ρf Qf ρp RG ~
V Rf H/W
(cm) (cm) (gm/cm-s) (gm/cc) (cc/s) (gm/cc) (cm/s)
0.034212 1.7 0.01115 0.999 36.778 2.65 521.1645 58.37416 4184.12 2.141732
60/40 0.034212 2.3 0.01115 0.999 58.289 2.65 521.1645 92.51649 6631.36 2.897638
Brady 0.034212 5.6 0.01115 0.999 133.295 2.65 521.1645 211.5662 15164.55 7.055118
0.034212 7.8 0.01115 0.999 232.588 2.65 521.1645 369.1644 26460.80 9.826772
0.056043 2.3 0.01115 0.999 46.556 2.65 2290.822 73.89383 5296.48 2.897638
20/40 0.056043 5.2 0.01115 0.999 133.106 2.65 2290.822 211.2663 15142.90 6.551181
Ottawa 0.056043 8.2 0.01115 0.999 227.542 2.65 2290.822 361.1554 25886.49 10.33071
0.06 1.4 0.01115 0.999 7.885 1.05 86.83778 12.5151 897.05 1.76378
20/40 0.06 2 0.01115 0.999 10.409 1.05 86.83778 16.5212 1184.19 2.519685
Light 0.06 3.9 0.01115 0.999 31.92 1.05 86.83778 50.66353 3631.42 4.913386
Beads 0.06 8.5 0.01115 0.999 128.438 1.05 86.83778 203.8572 14611.89 10.70866
0.06 12 0.01115 0.999 226.217 1.05 86.83778 359.0523 25735.84 15.11811
0.094946 1.5 0.01 0.998 31.542 2.73 14513.72 50.06356 3997.08 1.889764
16/20 0.094946 2.2 0.01 0.998 50.467 2.73 14513.72 80.10138 6395.30 2.771654
Carbolite 0.094946 9.9 0.01 0.998 258.642 2.73 14513.72 410.5174 32775.74 12.47244
0.094946 1.7 0.00378 0.972 36.778 2.73 100415.8 58.37416 12008.41 2.141732
16/20 0.094946 2.3 0.00378 0.972 58.289 2.73 100415.8 92.51649 19031.98 2.897638
Carbolite 0.094946 5.6 0.00378 0.972 133.295 2.73 100415.8 211.5662 43522.25 7.055118
0.094946 7.8 0.00378 0.972 232.588 2.73 100415.8 369.1644 75942.48 9.826772

Proppants d H η ρf Qf ρp RG ~
V Rf H/W
(cm) (cm) (gm/cm-s) (gm/cc) (cc/s) (gm/cc) (cm/s)
(continued)
16/30 0.088437 0.4 0.01115 0.999 10.535 3.45 13363.76 16.72119 1198.53 0.503937
Banrite 0.088437 0.6 0.01115 0.999 13.878 3.45 13363.76 22.02721 1578.85 0.755906
0.088437 1.3 0.01115 0.999 29.145 3.45 13363.76 46.25904 3315.71 1.637795
0.088437 3.5 0.01115 0.999 100.681 3.45 13363.76 159.8012 11454.08 4.409449
0.088437 8.3 0.01115 0.999 261.796 3.45 13363.76 415.5234 29783.50 10.45669
0.109021 1.3 0.01015 0.998 28.955 2.65 20342.9 45.95747 3615.03 1.637795
12/20 0.109021 2.5 0.01015 0.998 62.137 2.65 20342.9 98.62404 7757.81 3.149606
Badger 0.109021 5.8 0.01015 0.998 155.185 2.65 20342.9 246.3101 19374.85 7.307087
0.109021 9 0.01015 0.998 290.814 2.65 20342.9 461.5809 36308.13 11.33858
Table XV.1: Data from experiments on the initiation of sediment motion (Erosion Case).
In the erosion case, three dimensionless parameters Rf, H/W, and RG enter the power law
correlation
H
= a ( RG ) R mf ( RG ) (XV.13)
W
where a and m are function of RG.1 The values of a and m may be regarded as constants for
521.37 ≤ RG ≤ 20342.9 (table XV.2).
RG 86.838 521.37 — 20342.9 100415.8

a 0.0304 0.001 0.0006
m 0.614 0.90432 0.8672
Table XV.2. The prefactor a(RG) and exponent m(RG) in the power law correlations for the erosion case.
Bi-power law correlations for the bed load transport case (Wang et al. 2002)
Bed load transport is another name for the transport of sediments. In bed load transport cases,
both fluids and proppants play important role in determining H1 and H2. Therefore we seek
correlations for H1/W and H2/W in terms of Rf and Rp with the coefficients as functions of RG
1
Shield’s (1936) curve also gives the critical condition for the initiation of sediment motion. The Shields parameter
τ
S is defined as: S = , where τ is a measure of the shear stress on the particle bed. If we take
[ ρ p − ρ f ]gd
~ ηV
~ R f [d ]
τ = ηV , then S = = W . From the Shield’s 1936 curve one obtains (see also, Vanoni
W [ ρ p − ρ f ]gd 2
RG
1975) S = fs ( R f [ d ] ). (XV.13), applicable for proppant transport in narrow channels, has W/H as another
W
parameter. Nothing close to the bi-power law correlations has been put forward for sediment transport.

and λ. To create correlations, we need data and a data structure. An example of the way the data
is structured for processing correlations is given for 20/40 Ottawa in water in table XV.3.
ρf: 1.0gm/cc, ρp: 2.65gm/cc, d: 0.06cm, η: 0.01poise, W: 0.79375cm, RG: 3496.28, λ: 0.000451.
Qp Qf Rf Rp H1 H2 H1/W H2/W
(cc/s) (cc/s) (cm) (cm)
40 244.1 30752.76 13354.33 2.3 0.8 2.897638 1.007874
45.7 242.9 30601.57 15257.32 2.6 0.7 3.275591 0.88189
28.6 250.4 31546.46 9548.346 2.3 1 2.897638 1.259843
11.4 249.8 31470.87 3805.984 2.4 1.5 3.023622 1.889764
11.4 313.5 39496.06 3805.984 3 2.1 3.779528 2.645669
34.3 304.7 38387.4 11451.34 2.9 1.5 3.653543 1.889764
11.4 314.8 39659.84 3805.984 3.1 2.3 3.905512 2.897638
45.7 303.4 38223.62 15257.32 3 1.4 3.779528 1.76378
40 305.3 38462.99 13354.33 3 1.5 3.779528 1.889764
28.6 306 38551.18 9548.346 2.9 1.6 3.653543 2.015748
22.8 306 38551.18 7611.969 2.8 1.7 3.527559 2.141732
17.1 315.4 39735.43 5708.976 3.1 2 3.905512 2.519685
5.7 314.2 39584.25 1902.992 3.5 2.9 4.409449 3.653543
2.9 313.5 39496.06 968.189 4.1 3.6 5.165354 4.535433
1.4 312.9 39420.47 467.4016 5.1 5 6.425197 6.299213
0.4 311.6 39256.69 133.5433 5.8 5.7 7.307087 7.181102
Table XV.3: Experimental data for the bed load transport case with 20/40 Ottawa and water.
We look for correlations in the bi-power law form with five dimensionless parameters
involved:
H1 m ( R ,λ )
= c1 ( RG ) R f 1 G R pn1 ( RG ) ; (XV.14)
W
H2 m ( R ,λ )
= c2 ( RG ) R f 2 G R pn2 ( RG ) . (XV.15)
W
Following are the procedures we used to achieve the bi-power correlations: (1) Different
kinds of proppant and fluid are used in bed load transport experiments and lead to different
values of RG and λ. For each single case, we develop bi-power law correlations of H1 and H2. (2)
The prefactors and exponents in these correlations are functions of RG and λ. We implement
curve-fitting to find analytical expression for these coefficients. (3) Curve fitting implies that c1,
n1, c2 and n2 can be reasonably approximated by logarithmic functions of RG. While the trend of
m1, m2 is less obvious. (4) We use the predicted c1, n1, c2, and n2 by the logarithmic functions of
RG and vary m1, m2 in the bi-power law correlations to match the measured H1 and H2
consistently. The new m1 and m2 turn out to be also logarithmic functions of RG, but with slopes
and intercepts as functions of λ. (5) With the explicit and analytical expressions for all the
coefficients in the bi-power law known: c1(RG), c2(RG), n1(RG), n2(RG), m1(RG, λ), m2(RG, λ) , we
predict H1 and H2 and compare them with the experimentally measured values. The analytical
expressions c1(RG), c2(RG), n1(RG), n2(RG), m1(RG, λ) and m2(RG, λ) can be adjusted to obtain the
best fit for H1 and H2. Hence, we obtain the final form for the analytical expressions: (XV.16) –

(XV.21). Then they are inserted to (XV.14) and (XV.15), giving rise to (XV.22) and (XV.23) as
the final form for the bi-power law correlations.
Next, we present the analytical expressions for the prefactors and exponents in the bi-power
law correlations. These expressions are plotted in figure XV.6–XV.9.
c1 = −0.000230175 ln( RG ) + 0.0029193 ; (XV.16)

c2 = −0.000114966 ln( RG ) + 0.001328763 ; (XV.17)
0.0025
0.002
c1= -0.000230Ln(RG) + 0.002919
c1 and c 2
0.0015 c1
c2
0.001
0.0005
c 2 = -0.000115Ln(RG ) + 0.001329
0
10 100 1000 10000 100000
RG
Figure XV.6. Prefactors c1 and c2 as logarithmic functions of RG.
n1 = −0.01720815 ln( RG ) − 0.12002233 ; (XV.18)

n2 = -0.0071596ln(RG ) - 0.3041344 . (XV.19)
-0.15
-0.2
n 1= -0.0172Ln(RG) - 0.12
n1 and n 2
-0.25
n1
n2
-0.3
-0.35
n 2= -0.00716Ln(RG) - 0.3041
-0.4
10 100 1000 10000 100000
RG
Figure XV.7. Exponents n1 and n2 as logarithmic functions of RG.

m1 = 1.2 − 0.001257 λ−0.42785 [15 .2 − ln( RG )] (XV.20)
1.25
1.2
1.15
1.1 10 cp glycol
m1
5 cp glycol
1.05
water
1 150°F water
0.95
0.9
10 100 1000 10000 100000 100000 1E+07
RG 0
Figure XV.8. Exponent m1 as a logarithmic function of RG with the slopes and intercepts as functions of λ.
m2 = 1.2012713 − 1.295 × 10 -6 λ-1.28089 [11.667387 − ln( RG )] (XV.21)
1.22
1.2
1.18
1.16
10 cp glycol
1.14
m2
5 cp glycol
1.12
water
1.1
150°F water
1.08
1.06
1.04
10 100 1000 10000 100000 1000000
RG
Figure XV.9. Exponent m2 as a logarithmic function of RG with the slopes and intercepts as functions of λ.
From figure XV.6–XV.9, we can see that c1, n1, c2 and n2 can be represented by logarithmic
functions of RG, while m1 and m2 are logarithmic functions of RG with slopes and intercepts as
functions of λ. Equation (XV.20) implies that for any λ, the logarithmic curve m1(RG,
λ = constant) passes through the point (m1=1.2, ln(RG)=15.2). Equation (XV.21) shows that such

a point for m2(RG, λ) is (m2=1.201, ln(RG)=11.67). In figure XV.8 and XV.9, we can see the two
points.
In table XV.4, the bed load transport experiments with the corresponding RG and λ are listed.
Note that the proppant and fluid used in these experiments and their properties can be found in
the table XV.B2. c1, c2, m1, m2, n1 and n2 listed in table XV.4 are predicted by (XV.16)–(XV.21)
corresponding to RG and λ listed in the first and second columns. These c1, c2, m1, m2, n1 and n2
have been plotted in figure XV.6–XV.9, indicated by points.
RG λ c1 c2 n1 n2 m1 m2
27.51 0.00396 0.002156 0.000999 -0.1771 -0.33101 1.04072 1.18938
162.2879 0.00396 0.001751 0.000781 -0.2074 -0.34528 1.064336 1.192695
172.895567 0.00396 0.001733 0.000772 -0.2087 -0.3459 1.065354 1.192838
25.8172784 0.002034 0.002171 0.001007 -0.176 -0.3305 0.987032 1.168948
107.151931 0.002034 0.001843 0.000831 -0.2005 -0.34202 1.012398 1.175391
643.994554 0.002034 0.001431 0.000609 -0.2313 -0.35655 1.044363 1.18351
648.07 0.000452 0.001429 0.000608 -0.2314 -0.3566 0.90401 1.057401
2818.38293 0.000452 0.001091 0.000426 -0.2567 -0.3685 0.953871 1.105684
3496.28 0.000452 0.001041 0.000400 -0.2604 -0.37025 0.961182 1.112763
8903.2 0.000452 0.000826 0.000284 -0.2765 -0.37781 0.992887 1.143466
15921.522 0.000452 0.000692 0.000212 -0.2865 -0.38252 1.012604 1.162559
12755.3536 0.000209 0.000743 0.00024 -0.2827 -0.38073 0.929031 1.070145
73483.2936 0.000209 0.00034 2.31E-05 -0.3128 -0.39491 1.011605 1.228776
Table XV.4: RG and λ for bed load transport experiments and the corresponding c1, c2, m1, m2, n1 and n2 predicted
by (XV.16)–(XV.21).
Inserting the analytical expressions (XV.16)–(XV.21) into (XV.14) and (XV.15), we get the
final form for the bi-power law correlations:
H1 − 0.428
= [−0.000230 ln(RG ) + 0.00292]R1f .2−0.00126λ [15.2−ln( RG )] R [p−0.0172 ln( RG ) −0.120] , (XV.22)
W
H2 − 6 −1.28
= [−0.000115ln(RG ) + 0.00133]R1f .2−1.295×10 λ [11.67−ln( RG )] R[p−0.0072n( RG )−0.304] . (XV.23)
W
We emphasize that equations (XV.22) and (XV.23) are explicit and predictive correlations for
proppant transport. By (XV.22) and (XV.23), H1 and H2 can be predicted from the prescribed
parameters: ρf, ρp, d,η, W, Qp, Qf.
We predict H1/W and H2/W by (XV.22) and (XV.23). In figure XV.10 and XV.11, we plot
the predicted values against the experimentally measured data. Ideally, all the points should be
on the straight-line y=x. It can be seen that the predicted values are in good agreement with the
experimental data.

16 RG=3496.28
RG=15921.522
14
RG=2818.39
12
RG=8903.2
Predicted H 1/W
10 RG=648.07
8 RG=27.51
RG=107.15
6
RG=12755.35
4
RG=73483.29
2 RG=160.25
0 RG=172.90
0 2 4 6 8 10 12 14 16 RG=643.99
Measured H 1/W
RG=25.82
Figure XV.10. The predicted values of H1/W by equation (XV.22) versus the experimentally measured values for the
cases listed in table XV.4.
y=x
14
RG=25.82
12 RG=27.51
RG=107.15
2 /W
10 RG=160.25
RG=162.29
2/W
RG=172.90
H
predicted HH
8
8
predicted
RG=643.99
predicted
RG=648.07
6
6
RG=2818.38
4
4 RG=3496.28
RG=8903.20
2
2 RG=12755.35
RG=15921
0
0 RG=73483.29
0 5 10
0 5 10
measuredH
measured H /W
H 2/W
measured 2
Figure XV.11. The predicted values of H2/W by equation (XV.23) versus the experimentally measured values for the
cases listed in table XV.4.

To test this correlation, experiments were conducted in the slot which is 16 feet long and 4
feet high. (Note that the correlations are extracted from experiments conducted in the slot which
is 8 feet long and 1 foot high.) Following are the prescribed parameters for the experiments in the
16-foot-long slot: 20/40 Ottawa with water, ρf : 0.997gm/cc, ρp : 2.645gm/cc, d: 0.0548cm, η:
0.00998poise, W : 0.79375cm, RG: 2663.189, λ: 0.000452.
Rf Rp Predicted Predicted Predicted Predicted Measured Measured

H1/W H2/W H1(cm) H2(cm) H1(cm) H2(cm)
514437.8 14327.87 28.81 25.79502 22.86977 20.4748 24.5 24.30
570181.8 38207.65 24.73 20.14101 19.62663 15.98693 20.2 19.2
Table XV.5. The predicted H1 and H2 in comparison with the measured H1 and H2 in experiments conducted in the
slot which is 16 feet long and 4 feet high.
We can see that the agreement between predicted values and measured values is encouraging.
We believe that our correlation provides a promising way to predict transport of proppant.
3.4 Logistic dose curve fitting for H1/W and H2/W

The bi-power law correlation gives good prediction of H1/W and H2/W for the bed load
transport case. However, it is not compatible with the erosion case. When Rp approaches zero,
H1/W and H2/W tend to infinity. Therefore we need a different correlation to account for the
transition region from the erosion case to the bed load transport case.
We fit the data for H1/W and H2/W to a logistic dose curve (see Appendix A for details) to
determine a function valid in the transition region; this fitting effectively combines the power
law for the erosion case and the bi-power law for the bed load transport case. We seek to
determine the function:
H1 1
= C1 R Mf1 (XV.24)
W Rp
(1 + ) N1
T1
and
H2 1
= C2 R Mf 2 . (XV.25)
W Rp
(1 + ) N2
T2
When Rp=0 (the erosion case), (XV.24) and (XV.25) reduce to:
H1
= C1 R Mf1 , (XV.26)
W
H2
= C 2 R Mf 2 . (XV.27)
W
For the erosion case, H = H1 = H2; hence, C1=C2, M1=M2 and we recover the power law
correlation (XV.13) for the erosion case. When Rp>>T1 and Rp>>T2, (XV.24) and (XV.25)
reduce to:

H1
= (C1T1N1 ) R Mf 1 R p− N1 , (XV.28)
W
H2
= (C 2T2N 2 ) R Mf 2 R p− N 2 . (XV.29)
W
Therefore, we recover the bi-power law correlations (XV.14) and (XV.15) for the bed load
transport case.
Comparing (XV.28) and (XV.29) to (XV.14) and (XV.15), we observe that M1 and M2
should be functions of both RG and λ; hence, the exponent M1=M2 in the power law correlations
for the erosion case should be functions of RG and λ. However, most of the erosion experiments
were conducted using water at different temperatures and lead to a small range of λ (See table
XV.B1); we do not have enough data to find an analytical expression M1(RG, λ)=M2(RG, λ) .
However, we find that C1, M1, T1, N1, C2, M2, T2 and N2 can be reasonably approximated by
functions of a single variable RG; hence, the coefficients in (XV.24) and (XV.25) are functions of
RG only.
We do not have data for erosion and bed load transport with the same RG (See table XV.B1
and XV.B2) so that we use (XV.24) and (XV.25) to fit data from erosion and bed load transport
with different but close RG:
1: Erosion case with RG=86 and bed load transport case with RG=109,
4: Erosion case with RG=13363,14513 and bed load transport case with RG=12229.
The results of fitting erosion with RG=86 and bed load transport with RG=109 are presented
below:
H1 1
= 0.007846 R 0f .748778 , (σ2 = 0.951) (XV.30)
W Rp
(1 + ) 0.11166
68.23195
H2 1
= 0.007846 R 0f .748778 , (σ2 = 0.946) (XV.31)
W Rp
(1 + ) 0.155747
59.2391
By (XV.30) and (XV.31), H1/W and H2/W are computed and tabulated in table XV.6.
RG Rf Rp H1/W H2/W computed error computed error
H1/W H2/W
86.83778 897.05 0 1.76378 1.76378 1.275422 0.238493 1.275422 0.238493
86.83778 1184.19 0 2.519685 2.519685 1.570219 0.901486 1.570219 0.901486
86.83778 3631.42 0 4.913386 4.913386 3.633766 1.637427 3.633766 1.637427
86.83778 14611.89 0 10.70866 10.70866 10.30601 0.162123 10.30601 0.162123
86.83778 25735.84 0 15.11811 15.11811 15.74579 0.393979 15.74579 0.393979
109.48 5202.438 570.7087 2.863739 2.301595 3.704969 0.707668 3.291238 0.979393
109.48 17341.46 1902.362 8.183906 7.060129 8.047953 0.018483 6.792796 0.071467
109.48 8670.73 95.1182 7.229334 6.502897 6.324757 0.818258 6.006224 0.246684

(continued)
109.48 8670.73 142.6773 7.168748 6.437499 6.146838 1.044302 5.76016 0.458789
109.48 8670.73 285.3543 5.53511 4.795559 5.802219 0.071347 5.300052 0.254513
109.48 8670.73 570.7087 4.594431 3.846482 5.431254 0.700274 4.82475 0.957008
109.48 8670.73 951.1811 4.282596 3.528398 5.155187 0.761415 4.482455 0.910225
109.48 8670.73 1141.417 4.220726 3.464284 5.057626 0.700402 4.36363 0.808822
109.48 8670.73 1141.417 4.283718 3.527276 5.057626 0.598934 4.36363 0.699488
109.48 26012.19 1902.362 11.9804 10.57071 10.90282 1.161175 9.202417 1.872216
Table XV.6: Data structure for the combination of erosion case with RG=86 and bed load transport case with
RG=109. H1/W and H2/W calculated by equation (XV.30) and (XV.31) are listed in comparison with the
experimentally measured values. The error is computed by (calculated value - measured value)2.
18
16 y=x
σ2 = 0.951
14
12
Calculated H 1/W
10
8
6
4
2
0
0 2 4 6 8 10 12 14 16
Experimental H 1/W
Figure XV.12(a): Experimental H1/W vs. calculated H1/W using (XV.30).
18
16 y=x
14 σ2 = 0.946
12
Calculated H 2/W
10
0
0 2 4 6 8 10 12 14 16
Experimental H 2/W
Figure XV.12(b): Experimental H2/W vs. calculated H2/W using (XV.31).

We can see that the computed H1/W and H2/W are in good agreement with the experimentally
observed values. Cases 2, 3 and 4 are processed in the same manner and resultant coefficients C1,
M1, T1, N1, C2, M2, T2, N2 and σ-squared values are listed in table XV.7.
RG C1 M1 T1 N1 σ2 C2 M2 T2 N2 σ2
109.48 0.007846 0.748778 68.232 0.112 0.951 0.007846 0.748778 59.2391 0.155747 0.946
648.1 0.000856 0.923591 4.173 0.190 0.987 0.000856 0.923591 31.60245 0.319314 0.981
2761.9 0.000184 1.080253 3.332 0.234 0.964 0.000184 1.080253 9.694597 0.299927 0.929
12229 8.94E-06 1.361329 0.174 0.133 0.9453 8.94E-06 1.361329 5.008106 0.278326 0.9448
Table XV.7: The coefficients in the logistic dose curve fitting of H1/W and H2/W for the data from erosion and bed
load transport with close RG. We only list the RG of the bed load transport case in the first column.
Next we obtain the functions C1(RG), M1(RG), T1(RG), N1(RG), C2(RG), M2(RG), T2(RG), and
N2(RG).
0.1
0.01
-1 .39 71
C 1= C 2 = 6.8405R G
σ 2 = 0.9791
C1= C2
0.001
0.0001
0.00001
0.000001
10 100 1000 10000 100000
RG
Figure XV.13: The coefficient C1 = C2 as a function of RG.
C1 = C2 = 6.8405 RG−1.3971 , (σ2 = 0.9791) (XV.32)

10
M 1=M 2 1
0.1248
M 1=M 2 = 0.4126 R G
σ = 0.9939
2
0.1
10 100 1000 10000 100000
RG
Figure XV.14: The coefficient M1 = M2 as a function of RG.
M 1 = M 2 = 0.4126 RG0.1248 . (XV.33)

2
(σ = 0.9939)
100
T 2 = 861.55R G -0.5479
σ2 = 0.9758
10
T 1 and T 2
T2
T 1 = 14730R G -1.1677 T1
1 σ2 = 0.9302
0.1
10 100 1000 10000 100000
RG
Figure XV.15: The coefficient T1 and T2 as functions of RG.
T1 = 14730 RG−1.1677 ; (σ2 = 0.9302) (XV.34)
T2 = 861.55 RG−0.5479 , (σ2 = 0.9758). (XV.35)

0.35
0.3
0.25
N2
N1 and N2
0.2
N1
0.15
0.1
0.05
0
0 2000 4000 6000 8000 10000 12000 14000
RG
Figure XV.16: The coefficient N1 and N2 plotted against RG.
We use a natural cubic spline to interpolate N1(RG) and N2(RG) and the results are plotted in
figure XV.17. Because we do not have enough data, the spline interpolation is not reliable and
could be siginificantly changed when more data become available.
0.4
0.35 n1
n2
0.3
n1 and n 2
0.25
0.2
0.15
0.1
0.05
0
0 5000 10000
RG
Figure XV.17. Use natural cubic spline to interpolate N1(RG) and N2(RG).
The resultant expressions of the spline interpolation are:

N1 = A1 + B1 RG + C1RG2 + D1 RG3 (XV.36)
where the values of A1, B1, C1, D1 are listed in the following table:

Range 109.48—648.1 648.1—2761.9 2761.9—12229
A1 9.4344824626E-02 7.9038758235E-02 3.2815903993E-01
B1 1.5713752027E-04 2.2798800770E-04 -4.2608636465E-05
C1 1.4612470982E-08 -9.4707834681E-08 3.2669847295E-09
D1 -4.4490532619E-11 1.1735517774E-11 -8.9050201147E-14
Table XV.8: The coefficients in the spline intopolation for N1(RG).
N 2 = A2 + B2 RG + C 2 RG2 + D2 RG3 (XV.37)

where the values of A2, B2, C2, D2 are listed in the following table:
Range 109.48—648.1 648.1—2761.9 2761.9—12229

A2 1.1903363556E-01 7.9383985185E-02 7.6244366512E-01
B2 3.3261384535E-04 5.1614872518E-04 -2.2579671537E-04
C2 3.7393518981E-08 -2.4579562890E-07 2.284026181E-08
D2 -1.1385189939E-10 3.1799059083E-11 -6.2257098948E-13
Table XV.9: The coefficients in the spline intopolation for N2(RG).
The final correlations:

H1 1
= 6.8405 RG−1.3971 R 0f .4126 RG
0.1248
W Rp , (XV.38)
(1 + −1.1677
) N1 ( RG )
14730 RG
H2 1
= 6.8405 RG−1.3971 R 0f .4126 RG
0.1248
W Rp (XV.39)
(1 + −0.5479
) N 2 ( RG )
861.55 RG
where N1(RG) and N2(RG) are expressed in (30) and (31).
Correlations (XV.38) and (XV.39) can be used to predict H1/W and H2/W with the prescribed
parameters W, ρf, ρp, d η, Qf and Qp. These correlations are compatible with both the power law
for the erosion case and the bi-power law for the bed load transport case.
We developed the logistic dose curve by fitting the data for erosion and bed load transport
using equation (XV.24) and (XV.25). The curve should depend on RG and λ but we do not have
enough data to determine how it depends on λ. This fitting using the dose curve is independent
of the previous power law and bi-power law correlations and the resultant equation (XV.38) and
(XV.39) do not reduce precisely to equation (XV.13), (XV.22) and (XV.23) at the respective
limits. We emphasize that the logistic dose curve represents a correlation for transition situations
between erosion and bed load transport; it is not as accurate as the power law (XV.5) for erosion
or the bi-power law (XV.22) and (XV.23) for bed load transport.
§ Summary
We believe that research leading to optimal techniques of processing data for correlations
from real and numerical experiments is founded on the far from obvious property of self
similarity (power laws) in the flow of dispersions. The basis for this belief are the excellent

correlations of experiments on fluidization and sedimentation done by Richardson & Zaki and
the correlations for sediment transport in horizontal channels obtained from our numerical
simulations and the analysis of the experimental data from STIM-LAB.
Results of two dimensional simulations of solid-liquid flows give rise to straight lines in log-
log plots of the relevant dimensionless Reynolds numbers. Power laws are also obtained from the
analysis of experimental data. The extent and apparent universality of this property is remarkable
and shows that the flow of these dispersions are governed by a hidden property of self similarity
leading to power laws. These power laws make a powerful connection between sophisticated
high performance computation, experiments and the world of engineering correlations.
The correlations obtained can be used as predictive tools or as a basis for models for
sediment transport in simulators used for design purposes.
In a sense our results here realize the opportunity which is presented by digital technology
for implementing the old tried and true method of correlations to big data sets. Our mantra is “the
secrets are in the data.” The same method works well for data from numerical and from real
experiments. We have used the following procedure not only for this but also for other
multiphase flow processes. First we propose candidates for controlling dimensionless parameters
and list the data required to form these numbers in a spreadsheet. Then we identify two
parameters and plot the results of the experiments for those two in log-log plots under conditions
in which other parameters are fixed. We have a good choice when the plots come up as straight
lines in the log-log plot. Here and elsewhere we have had excellent results in this search using
the parameters suggested by making the governing sets of PDE’s dimensionless. The results of
this kind of power law processing is that the slopes and intercepts of the straight lines in log-log
plots, or the prefactors and exponents of the power laws these lines imply, depend on the
parameters we have fixed. When we look at the variation of these parameters, the prefactors and
exponents sometimes are expressible as power laws and sometimes they are not. In any case we
may and do implement curve-fitting procedures for the prefactors and exponents to present
explicit formulas in analytic form for the prediction of future events.
We might add that the search for the governing dimensionless numbers in multiphase flow is
a way to achieve a deep understanding of the underlying physics. The method of correlations is
an excellent procedure to guide the research because the data doesn’t lie.

§ Appendix A (contribution of B. Baree)
Fitting power-law data with transition regions by a continuous function: General framework
and application to the Richardson-Zaki correlation
Many data sets representing naturally occurring phenomena can be described using a
Sigmoidal distribution function. One such function that is particularly useful in fitting physical
data is the logistic dose response curve given by:
(b − a)
y =a+ d
(XV.A1)
æ æ x öc ö
ç1 + ç ÷ ÷
ç ètø ÷
è ø
In this equation each of the constant terms or coefficients (a, b, c, d and t) have readily
apparent physical significance, which allows data modeling to be accomplished almost by
inspection.
10000
1000
100 y =1+
(1000 − 1)
êë ( )
é1 + x 2 ù
10 úû
y
10
0.1
0.01 0.1 1 10 100 1000 10000 100000 1000000
x
Figure XV.A1. A typical logistic dose response curve.
Figure XV.A1 shows the dose-response function for a = 1, b = 1000, c = 2, d = 1, and t = 10.
As can be seen, the coefficients a and b represent the values of the lower and upper plateaus of
the function, respectively, or its range. The coefficient t defines the value of the independent
variable (x) where the function deviates from the constant first plateau value. The sharpness of
curvature during the deviation from the first plateau is determined by the coefficient c. The slope
of the power-law straight line in transition from the first plateau to the second plateau is
determined by the product of coefficients c and d. The slope in this example is negative because
both exponents are positive in the denominator of the rational fraction.

10000000
1000000
1
100000 Curve C : y =
( )
−2
10000 é1 + x 2 ù
êë 10 úû
1000
100
10
1
y
0.1
0.01
0.001
1
0.0001 Curve A : y =
( )
2
0.00001
é1 + x 2 ù
Curve B : êë 10 úû
0.000001
1
y=
( )
0.0000001 2
é1 + x −2 ù
0.00000001 êë 10 úû
0.000000001
0.01 0.1 1 10 100 1000 10000 100000
x
Figure XV.A2. Effects of changing signs of coefficients c and d in the logistic dose response curve.
The effects of changing the signs of the exponents can be examined (figure XV.A2). If the
sign of coefficient c is changed, the plot is essentially rotated about a line parallel to the y-axis
through the transition value t (compare curves A and B in figure XV.A2). If the sign of the
coefficient d is changed, the plot is rotated about a line parallel to the x-axis through the upper
bound b (compare curves A and C in figure XV.A2). These relationships allow construction of a
transition function in any general form. Another useful property of this function is that the bound
corresponding to the coefficient a can be eliminated by setting its value to zero. With a = 0 the
function yields a horizontal line at the upper bound value b and a power-law line of slope c×d
which extends to infinity (Curve A, figure XV.A2). Various functions can then be modeled by
products of functions with specified power-law slopes and transition points (figure XV.A3). For
curve A in figure XV.A3 c1 = 2, d1 = 1 and for curve B c2 = 1.6, d2 = –1. The final power-law
slope in the product is then c1×d1 + c2×d2.

10000000
1000000
1
Curve B : y =
[1 + x ]
100000
1.6 −1
10000
1000
100
Product of curves A & B
10
1
y
0.1
0.01
0.001 1
Curve A : y =
0.0001
êë ( )
é1 + x 2 ù
10 úû
0.00001
0.000001
0.0000001
0.01 0.1 1 10 100 1000 10000 100000 1000000
x
Figure XV.A3. Obtaining a curve from the product of two different logistic dose response curves.
Combinations of these functions can be used in various forms to model many commonly
observed phenomena. The logistic dose response curve can also be multiplied by a linear power
law function to impose an overall slope to the function. Quite complex systems can be modeled
by combining rational fractions or products of multiple functions.
This method has been used to model the Richardson-Zaki correlation that relates bed
fluidization velocity to the solids volume-fraction of particles in suspension. The Richardson-
Zaki correlation is given by (XV.1). Specifically, the various functions representing the exponent
n are (Richardson & Zaki, 1954)
æ dö
n = ç 4.65 + 19.5 ÷ when R0 < 0.2;
è Dø
æ dö
n = ç 4.35 + 17.5 ÷ R0−0.03 when 0.2 < R0 < 1;
è Dø
æ dö
n = ç 4.45 + 18 ÷ R0−0.1 when 1 < R0 < 200 (XV.A2)
è Dø
n = 4.45R0−0.1 when 200 < R0 < 500;
n = 2.39 when 500 < R0
In these relations d is particle diameter, and D is the diameter of the fluidization column.
Note that in (XV.A2) the value of n at the transition points is not unique. Nevertheless, these

functions can be replaced with a single continuous form of the logistic dose response curve
where R0 is the independent variable and n is the dependent result (figure XV.A4).
10
n(0.2<R<1)
n(1<R<200)
n(200<R<500)
n(R<0.2)
n(R>500)
n(general)
n
1
0.1 1 10 100 1000 10000
R0
Figure XV.A4. A continuous logistic dose response curve for the Richardson-Zaki exponent n for d/D = 0.
Figure XV.A4 shows a continuous curve for the Richardson-Zaki exponent n. The
continuous form of the function is formed assuming that n should not decrease below a value of
2.39 for any value of Re. The continuous form is generated by the equation
n = 2.39 +
(2.26 + 19.5 d D ) (XV.A3)
1.1
é æ Ro ö 0.7 ù
ê1 + ç ÷ ú
ëê è T ø ûú
This function sets a minimum value of n = 2.39 and a maximum that is a function of the ratio
of particle size to vessel diameter (first of (XV.A2)). The transition value T is also a weak
function of the diameter ratio d/D, and is given by
12.0
T = 1+ (XV.A4)
æ æ d öö
ç ç D ÷÷
çç1 + çç 0.1 ÷÷ ÷÷
è è øø
The final calculation of n is then given by a combination of (XV.A3) and (XV.A4). Rowe
1987 obtained an empirical equation for the Richardson-Zaki exponent by using the logistic
curve for d/D = 0. We verified that there is good quantitative agreement between (XV.A3) (for
d/D = 0) and Rowe’s equation.

§ Appendix B
Following tables give description to the proppant and fluid used in the erosion and bed load
transport experiments.
Proppant Fluid ρf ρp d η W RG λ
(gm/cc) gm/cc) (cm) (poise) (cm)
60/4 Brady water 0.999 2.65 0.0342 0.01115 0.79375 521.1554 0.000504
20/4 Ottawa water 0.999 2.65 0.056 0.01115 0.79375 2290.846 0.000504
20/4 Light Beads water 0.999 1.05 0.06 0.01115 0.79375 86.83778 0.000504
16/30Banrite water 0.999 3.45 0.0884 0.01115 0.79375 13363.8 0.000504
12/20Badger water 0.998 2.65 0.109 0.01015 0.79375 20342.67 0.000459
16/20 Carbolite water 0.998 2.73 0.0949 0.01 0.79375 14513.68 0.000452
16/20 Carbolite 180 °F water 0.972 2.73 0.0949 0.00378 0.79375 100415.5 0.000176
Table XV.B1: Proppant and fluid parameters in erosion experiments.
Proppant Fluid ρf ρp d η W RG λ
(gm/cc) gm/cc) (cm) (poise) (cm)
20/40 sand 10cp glycol 1.11 2.645 0.0548 0.1 0.79375 27.50692 0.004067
16/30 ceramic 10cp glycol 1.14 2.73 0.097 0.1 0.79375 162.2879 0.003960
12/20 sand 10cp glycol 1.14 2.645 0.1009 0.1 0.79375 172.8956 0.003960
40/60 Sand 5cp glycol 1.11 2.645 0.0338 0.05 0.79375 25.81728 0.002034
20/40 sand 5cp glycol 1.11 2.73 0.097 0.05 0.79375 643.9946 0.002034
20/40 sand 5cp glycol 1.091 2.65 0.0548 0.05 0.79375 109.4829 0.002069
40/60 sand water 0.997 2.65 0.0338 0.0098 0.79375 648.0661 0.000444
20/40 sand water 0.997 2.65 0.0548 0.0098 0.79375 2761.919 0.000444
20/40 sand water 1 2.65 0.06 0.01 0.79375 3496.284 0.000451
20/40 Bauxite water 0.997 3.45 0.0709 0.0098 0.79375 8903.244 0.000444
16/30 Ceramic water 1 2.71 0.09 0.01 0.79375 12229.05 0.000451
16/30 Ceramic water 1 2.73 0.097 0.0098 0.79375 16127.91 0.000442
20/40 Sand 150 oF water 0.981 2.645 0.0548 0.004545 0.79375 12755.35 0.000209
16/30 Ceramic 150 oF water 0.981 2.73 0.097 0.004545 0.79375 74352.82 0.000209
Table XV.B2: Proppant and fluid parameters in bed load transport experiments.

XV Bi-power law correlations for sediment transport in pressure driven channel flows 207
§ Analogy between fluidization by drag and lift............................................................................................ 207
§ Direct numerical simulation (DNS) of solid-liquid flows........................................................................... 209
§ Experimental setup ..................................................................................................................................... 211
§ Experimental correlations for sediment transport....................................................................................... 213
Dimensionless parameters..................................................................................................................... 213
Power law correlations for the erosion case (Patankar et al. 2002)....................................................... 213
Bi-power law correlations for the bed load transport case (Wang et al. 2002) ..................................... 215
§ Summary..................................................................................................................................................... 227
§ Appendix A (contribution of B. Baree) ...................................................................................................... 229
Fitting power-law data with transition regions by a continuous function:
General framework and application to the Richardson-Zaki correlation ........................................... 229
§ Appendix B................................................................................................................................................. 233
(DISCARD THIS PAGE (only used to generate table of contents)

XVI Fluid Dynamics of Floating Particles
This chapter takes up a new problem for direct numerical simulation which includes the
topics of fluid dynamics of heavier-than-liquid floating particles driven to self-assembled
aggregates by capillary forces. This is a problem of three-phase flow; liquid/gas/solid. The
material in this chapter is taken from the two papers listed below. Prior literature is found locally
at the end of the two subsections corresponding to the two papers.
D.D. Joseph, J. Wang, R. Bai, and B.H. Yang, 2003. Particle motion in a liquid film rimming the
inside of a partially filled rotating cylinder. J. Fluid. Mech., 496, 139-163.
P. Singh and D.D. Joseph, 2005. Fluid dynamics of floating particles. J. Fluid Mech., 530, 31-80.
• Self-aggregation, clustering and bonding of particles in a liquid film

rimming the inside of a partially filled rotating cylinder
Abstract
Both lighter- and hydrophobic heavier-than-liquid particles will float on liquid-air surfaces.
Capillary forces cause the particles to cluster in typical situations identified here. This kind of
clustering causes particles to segregate into islands and bands of high concentrations in thin
liquid films rimming the inside of a slowly rotating cylinder partially filled with liquid. A second
regime of particle segregation, driven by secondary motions induced by off-center gas bubbles in
a more rapidly rotating cylinder at higher filling levels, is identified. A third regime of
segregation of bi-disperse suspensions is found in which two layers of heavier-than-liquid
particles that stratify when there is no rotation, segregate into alternate bands of particles when
there is rotation*.
I. Capillary forces
The deformation of the air-liquid interface due to floating light particles or due to trapped
heavy small particles gives rise to capillary forces on the particles. These forces may be
qualitatively understood from simple arguments. Two kinds of forces act on particles: forces due
*
Movies of the experiments reported in this paper can be viewed at the web address
http://www.aem.umn.edu/research/particles/rtcylinderpaper/ .
Printed 3/28/2005 9:58:00 PM XVI-1 DDJ/2002/papers/MotionParticles/ChapXVI

to gravity and forces due to the action of contact angles. These two kinds of forces are at play in
the vertical force balance but require a somewhat more elaborate explanation for horizontal force
balance. The effects of gravity are usually paramount for heavier-than-liquid floating particles in
which one particle will fall into the depression of the second. A heavier-than-liquid particle will
fall down a downward sloping meniscus while an upward buoyant particle will rise. Capillary
forces cause particles to cluster, as shown in figure 4.
In this section, we shall review the nature of capillary forces which cause the particles to
cluster; in section II we show how these forces produce islands and bands of segregated particles
in a thin liquid film rimming the inside of a slowly rotating cylinder.
I-1. Vertical forces

The simplest analysis relevant to understanding the forces on small particles is the vertical
force balance on a sphere floating on the interface between fluids which, for convenience, is here
called water and air. This analysis was given first by Princen (1969), then by Rapacchietta and
Neumann (1977) and by Kotah, Fujita and Imazu (1992), who used the floating ball to measure
contact angles.
Hydrophilic particle Tangent to point of contact
Hydrophobic particle Tangent to
point of
contact
Air Va Tangent to
meniscus Air Va
R R h2
θc
Water θ1 h2
Vw α Vw θc α
Water Tangent to
Contact θ1
g meniscus
angle
Figure 1. Hydrophobic and hydrophilic particles at equilibrium.
The capillary force Fc , is a function of the radius of particle R , the surface tension
coefficient γ , the filling angle (position of the contact ring) θ c and the contact angle α (see
figure 1), given by,
Fc = 2π (Rsin θ c )γ sin[θ c − (π − α )] = −2πRγ sin θ c sin(θ c + α ) (1)
for both the hydrophobic and hydrophilic cases.

To have mechanical equilibrium, the capillary force plus the vertical resultant of pressure
around the sphere must balance the gravity vertical component,
Fc + Fp = G (2)
where G = 43 ρ p πR 3 g is the weight of the particle, and the vertical resultant of pressure around the
sphere can be written as

Fp = ρl gVw + ρ a g(V − Vw ) − ( ρl − ρ a )gh2 A . (3)
where ρ l and ρ a are densities of the liquid and the air, respectively; h2 is the depression
generated by the particle, with positive value in the case shown in figure 1(a) and negative value
shown in figure 1(b); V = 43 πR 3 is the volume of the sphere, Vw = πR 3 (23 − cos θ c + 13 cos 3 θ c ) is the
volume of the sphere immersed in the water and A = π (Rsin θ c ) is the area of the ring of
2
contact. The first two terms at the right hand side of (3) are in agreement with Archimedes’
principle, while the last term accounts for the meniscus effect. When a meniscus is present, the
buoyancy calculated by Archimedes’ principle ρ l gVw + ρ a g(V − Vw ) not only lifts the sphere, but
also the fluid in the meniscus.
Inserting (1) and (3) into (2), we get the vertical force balance,
⎡4 ⎛2 1 ⎞ ⎛2 1 ⎞⎤
B⎢ 1 ψ − ⎜ − cos θ c + cos 3
θ c ⎟ − ψ 2 ⎜ + cos θ c − cos 3
θ ⎟
c ⎥
sin θ c sin(θ c + α ) = − ⎢ 3 ⎝3 3 ⎠ ⎝3 3 ⎠⎥ (4)
2
⎢⎣+ (1− ψ 2 )(cos θ c − cosθ1)sin 2 θ c ⎥⎦
where cosθ c − cosθ1 = h2 R  θ1 is measured from the point of extension of the flat meniscus
as indicated in figure 1. B = ρ l R 2 g γ is the bond number and ψ1 = ρ p ρ l and ψ 2 = ρ a ρ l are the
dimensionless control parameters.

It can be inferred from (4) that the left side of the equation, consequently, the right side, lies
in the range −1 ≤ sin θC sin(θC + α ) ≤ 1. Obviously this equation cannot be solved if the particles
are too large or too heavy. However, it can be concluded that small hydrophobic particles can
always be suspended in fluid surfaces no matter how heavy they may be, as long as ρ p R 2 g γ is
small enough. Moreover, in the limit of ρ p R 2 g γ → 0, sinθ c sin(α + θ c ) = 0 and the particles sit
on the top of the fluid or are held in place by capillarity.

If the particle is irregular with sharp corners the capillarity argument fails. Liquid-air surfaces
bind at razor sharp corners; the physics associated with this strong bond are not understood.
Razor blades and straight pins can float on water-air surfaces pinned at the sharp surface.
Equation (4) suggests that hydrophobic nanoparticles can float on the surface no matter how
heavy they are. However, even though the formula does not predict that hydrophilic particles will
sink, they will sink because of a not-understood wetting instability. If heavy nanoparticles are put
into the liquid, they will not rise. The surface layer on a liquid which gives rise to surface tension
is very small, but not zero; likewise, the concept of contact angles on nanoparticles may lose
meaning.
I-2. Horizontal forces

The deformation of a liquid-fluid interface due to trapped small particles gives rise to lateral
capillary forces exerted on the particle. A simple explanation is shown in figure 2. For a heavier-
than-liquid particle, the meniscus is below the undisturbed level. The particles will tilt causing an
imbalance of the horizontal component of capillary forces pulling the spheres together. Lighter-
than-water hydrophilic particles will rise into the elevated section of the meniscus and come
together.
(a)
Fc
π−α Fc
π−α
(b) F1h
F2h
(c)
Figure 2. Spherical particles in water, (a) heavier-than-water hydrophobic particles. (b) Lighter-than-
water hydrophilic particles. (c) If for any reason, the particle tilts with the two contact angles equal,
a horizontal force imbalance will result. In the figure, the vector Fc indicates the magnitude of the
capillary force and F1h and F2h are horizontal components, F2h > F1h .
There are several ways to isolate the effects of capillarity uninfluenced by gravity. Poynting
and Thompson (1913) investigated the capillary effect by considering two vertical plates
immersed in a liquid, the space between the plates is a two dimensional capillary tube. If the
plates are hydrophobic, the level in the capillary gap sinks below the liquid outside; if the plates

are hydrophilic the levels will rise. Their argument about the nature of horizontal forces on the
plates is given in the caption of figure 3. Repulsion between plates with different wetting
properties is rather short range because it stops when the meniscus between plates gets flat.
(a) (b) (c)
pa > p
L
p
L
p >p
L a
pa
Figure 3. (After Poynting and Thompson 1913). Horizontal forces associated with the fall (a) of liquid
between hydrophobic plates and the rise (b) of liquid between hydrophilic plates. In (c) one plate is
hydrophilic and the other hydrophobic. The contacts on both sides of a plate are the same and the
tension γ is constant. They argue that the net horizontal force due to γ can be calculated at flat
places; so that there is no net horizontal component of the tension. In (a) and (b) the pressures are
such that they push the plates together; there is no net attractive force in (c). If the plates (c) are so
close that there is no flat place, then the horizontal projection γ sin α of the interface midway between
the plates is smaller than the horizontal component outside the plates and the plates are pulled apart;
they repel. They note that “…small bodies, such as straw or pieces of cork, floating on the surface of
a liquid often attract each other in clusters; this occurs when the bodies are all wet by the liquid and
also when none of them is wet; if one body is wet and one is not wet, they repel each other.” (It may
help here to note that if one face of the plate is hydrophobic and the other hydrophilic, the contact
angles will put the plates in tension, tending to pull them apart.)
Another way to take away the effects of gravity is to support the particles on a substrate. In
this case the horizontal forces are due to capillary effects alone. Katoh, Fujita and Imazu (1992)
studied the motion of a particle floating on a liquid meniscus which could be interpreted as
motion on a substrate because the foaming phlystyrol particles used by them are an order of
magnitude lighter than water, and minimize the effects of gravity compared to capillarity. Their
experimental results are completely consistent with the predictions of Poynting and Thompson
(1913): when the sphere and the wall are alike with respect to wetting; say both are hydrophobic
or hydrophilic, the wall and sphere attracts; when they are unlike the sphere and wall repel.

Despite the well-established importance of the capillary meniscus forces there are only a few
theoretical works devoted to them. Nicolson (1949) was the first to derive an analytical
expression for the capillary force between two floating bubbles by using the superposition of
approximation to solve the Laplace equation of capillarity. A similar approximate method was
applied by Chan, Henry and White (1981) to floating spheres and horizontal cylinders. For
horizontal cylinders the alternative approaches were proposed by Gifford and Scriven (1971) and
by Fortes (1982). The theoretical works are based on solutions of the Laplace equations for
capillary menisci of translational or rotational symmetry, where the Laplace equation reduces to
an ordinary differential equation.
An analytical solution of the Laplace partial differential equation in bipolar coordinates was
proposed by Kralchevsky, Paunov, Ivanov and Nagayama (1992), and Kralchevsky, Paunov,
Denkov, Ivanov and Nagayama (1993) for the case of small particles and small meniscus slope.
This solution provides expressions for calculating the capillary meniscus force between two
vertical cylinders, between two spheres partially immersed in a liquid layer and between a
vertical cylinder and a sphere. A review is recently presented by Kralchevsky and K. Nagayama
(2000).
Their theory (see Kralchevsky and Nagayama, 2000), which has been validated in
experiments, provides the following asymptotic expression for calculating the lateral capillary
force between two particles of radii R1 and R2 separated by a center-to-center distance L,
[ ]
F = −2π γ Q1 Q2 qK1 (qL) 1+ O(q 2 Rk2 ) when L >> rk (5)
where rk = Rk sin θ c (k=1 and 2) are the radii of the two contact lines (see figure 1);
Qk = rk sinψ k (k = 1 and 2) (6)
with ψ k being the meniscus slope angles with respected to the horizontal plane at the contact
point ( ψ > 0 for floating light particles, and ψ < 0 for heavy particles);
q= (ρl − ρa )g γ (7) 1
is the inverse of the capillary length; in addition K1(x ) is the modified Bessel function of the first
order. Therefore, the lateral capillary force between two identical particles is
F = −2π γ Q1Q2 L , (7)2

when the distance between them is much smaller than the capillary length ( q−1 = 2.7mm for
water-air interface).
I-3. Particle clustering

Due to the attractive lateral capillary forces between similar particles floating on a liquid
surface, particles tend to cluster. The dynamic behavior of clustering is not well characterized.
Gifford and Scriven (1971) noted that “casual observations… show that floating needles and
many other sorts of particles do indeed come together with astonishing acceleration. The
unsteady flow fields that are generated challenge analysis by both experiment and theory. They
will have to be understood before the common-place ‘capillary attraction’ can be more than a
mere label, so far as dynamic processes are concerned.”
There are a small number of theoretical studies of the drag and diffusion coefficient of a
spherical particle attached to a fluid interface (Brenner and Leal 1978, 1982; Goldman, Cox and
Brenner 1967; Schneider, O’Neill and Brenner 1973; Majumdar, O’Neill and Brenner 1974—
which may be collectively designated as Brenner et al—and Wakiya 1957; Redoev, Nedjalkov
and Djakovich 1992; Danov, Aust, Durst and Lange 1995). A recent study of Saif (2002)
develops a theory of capillary interactions between solid plates forming menisci on the surface of
a liquid.
The only experimental determination of drag coefficients for particles of any size were
performed by Petkov, Denkov, Danov, Velev, Aust and Durst (1995) for particles of sub-
millimeter radius by measuring the particle velocity under the action of well defined external
force. They showed that the capillary interactions are quite strong and very long range.
Accelerations, which are very great under many conditions of interest, have not been studied
before.
We found that the initially randomly distributed particles floating on a liquid surface tend to
cluster due to the attractive lateral capillary forces between the particles. It is generally observed
that the particles initially form small clusters, then the small clusters slowly merge into bigger
ones; and eventually the bigger ones are assembled into a giant cluster. This self-assembly
process is shown in the pictures of figure 4. The procedure by which we obtain dispersions like
those shown at 3 min in figure 4 is noteworthy. We create such dispersions by pouring particles
on the liquid, nothing complicated, just like a salt shaker. As soon as the particles hit the liquid

surface they disperse rapidly leading to dispersions like that at 3 min in figure 4. The dispersions
then attract. This initial repulsion, followed by attraction, is more or less universal and we have
not seen it mentioned in the literature.
Elapsed time: 3 minutes 10 minutes
30 minutes 2 hours
12 hours 24 hours
Figure 4. Free motions leading to self-assembly of floating particles (sands in 1% aqueous polyox
solution).

Experiments on particle clustering due to capillarity were carried out in glass petrie dishes
with diameters ranging from 5 to 15 cm. Three different liquids (table 1) and two different
hydrophobic particles (table 2) were used. Hydrophobic particles did not collect on the glass
side-walls of the petri dishes so that end effects arising from walls in dishes of different sizes
were not important.
Type of liquid Glycerin Soybean oil Triton mixture
Density ρ (g/cm3) 1.173 0.915 1.241
Viscosity µ (cp) 1490 282 2950
Surface Tension (mN/m) 41.46 24.28 33.15

(measured with a spinning-drop tensiometer)
Table 1. Physical parameters for liquids.
Type of particle Polymer particle Nylon particle
Density ρ (g/cm3) 1.034 1.170
Diameter dp (cm) 0.065 0.314
Table 2. Physical parameters for particles.
Clustering, of the type shown in figure 4, was observed for both types of the particles in
glycerin and in the Triton-water mixture. These particles could not be suspended in soybean oil,
because they were too heavy. However, they could be held by capillarity in the hanging film
shown in figure 7.
Rate of approach experiments were conducted for two identical particles attracted by
capillary forces on the Glycerin-air and Triton mixture-air interface. The distance between the
two particles was measured with a video camera as a function of time until they touch, as shown
in figures 5 and 6. In general, the approaching of the particles takes hours, agreeing with the time
for cluster formation. The curves are not similar near the time of final approach. The approach
velocity depends strongly on liquid properties. Particles in a 6000 cp Triton mixture barely move
even when they are placed very close together. On the other hand, the rate of approach of
hydrophobic particles on water-air surfaces is surprisingly fast in the final times. The estimated
final approaching velocities are 0.2 µm s and 0.025 mm s for the data shown in figures 5 and 6,
respectively. The approach velocity is smaller for the smaller particles, but the data was erratic
and quantitative results were not obtained.

14
Distance between surface (mm)

12
10
8
6
4
2
0
0 100 200 300 400 500 600 700 800 900 1000
Time (minute)
Figure 5. Distance between the two identical particles. Triton mixture (2950 cp) and nylon particles were
used.
14
Distance between surface (mm)
12
10
8
6
4
2
0
0 20 40 60 80 100 120
Time (minute)
Figure 6. Distance between the two identical particles. Glycerin and nylon particles were used.
The photograph in figure 7 shows the aggregation of polymer particles in a hanging glycerin
film on the bottom of a flat glass plate taken from the top of the plate. The particles are
encapsulated by glycerin and drawn together in hanging drops of glycerin robustly stable for
months. This hanging drop configuration is shown in figure 9 as a cartoon in side view.

Figure 7. Aggregates of polymer particles in glycerin drops hanging from the bottom of a glass plate.
The ability of self-assembly of particles under the action of lateral capillary forces has been
used by Bowden et al. (1997, 1999) and Grzybowski et al. (2001). They assembled topologically
complex mesoscale (from millimeter to micrometer size) objects into ordered two-dimensional
arrays by floating the objects at the interface between perfluorodecalin (hydrophobic) and water.
The structure of the arrays was manipulated by the design of the shape of the assembling objects
and wettability of their surfaces. They modeled the self-assembly process as the minimization of
the total interfacial free energy (the sum of the capillary energy and the gravitational energy) of
the liquid-liquid interface.
II. Particle aggregation in a liquid film rimming a rotating cylinder

Tirumkudulu, Tripathi and Acrivos (1999) first reported particle segregation in a suspension
of monodispersed neutrally buoyant spheres in a Newtonian liquid medium being sheared in a
partially filled horizontal Couette device. They found that the suspension separates itself into
alternating regions of high and low particle concentration along the length of the tube. In a
following study, Tirumkudulu, Mileo and Acrivos (2000) (hereafter denoted as TMA 2000)
observed that under certain circumstances particles which are initially uniformly mixed in a film
rimming a horizontal rotating cylinder will also be drawn into cylindrical bands of high particle
concentration separated by regions of pure liquid. They did not offer a quantitative explanation
of this phenomenon but suggested that the cause might be found in changes of the effective
viscosity of the suspension induced by fluctuations of concentration. A theory relying on the

shear induced diffusion of particles, concentration-dependent viscosity and the existence of a
free surface was developed by Govindarajan, Nott and Ramaswamy (2001) to provide an
explanation of the above mentioned experiments. However, quantitative comparison with the
experimental data was not provided. A latest experimental result was reported by Timberlake and
Morris (2002) in which concentration band dynamics was studied using a partially filled Couette
device. They showed that the particle migration process observed in experiments was much
faster than that predicated by the shear induced diffusion theory, about 40 times faster in one
case examined, suggested strong evidence against shear-induced diffusion as the mechanism
responsible for the observed segregation.
We carried out similar experiments and identified two regimes in which particles segregate; a
low-speed, low-Reynolds number regime, in which particles are segregated at thin places on the
rimming film by lateral capillary forces, and a high-speed regime associated with the formation
of bubbles (Balmer 1970, Karweit and Corrsin 1975, Preziosi and Joseph 1988 among others).
The segregation at low Reynolds numbers occurs in the parameter ranges similar to those studied
by TMA 2000. The high-speed segregation has not been noted before.
II-1. The ratio of the minimum film thickness to the particle diameter
The segregation of particles due to capillarity occurs in the thin part of the film rimming the
rotating cylinder near the top of the cylinder. The critical parameter for this appears to be the
ratio Dmin/dp, where Dmin is the minimum film thickness which is near the top of the cylinder to
the side in which the gravity and the vertical component of rotation point downward (figure 8).
R R
Ω Ω
D(θ) D(θ)
θ g θ
Figure 8. Film profile in rimming flow inside a rotating cylinder. (a) For small β. (b) For β larger than a
critical value ( β c = 1.414 ).
We find that Dmin/dp is O(1) and the rotational speed of the cylinder must be slow enough relative
to the speed of capillary attraction to allow clusters to form more rapidly than they disperse in

the pool of liquid at the bottom of the rotating cylinder. TMA (2000) identified the relevant
rotation parameter
gR
β =F (8)
νΩ
where F is the fill ratio, i.e., the ratio of the total volume of liquid inside the cylinder to the
volume of the cylinder; R is the radius of the cylinder; Ω is its angular rotational speed; ν is the
kinetic viscosity of the liquid; g is the gravitational acceleration. The parameter β is the only
dimensionless parameter to arise in lubrication theory; its relation to the filling parameter F is
subtle and needs clarification. When β < βc = 1.4142, a smooth film exists inside the rotating
cylinder. However, when β is increased beyond the critical value βc, smooth solution of the
lubrication equations does not exist, and a bump is formed near the bottom of the cylinder where
the film thickness varies rapidly, as is shown in figure 8 and in figures 1 and 3 of TMA (2000).
Lubrication theory can be used to compute Dmin when β < βc; it can also be used to compute the
minimum film thickness in the region above the pool of liquid when β > βc by a procedure which
will be discussed below, but the solution is not uniformly around the cylinder in the pool below.
The critical condition for the existence of a smooth solution of the lubrication equation for a thin
film on the exterior of a rotating rod was called the “run-off” condition by Preziosi and Joseph
(1988) and the critical condition for rimming flow was called a “run-on” condition. The run-off
and run-on conditions were verified in experiment reported by Preziosi and Joseph (1988).
Many authors have published analyses of lubrication flows of liquids running inside of a
rotating cylinder: Diebler and Cerro (1970), Moffatt (1977), Preziosi and Joseph (1988), Johnson
(1988), O’Brien and Gath (1998) and Tirumkudulu and Acrivos (2001).
The segregation of particles by capillary forces does not correlate with β but the ratio Dmin/dp
is important and lubrication theory can be used to compute Dmin by the method described below.
The physical parameters are the cylinder radius R, the kinematic viscosity of the fluid ν, the
angular speed of rotation Ω, the volume fraction of the fluid F, the acceleration of gravity g, the
liquid flux Q, and the thickness of the film D. Preziosi & Joseph (1988) obtained the following
equation for D using a lubrication theory:
g1 3 g1 3
Q = ΩRD − D cos θ = ΩRD0 − D0 (9)
ν3 ν3

where D0 is the film thickness at θ = 0. Let h = D/R and S = gR/νΩ, (9) can be written in the
following non-dimensional form:
Q 1 1
2
= h − Sh 3 cos θ = h0 − Sh03 (10)
ΩR 3 3
where ho is the maximum film thickness; ho = h(θ ) at θ = 0 (see figure 8). Preziosi & Joseph
gave the condition under which equation (10) is solvable,
h02 S < 1 . (11)
Expression (11) is also the critical criterion of run-on.

1/ 2 1/ 2
⎛ g ⎞ Q 1/ 2
O’Brien and Gath (1998) defined η = D⎛⎜
g ⎞ D 1/ 2
⎟ = S and q = Q⎜ 3 3 ⎟ = S .
⎝ Rν ⎠
Ω R ⎝Ω R ν ⎠ ΩR 2
Under such definitions, (10) becomes:
1 1
q = η − η 3 cos θ = η 0 − η 03 (12)
3 3
O’Brien and Gath gave the condition under which equation (12) is solvable:
0 < q < 2/3 (13)
Note that when q = 2/3, the solution of (12) is η0 = h0 S 1 / 2 =1. Hence, the run-on criterion (11) is
equivalent to (13).
The fluid fraction F can be computed by integrating D(θ ):
1 π β 1 π
F= ∫ D(θ )dθ ≡ 1 / 2 , where β = ∫−π η (θ )dθ . (14)
πR −π
S π
By virtue of (12) and (14), the value of β corresponding to q = 2/3 is β =1.4142.
Therefore, the three run-on conditions are equivalent: h02 S < 1 , 0 < q < 2/3, and β < 1.4142.
When β < 1.4142, all the fluid join the circulation. The fluid flux is obtained from F,
Q F
2
= . (15)
ΩR 2
Hence, equation (10) can be solved for the film profile, h(θ ),
h − 13 Sh 3 cosθ = 12 F , (16)
and the minimum film thickness can be obtained from θ = π .

The maximum film thickness ho at θ = 0 is an increasing function of β with a maximum at
β = βc. When β > β c = 1.4142 , there are places on the cylinder where the thickness of the layer is
larger than the critical value, and the excess fluid will collect under the bump. However, it may
be assumed that ho remains at θ = 0 (see figure 8); it is the maximum thickness of the film above
the bump. This assumption could not be strictly correct; Ruschak and Scriven (1976) showed
that under a perturbation of the thin film condition used to justify lubrication theory the position
of the maximum thickness rotates into the first quadrant. We assume that the maximum film
thickness that can be maintained by rotation is determined by the critical run-on condition (11),
h0 = 1 S = νΩ gR . Therefore, we can calculate the actual fluid flux in the circulation by
Q S 2
= h0 − h03 = . (17)
ΩR 2
3 3 S
Then the minimum film thickness at θ = π is determined from the volume conservation (10),
Q S 3 2
= hmin + hmin =
ΩR 2
3 3 S
or,
0.596 νΩ
hmin = = 0.596 . (18)
S gR
Tirumkudulu & Acrivos (2001) solved the film profile using lubrication analysis and numerical
computation from the full Stokes equations, and compared their solutions with experimental
gR
measurements. They found that ηmin = hmin = 0.6 , (in their Figure 5), which agrees perfectly
νΩ
with our expression (18) based on a much simpler argument. Therefore, when β is greater than
β c, the minimum film thickness Dmin listed in table 3 are evaluated from (18). Corresponding to a
specified filling level F, we may also determine the average film thickness Da from
R 2 − (R − Da)
2
F=
R2
(
, that is, Da = 1− 1− F R . ) (19)
II-2. Particle segregation in aqueous Triton mixtures

TMA 2000 found particle segregation in monodispersed sheared suspensions in a partially
filled rotating horizontal cylinder when the filling fractions (liquid volume/total volume) were

small 0.1 ≤ F ≤ 0.15. The particle concentrations for the uniform mixtures were 5% and 15%.
The values of β in experiments reported in TMA 2000 were all greater than β c .
Systematic experiments on clustering of particles into bands were carried out using the
polymer particle whose properties are described in table 2. In these experiments we used the
same fluid as TMA. The liquid is a mixture of Triton X 100, ZnCl2 and water in combinations
used to control viscosity. The high viscosity mixture is in the range of 20-60 poise and a density
in the range 1.1~1.5 g/cm3. The viscosity of the mixture is sensitive to temperature which was
maintained at 68 ± 2°F in our experiments.
The experiments were conducted in two different rotating cylinders; one is 30 cm long and
the inside diameter is 2.792 cm; the other is 15 cm long with the same inside diameter. The
cylinder is supported horizontally and is driven at constant rotational speed Ω by a motor. For
these experiments, the Reynolds number Re = (ΩDa2 ) ν , where Ω is the angular velocity of the
cylinder, Da the mean thickness of the film, and ν the kinetic viscosity of the pure liquid, is
always very small (less than 10-2). Inertial effects were generally negligible.
F R Ω µ ρ ν β Da Dmin Dmin /dp

(cm) (rpm) 3 2
(poise) (g/cm ) (cm /s) (cm) (cm)
TMA 1 0.150 1.270 1.40 40.00 1.172 34.13 2.36 0.099 0.0480 1.04
TMA 2 0.125 5.000 2.80 49.00 1.172 41.81 2.50 0.323 0.149 3.24
M1 0.151 1.396 1.65 51.95 1.241 41.86 2.08 0.110 0.0605 0.931
M2 0.140 1.396 1.65 29.50 1.332 22.15 2.05 0.101 0.0440 0.677
M3 0.150 1.396 1.10 51.95 1.241 41.86 2.53 0.110 0.0494 0.760
M4 0.145 1.396 10.9 48.50 1.212 40.02 2.11 0.105 0.0966 1.49
M5 0.061 1.396 1.76 44.34 1.203 36.86 0.87 0.043 0.0403 0.620
M6 0.061 1.396 3.13 44.34 1.203 36.86 0.65 0.043 0.0412 0.634
M7 0.061 1.396 6.00 44.34 1.203 36.86 0.47 0.043 0.0418 0.648
M8 0.061 1.396 10.0 44.34 1.203 36.86 0.37 0.043 0.042 0.646
M9 0.046 1.396 38.71 2.377 1.498 1.587 0.67 0.032 0.0310 0.477
M10 0.046 1.396 51.10 2.377 1.498 1.587 0.58 0.032 0.0313 0.482
Table 3. Parameters for experiments reported in TMA (2000) and for our experiments (M1 through M10).
In table 3 we listed the parameters for the experiments using Triton mixtures reported by
TMA and our experiments using the polymer particles (table2) colored blue for visualization.

The derived parameters (β ,Da ,Dmin ,Dmin d p ) are based on the properties of the pure liquid. The
particles used by TMA1 and TMA2 were neutrally buoyant with density 1.172 g cm 3 , and
diameter d p = 0.04625 ± 0.00375 cm in concentrations of 15%. The values of β for TMA1 and
TMA2 based on the viscosity of the homogenous suspension would be 1.8 and 1.9, respectively.
In our experiments, M1-M10, the density of the particles ρ p = 1.034 g cm 3 is less than fluid
density; the diameters of the particles are more dispersed with an average particle diameter
d p = 0.065 cm . The concentrations of the particles range from 2% to 7%. The cylinder length is
15 cm from M4 and M5; otherwise the cylinder length is 30 cm. The values of the minimum film
thickness, Dmin d p , listed in table 3 are all of O(1), and are consistent with the observations of
particle segregation driven by capillarity of the thin films. Although our particles are not
neutrally buoyant, the sedimentation of the particles in the liquid used in the experiments can be
neglected, since the sedimentation velocity of the particles in those liquids is of the order of
10 nm s .
Top of cylinder
Air
Figure 9. Capillary attraction of two particles hanging in a film at the top of a stationary rotating
cylinder. The liquid film is the top section (the gray area). The air fills the other space.
After the cylinder is partially filled with the uniform suspension, it is turned a few times by
hand and then put to rest so that the suspension covers the whole inside cylinder. It was observed
that particles trapped in the thin film at the top of the cylinder move rather rapidly together under
the action of capillarity (see figure 9). A similar kind of dynamics prevails when the cylinder
rotates continuously at a constant velocity. In general the trapped particles are completely wet by
the liquid as they pass through the deep pool at the bottom of the rotating cylinder. The
segregation of particles generally occurs slowly. It takes a long time (hours) for the particles to
reach the final steady band formation. The particle segregation occurs in a number of stages. The
first is the formation of many small particle clusters, which were nucleated randomly along the
cylinder after a few minutes. As time passes by, small particle clusters merge into larger ones.

Eventually they form into several comparatively large blocks which are often far from each
other. These large blocks are quite stable, and stay separated. Meanwhile, they are gradually
stretched thinner and longer in the flow direction, and eventually form cylindrical bands. Some
bands may merge. The final formation of the bands is frequently uniform along the cylinder axis.
The bands are not robustly stable; they may born, move, break and reform. It can be said that
uniform dispersion is robustly unstable and clusters are robustly stable. The snap shots of the
particle band formation are shown in figure 10.
(a) Uniform distribution of particles at the beginning
(b) Particle clusters
(c) Larger clusters
(d) Particle bands

Figure 10. Process of particle bands formation. The experiment is performed with high viscous Triton-
mixture fluid under low rotating speed and low filling level (case M7).

In our experiments, it is observed that the formation of particle clusters or bands is easier in
more viscous liquid at a low rotational speed. In experiments with a high-viscosity Triton
mixture, we had to restrict to low rotating speed to get bands to separate. The rotational speed of
the cylinder has to be slow enough such that the time needed for capillary attraction is
comparable to the residence time for the particles in the thin part of the rimming film.
For the low-viscosity Triton mixtures, it is possible to get band separation by decreasing the
filling level and at the same time increasing the rotational speed. The band formation was
observed for the range of small β values (0.5~0.7). However, the rotational speed should not be
too high, since then the particle clusters may not form, and even if particle clusters form they are
not stable. The process of particle segregation in low-viscosity Triton mixtures (shown in figure
11) is similar to the case with high-viscous Triton-mixture (figure 10). Decreasing the filling
level F as well as increasing the rotating speed is not always effective for low-viscosity fluids;
we repeated the experiment with soybean oil and glycerin with relatively high rotating speeds
(corresponding to β =1.0~1.2), and even though bands did not form, particle clusters were
always generated by capillary induced “anti-diffusion”.
(a) Initial distribution (b) Particle clusters
(c) Particle blocks (d) Initial bands
(e) Developing bands (f) Final bands

Figure 11. Process of particle bands formation. The experiment is performed with low viscous Triton-
mixture fluid under relatively high rotating speed and low filling level (case M10).

Experiment M3 (table 3) is a high filling level case. Band separation occurred, however,
there were many particles in the liquid sections between particle bands, in contrast to the case
with low-filling where clear pure liquid between the particle bands were observed. The
segregation of particles into cylindrical bands may take hours even days. The achieved
configurations are stable for times of the order of hours and days. There is slight secondary flow
and transverse movement of particles can be seen between particle bands.
Figure 12. Particles which are initially distributed uniformly in a film rimming a rotating cylinder
segregate into cylindrical bands (case M3). The formation of the bands takes hours.
In table 4, we list values for the times of formation of small clusters tw1, large clusters (called
blocks) tw2 and bands tw3 and the distance between bands as a function of the filling level and
angular velocity. In general, for the same filling level, clusters form faster with the rotational
speed increasing.
Filling level Rotating Waiting time for Waiting time Waiting time Average distance
F speed particle clusters for large blocks for bands between bands
Ω (rpm) t w1 (hour) t w 2 (hour) t w3 (hour) l (cm)
0.061 1.76 1.1 2.6 6.6 6.2
0.061 3.13 0.8 2.1 4.7 5.5

0.061 6.00 0.6 2.2 4.0 3.9
0.061 10.0 0.5 1.8 3.5 4.5
0.046 38.71 0.10 0.25 0.40 3.9
0.046 51.10 0.08 0.20 0.30 3.6
Table 4. Times and distance of cluster and band formation.

In this section we described cluster and band formation due to capillarity for lighter-than-
liquid polymer particles in small concentrations in a highly viscous Triton mixture under
conditions in which Dmin/dp = O(1).
II-3. Particle segregation in water

We did experiments in water using the same polymer particles as before. In figure 13 we
show cluster and band formation due to capillarity for heavier-than-water particles in large
concentration 20.7% with β = 15.94.
(a) 3 minutes after beginning of rotation
(b) 1.5 hours after beginning of rotation
(c) 6 hours after beginning of rotation
(d) 18 hours after beginning of rotation
Figure 13. Band formation of particles due to capillarity. The fluid is water (18°C) and the particles are
polymer particles with a density of 1.034 g/cm3 and a diameter of 0.065 cm. The filling level of the
fluid is 4.08% and the particle concentration is 20.7%. The rotating speed Ω = 8.57 rpm, β = 15.94,
average film thickness Da = 0.288 mm, minimum film thickness Dmin = 0.0213 mm, Dmin /dp = 0.033.

II-4. Particle segregation in glycerin
Clustering and band formation due to capillarity is very robust in thin films on the inside of a
rotating cylinder. In figures 14, 15 and 16 we show clusters and bands for three different cases.
In figure 14 the neutrally buoyant particles are very large and the concentration of particles is
54.2%. In figure 15 the particles are much heavier than glycerin, but bands form nonetheless. In
figure 16 the particles are lighter than glycerin and bands form as well.
Figure 14. Clustering of particles due to capillarity. The fluid is glycerin with a density of 1.173 g/cm3, a
viscosity of 1490 cp, and a surface tension of 41.46 mN/m. The particles are white Nylon particles
with a density of 1.170 g/cm3 and a diameter of 0.314 cm. The filling level of the fluid is 8.2% and the
particle concentration is 54.2%. The rotating speed Ω=5.45 rpm, β=1.127, average film thickness
Da=0.585 mm, minimum film thickness Dmin= 0.526 mm, Dmin /dp = 0.176. Clusters of particles form
about 5 minutes after starting of rotation. Bands form occasionally but are not stable.
(a) 2.5 hours after beginning of rotation.
(b) 16 hours after beginning of rotation.
Figure 15. Band formation of particles due to capillarity. The fluid is glycerin with a density of 1.173
g/cm3, a viscosity of 1490 cp, and a surface tension of 41.46 mN/m. The particles are 16/20 Naplite
sands with a density of 2.59 g/cm3 and a diameter of 0.959 mm. The filling level of the fluid is 7.85%
and the particle concentration is 13.0%. The rotating speed Ω =5.45 rpm, β=1.079, average film
thickness Da = 0.559 mm, minimum film thickness Dmin = 0.506 mm, Dmin /dp = 0.528.

Figure 16. Band formation of particles due to capillarity. The fluid is glycerin with a density of
1.173 g/cm3, a viscosity of 1490 cp, and a surface tension of 41.46 mN/m. The particles are polymer
particles with a density of 1.034 g/cm3 and a diameter of 0.65 mm. The filling level of the fluid is
10.5 % and the particle concentration is 4.8%. The rotating speed Ω = 5.45 rpm, β = 1.443, average
film thickness Da = 0.753 mm, minimum film thickness Dmin = 0.605 mm, Dmin /dp = 0.931. Band
formation 15 hours after beginning of rotation is shown in the figure.
III. Particle segregation due to the formation of bubbles

When a partially filled horizontal cylinder is rotated at rates which are not too high or too
low such that the effects of surface tension and gravity are both important, air bubbles separated
by disks of liquid will form. The bubbles are then not centered and can take different shapes
depending on conditions. The off-center bubbles pump the liquid to form the secondary motion
which is from the bubble to the liquid disks near the bubble surface and from the liquid disks to
the bubble near the wall (see figure 19). Particles are centrifuged to the wall if they are heavier
than the liquid; they are centrifuged to the surface of the bubble is they are heavier than the air
but lighter than the liquid. Driven by the secondary motion, lighter-than-liquid particles
segregate in the liquid disks; heavier-than-liquid particles segregate in the region circling the
bubbles when the bubbles are off the wall, and in the liquid disks when the bubbles touch the
wall.
III-1. Bubbles in a partially filled rotating cylinder

Rimming flow is a coating flow inside a partially filled rotating horizontal cylinder. The
filling level F is the volume fraction of liquid in the cylinder; when F is large there is very little
air in the cylinder. If the filling fraction is not too small, which is characterized by β > β c in
expression (9) and the flat bump of the film profile depicted in figure 8(b) becomes unstable, air
bubbles will form and the shape, numbers and position of these bubbles depend on F, the angular
velocity Ω, the surface tension γ, the viscosity of the liquid, the density difference between liquid
and gas ρl − ρ g , and the dimensions of the apparatus. The qualitative effects of all these
parameters are fairly well understood.

When the cylinder rotates so fast that the effects of centrifugal gravity Ω2 R (R is the cylinder
radius) overwhelm those of terrestrial gravity, Ω 2 R g >> 1 , all of the liquid is centrifuged and
rotates with the cylinder as a rigid body; in this case the air is centered and if the filling level is
not too small, bubbles will form under the action of an interfacial potential (see Preziosi and
Joseph 1987). An important parameter for this potential is
( ρ l − ρ g )Ω2 Rb3
J= (20)
γ
where Rb is the maximum radius of the bubble. This parameter J does not depend on gravity,
viscosity, filling level or the length of the apparatus. If J < 4 cigar shaped bubbles will form; the
bubbles are all identical but the number of them depends on the filling level and the history of
their creation. J = 4 is a limiting value for the drop shape parameters; when Ω is increased, the
maximum radius of the bubble decrease in such a way that J = 4; when the ratio of bubble length
to radius L Rb > 8 , the bubble shape is very closely approximated by a cylinder of constant
radius Rb bounded by two semi-spherical end caps (this is nearly achieved in figure 17f).
Equation J = 4 was derived from heuristic arguments by Vonnegut (1942). It is the working
formula for “spinning drop” tensiometer which are used to measure interfacial tension (see
Joseph et al. 1992).
As Ω is increased, Rb decreases with J = 4. Since the bubble volume is fixed the length L
increases and eventually all the bubbles collect end to end to form a long cylindrical column,
rigorously centered, which does not change under further increases of angular velocity.
Coming the other way, decreasing from large values of Ω2R/g, the length of the bubbles will
decrease and the maximum radius Rb will increase with J = 4. Eventually, when Ω2 R g ~ O(1)
the effect of terrestrial gravity becomes important, the bubble rise; secondary motions are
generated and velocity becomes important. Photographs which exhibit typical regimes are
displayed in figure 17 where we compare soybean oil whose viscosity is 282 cp with Triton
mixture whose viscosity is 2950 cp. The main effect of viscosity here is to maintain rigid motion
of the fluid under perturbations with gravity. The radius of the cylinder used in our experiments
is 0.64 cm and the ratio of centrifugal to terrestrial gravity are listed below:
Ω 200 300 600 1000

Ω2R/g 0.287 0.64 2.57 7.16
Table 5. Gravity ratio for different angular velocity.
When Ω = 1000 rpm, the bubble is rigorously centered and extends from end to end of the
cylinder. When Ω = 600 rpm, the effects of gravity are sensible. The perturbation of rigid motion
in the high viscosity Triton mixture is small and the secondary motions are much weaker than in
soybean oil. The configuration in figure 17(f) is essentially uninfluenced by gravity with
centered bubbles whose shape is determined by a potential lined up end to end. At lower value of
Ω, the bubbles rise but the secondary motions which distort the bubbles are less important in the
high viscosity fluid.
(a) Soybean oil, Ω = 200 rpm (b) Triton mixture, Ω = 200 rpm
(c) Soybean oil, Ω = 300 rpm (d) Triton mixture, Ω = 300 rpm
(e) Soybean oil, Ω = 600 rpm (f) Triton mixture, Ω = 600 rpm
(g) Soybean oil, Ω = 1000 rpm (h) Triton mixture, Ω = 900 rpm
Figure 17. Comparison of soybean oil and Triton mixture under same conditions
As the angular velocity is decreased to zero the liquid and air stratify, with all the air at the
top. Even in this case, a stationary liquid, the air may separate into bubbles induced by capillarity
if the filling level is not very high; if the filling level is near 100%, the very small amount of air
will rise to the top and form a single short bubble due to the restraining action of surface tension.

Many unusual shapes of bubbles may occur when Ω2 R g ~ O(1), as put into evidence and in
the papers by Balmer (1970), Sanders, Joseph and Beavers (1981) and Preziosi and Joseph
(1988).
The combined effects of the filling level and rotational speed on the formation of the bubbles
are of interest. When the gravity ratio parameter is small, Ω2 R g << 1, the liquid and air are
stratified with a thin film being dragged up by the rotating cylinder. If the filling level is large
enough, the thickness of the film dragged up increases as the cylinder rotates faster. Up to a
critical condition, the thick liquid film on the top of the cylinder cannot be maintained and part of
it falls down under gravity, subsequently the single air bubble breaks. On the other hand, when
the gravity ratio parameter is large, the air forms a rigorously centered cylindrical column
stretching from end to end of the cylinder. If the rotational speed decreases, the stabilizing effect
of the centrifugal acceleration decreases. Up to a point, the combined effects of the surface
capillarity and the terrestrial gravity will break up the air bubble into smaller ones.
The critical conditions under which the single air bubble breaks were determined
experimentally for the rimming flow of soybean oil. The two lines on a (F, Ω) plane in figure 18
indicate the critical conditions. When the filling level and rotational speed fall in the region
between the two lines, the bubble breakup is observed; otherwise, a single air bubble is stable.
Note that when F < 0.4, a single air bubble is stable at any rotational speed.
800
Ω1
700
Ω2
600
500
Ω (rpm)
400
300
200
100
0
0.5 0.6 0.7 0.8 0.9 1
F
Figure 18. Critical conditions for the single air bubble breakup in soybean oil. The experiments were
carried out in a cylinder of glass with inside diameter 1.28 cm and length 22.14 cm. When F < 0.4, a
single air bubble is stable at any rotational speed. For F ≥ 0.5, when the rotational speed is between
Ω 1 and Ω2, air bubble breakup is observed.

III-2. Particles segregation due to bubbles
In a system with several smaller air bubbles distributed along the length of the cylinder and
displaced off the axis of the rotation by the action of gravity, the motion of the particles
suspended in the liquid may be driven by the secondary motions associated with the pumping
effect around the off-center bubbles. The liquid passes by the above-center bubbles which are
relatively stationary and are pushed away from the places occupied by the air. The pumping
motion of the bubble sets up an eddy which will push the liquid from the bubble to the liquid
disks near the bubble surface and from the liquid disks to the bubble near the wall (see figure
19).
Figure 19. The eddies set up by the pumping motion of the off-center bubble. Liquid flows from the bubble
to the liquid disks near the bubble surface and from the liquid disks to the bubble near the wall.

(a) Ω = 200 rpm
(b) Ω = 300 rpm
(c) Ω = 600 rpm
(d) Ω = 1000 rpm
Figure 20. Particle segregation of resin particles ρp = 1.13 g/cm, average diameter dp = 0.065 cm,
concentration 15% in soybean oil, filling level 60%.
The changes in the nature of heavier-than-liquid particle segregation as Ω2R/g changes (see
table 5) are shown in four panels of figure 20 where we go from stratified flow (a) to uniform
flow (d). The experiments were carried out in a cylinder of glass with inside diameter 1.28 cm
and length 22.14 cm. Comparing figure 17(a, c) with figure 20(a, b) we see that the particles
promote bubble formation at low speeds in soybean oil. This effect may be due to the increase in
the effective density of the mixture which increases the value of J in (20) by replacing ρ l with
ρ c = ρ pφ + ρ l (1 − φ ) where φ is the particle fraction.
The heavier-than-liquid particles are centrifuged and segregated near the cylinder wall where
secondary motions are weakest. The eddies push the particles on the wall to the region circling
the bubble and away from the gap between bubbles when the rotational speed of the cylinder is
large enough to centrifuge the air away from the wall but not so large as to center it (figure 20c).

At lower speeds, the bubbles rise all the way to the wall and the particles on the wall are pushed
to the space between bubbles (figure 20a,b).
In figure 21 we show that for heavier-than-liquid particles, segregation does not depend
sensitively on the type of particle.
(a) 16/30 AcFrac PR, Ω = 300 rpm (b) Resin particle, Ω = 300 rpm
(c) 16/30 AcFrac PR, Ω = 600 rpm (d) Resin particle, Ω = 600 rpm
Figure 21. Segregation of two different particles in soybean oil. The AcFrac PR particles are hydrophilic,
ρp =1.64 g/cm, average diameter dp = 0.088 cm; the resin particles are hydrophobic, ρp = 1.13 g/cm,
average diameter dp = 0.065 cm.
(a) 200 rpm
(b) 600 rpm
Figure 22. Particle segregation of resin particles ρp = 1.13 g/cm, average diameter dp = 0.065 cm,
concentration 8.96% in glycerin, filling level 66.7%.
Figure 22 shows the segregation pattern of particles which are lighter than the liquid. The
lighter-than-liquid particles are centrifuged to the surface of the bubbles. The eddies described in
figure 19 will push the particles to the space between air bubbles. Compare figure 20(c) with
figure 22(b), the different patterns should be noted. The heavier-than-liquid particles segregate in
the region above the bubbles, whereas the lighter-than-liquid particles segregate in the space

between bubbles. In figure 22(b), some of particles do circulate around the bubbles because the
bubbles are almost centered and the secondary motions are weak.
III-3. Segregation of bi-disperse suspension in a partially filled rotating

cylinder
Preliminary experiments using suspensions of particles with two different weights show that
the rotating flow leads to segregation of the two types of particles into separate regions whose
exact form depends on the weight and concentration of particles and on other features which
have yet to be determined. Here we show that this kind of segregation does occur and is robust.
Figure 23 and 24 show two experiments of segregation of bi-disperse suspension in aqueous
glycerin solutions. The different concentrations of the brown resin particles cause different
patterns of segregation (figure 23a and figure 24a). The rotating flow finally leads to uniform
distribution of particles, with the heavy particles at the end of the cylinder and the light particles
at the middle of the cylinder (figure 23c and figure 24c). Figure 25 shows bi-disperse suspension
in water. The configuration shown in figure 25 is stable for hours.
(a) 30 minutes after beginning of rotation.
(b) two hours after beginning of rotation.
(c) 20 hours after beginning of rotation.

Figure 23. Segregation of two types of particles in a 47.8% aqueous glycerin solution, ρl = 1.09 g/cm3.
The filling level of the liquid is 0.354. The black is a silicon particle with ρp = 3.07 g/cm3, average
diameter dp = 0.05 cm, and its concentration is 21.7%. The brown is a resin particle with ρp = 1.13
g/cm3, average diameter dp = 0.065 cm, and its concentration is 12%. The rotational speed is 165
rpm.

(a) four minutes after beginning of rotation.
(b) 24 minutes after beginning of rotation.
(c) 36 minutes after beginning of rotation.
Figure 24. Segregation of two types of particles in a 48.7% aqueous glycerin solution. The filling level of
the liquid is 0.328. The black is a silicon particle with ρp= 3.07 g/cm3, average diameter
dp = 0.05 cm, and its concentration is 16%. The brown is a resin particle with ρp = 1.13 g/cm3,
average diameter dp = 0.065 cm, and its concentration is also 16%. The rotational speed is 160 rpm.
Figure 25. Segregation of two types of particles in water. The filling level is 0.357. The black is a silicon
particle with ρp = 3.07 g/cm3, average diameter dp = 0.05 cm, and its concentration is 4%. The brown
is a resin particle with ρp = 1.13 g/cm3, average diameter dp = 0.065 cm, and its concentration is
16%. The rotational speed is 306 rpm.
IV. Conclusion
The principal facts concerning capillary attraction and self-assembly of small lighter- and
heavier-than-fluid floating particles were reviewed. These facts were applied to explain the
clustering and segregation of bands of particles in a thin liquid film rimming the inside of a
partially filled, slowly rotating cylinder in situations resembling those first observed by
Tirumkudulu, Mileo and Acrivos (2000). In our experiments clustering and band formation
occurred under all kinds of conditions, for lighter- and heavier-than-liquids, for small particles

and large particles, and for low concentrations and high concentrations of particles. Uniform
dispersions of particles in thin films are robustly unstable to anti-diffusion due to capillarity, and
clusters which are self-assembled are robustly stable. The conditions required to support this
phenomenon are that the liquid film is thin relative to the particle size; the film should be thin, or
in any case, not much thicker than the particles. The rotation speed of the cylinder should be
slow enough that the time needed for sensible capillary attraction is comparable to the time of
residence of the particle in the thin part of the rimming film.
Particle segregation may also be generated by pumping secondary motions of fluid by off-
center gas bubbles, which arise when the gravity parameter Ω2R/g ≤ O(1) and the filling level is
not too small. Lighter-than-liquid particles segregate in the liquid disks between bubbles;
heavier-than-liquid particles segregate in the region above the bubbles when they are off the
wall, and in the liquid disks when the bubbles touch the wall.
A third regime of segregation of bi-disperse suspension of particles of different heavier-than-
liquid weights, which stratify when the cylinder is at rest, form into rings when the cylinder
rotates. Different forms of the ring appear to depend on the particle concentration and other
factors which have as yet to be determined.
Acknowledgement
The work of authors Joseph, Wang, Bai and Yang was supported by the NSF/CTS under
grant 9873236, and by the Engineering Research Programs at the Office of Basic Energy Science
of the DOE.

References
Balmer, R.T. The hydrocyst—a stability phenomenon in continuum mechanics, Nature (London)
227, 600 (1970).
Bowden, N., I.S. Choi, B.A. Grzybowski, G.M. Whitesides, Mesoscale self-assembly of
hexagonal plates using lateral capillary forces: synthesis using the “capillary bond”, J. Am.
Chem. Soc. 121, 5373-5391. (1999).
Bowden, N., A. Terfort, J. Carbeck, G.M. Whitesides, Self-assembly of mesoscale objects into
ordered two-dimensional arrays, Science, 276, 233-235. (1997).
Brenner, H. and L.G. Leal. A micromechanical derivation of Fick’s law for interfacial diffusion
of surfactant molecules, J. Colloid Interface Sci. 65, 191 (1978).
Brenner, H. and L.G. Leal. Conservation and constitutive equations for adsorbed species
undergoing surface diffusion and convection at a fluid-fluid interface, J. Colloid Interface
Sci. 88, 136 (1982).
Chan, D.Y.C., J.D. Henry Jr. and L.R. White. The interaction of colloidal particles collected at
the fluid interface, J. Colloid Interface Sci. 79, 410 (1981).
Danov, K.D., R. Aust, F. Durst and U. Lange. Influence of the surface viscosity on the
hydrodynamic resistance and surface diffusivity of a large Brownian particle, J. Colloid
and Interface Science, 175(1), Oct 36-45 (1995).
Fortes, M.A. Attraction and repulsion of floating particles, Can. J. Chem. 60, 2889 (1982).
Gifford, W.A. and L.E. Scriven. On the attraction of floating particles, Chem. Engrg. Sci. 26,
287-297 (1971).
Goldman, A.J., R.G. Cox and H. Brenner. Slow viscous motion of a sphere parallel to a plane
wall—I Motion through a quiescent fluid, Chem. Eng. Sci. 22, 637-651 (1967).
Govindarajan, R., P.R. Nott, and S. Ramaswamy. Theory of suspension segregation in partially
filled horizontal rotating cylinders, Phy. Fluids, 13(12), 3517-3520 (2001).
Grzybowski, B.A., N. Bowden, F. Arias, H. Yang, G.M. Whitesides, Modeling of menisci and
capillary forces from the millimeter to the micrometer size range, J. Phys. Chem. B 105,
404-412. (2001).
Joseph, D.D., M. Arney, G. Ma, 1992. Upper and lower bounds for interfacial tension using
spinning drop devices, J. Colloid Interface Sci., 148(1), 291-294.
Karweit, M.J. and S. Corsin. Observation of cellular patterns in a partly filled, horizontal,
rotating cylinder, Phys. Fluids 18, 111 (1975).
Katoh, K., H. Fujita and E. Imazu. Motion of a particle floating on a liquid meniscus surface, J.
Fluids Engrg. 114, 411 (1992).
Kralchevsky, P.A., V.N. Paunov, N.D. Denkov, I.B. Ivanov and K. Nagayama. Energetical and
force approaches to the capillary interactions between particles attached to a liquid-fluid
interface, J. Colloid and Interface Sci. 155, 420-437 (1993).

Kralchevsky, P.A., V.N. Paunov, I.B. Ivanov and K. Nagayama. Capillary meniscus interactions
between colloidal particles attached to a liquid-fluid interface, J. Colloid Interface Sci. 151,
79 – 94 (1992).
Kralchevsky, P.A. and K. Nagayama. Capillary interactions between particles bound to
interfaces, liquid films and biomembranes, Advances in Colloid and Interface Sci. 85, 145-
192 (2000).
Majumdar, S.R., M.E. O’Neill, and H. Brenner. Note on the slow rotation of a concave spherical
lens or bowl in two immiscible semi-infinite viscous fluids, Mathematika, 21, 147-154
(1974).
Nicolson, M.M. The interaction between floating particles, Proc. Cambridge Philosophical Soc.,
45, 288 (1949).
Petkov, J.T., N.D Denkov, K.D Danov, O.D. Velev, R. Aust and F Durst. Measurement of the
drag coefficient of spherical particles attached to fluid interfaces, J. Colloid and Interface
Science, 172, 147-154 (1995).
Poynting, J.H. and J.J. Thompson. A Text-book of Physics: Vol. 1, Properties of Matter, C.
Griffith & Co. Ltd (London) 153-155 (1913).
Preziosi, L. and D.D. Joseph. The run-off condition for coating and rimming flows, J. Fluid
Mech., 187, 99-113 (1988).
Princen, H.M. Equilibrium shape of interfaces, drops, and bubbles. Rigid and deformable
particles at interfaces, Surface and Colloid Science, E. Matijevie, ed., Interscience, New
York, Vol. 2, p.1-84 (1969).
Rapacchietta, A.V. and A.W. Neumann. Force and free-energy analyses of small particles at
fluid interfaces: II. Spheres, J. Colloid and Interface Sci., 59(3), 555-567 (1977).
Redoev, B., M. Nedjalkov and V. Djakovich. Brownian motion at liquid-gas interfaces. 1.
Diffusion coefficients of macroparticles at pure interfaces, Langmuir, 8, 2962 (1992).
Ruschak, K.J. and L.E. Scriven, Rimming flow of liquid in a rotating horizontal cylinder, J.
Fluid Mech. 76, 113-127 (1976).
Saif, T.A., On the capillary interaction between solid plates forming mensci on the surface of a
liquid, J. Fluid Mech. in press (2002).
Schneider, Y.C., M.E. O’Neill, and H. Brenner. On the slow viscous rotation of a body
straddling the interface between two immiscible semi-infinite fluids, Mathematika, 20, 175
(1973).
Timberlake, B.D., J.F. Morris, Concentration band dynamics in free-surface Couette flow of a
suspension, Phy. Fluids, 14(5), 1580-1589 (2002).
Tirumkudulu, M., A. Tripathi, A. Acrivos. Particle segregation in monodisperse sheared
suspensions, Phy. Fluids, 11(3), 507-509 (1999).
Tirumkudulu, M., A. Mileo, A. Acrivos. Particle segregation in monodisperse sheared
suspensions in a partially filled rotating horizontal cylinder, Phy. Fluids, 12(6), 1615
(2000).

Tirumkudulu, A. Acrivos. Coating flows within a rotating horizontal cylinder: Lubrication
analysis, numerical computations, and experimental measurements, Phys. Fluids, 13, 3517
(2001).
Vonnegut, B., Rotating Bubble Method for the Determination of Surface and Interfacial Tension,
Rev. Sci. Instrum. 13, 6-9, (1942).
Wakiya, S. Niigata Univ. College of Engng. Res. Rept. 6, Nagoaka, Japan (30 March 1957).

Table of contents
Abstract ........................................................................................................................................... 1
I. Capillary forces....................................................................................................................... 1
I-1. Vertical forces..................................................................................................................... 2
I-2. Horizontal forces................................................................................................................. 4
I-3. Particle clustering................................................................................................................ 7
II. Particle aggregation in a liquid film rimming a rotating cylinder ........................................ 11
II-1. The ratio of the minimum film thickness to the particle diameter.................................... 12
II-2. Particle segregation in aqueous Triton mixtures............................................................... 15
II-3. Particle segregation in water............................................................................................. 21
II-4. Particle segregation in glycerin......................................................................................... 22
III. Particle segregation due to the formation of bubbles: .......................................................... 23
III-1. Bubbles in a partially filled rotating cylinder: .................................................................. 23
III-2. Particles segregation due to bubbles:................................................................................... 27
III-3. Segregation of bi-disperse suspension in a partially filled rotating cylinder.................... 30
IV. Conclusion ............................................................................................................................ 31
References..................................................................................................................................... 32

• Fluid Dynamics of Floating Particles
Singh and Joseph (2005) have developed a numerical package to simulate particle motions in
fluid interfaces. The particles are moved in a direct simulation respecting the fundamental equations of
motion of fluids and solid particles without the use of models. The fluid-particle motion is resolved by the
method of distributed Lagrange multipliers and the interface is moved by the method of level sets. The
present work fills a gap since there are no other theoretical methods available to describe the nonlinear
fluid dynamics of capillary attraction.
Two different cases of constrained motions of floating particles are studied here. In the first case,
we study motions of floating spheres under the constraint that the contact angle is fixed by the Young-
Dupré law; the contact line must move when the contact angle is fixed. In the second case, we study
motion of disks (short cylinders) with flat ends in which the contact line is pinned at the sharp edge of the
disk; the contact angle must change when the disks move and this angle can change within the limits
specified by Gibbs extension to the Young-Dupré law. The fact that sharp edged particles cling to
interfaces independent of particle wettability is under appreciated and needs study.
The numerical scheme presented here is at present the only one which can move floating particles
in direct simulation. We simulate the evolution of single heavier-than-liquid spheres and disks to their
equilibrium depth and the evolution to clusters of two and fours spheres and two disks under lateral forces
collectively called capillary attraction. New experiments by Wang, Bai and Joseph (WBJ 2003) on the
equilibrium depth of floating disks pinned at the edge are presented and compared with analysis and
simulations.

Contents
Abstract
I. Introduction
II. Floating particles which should sink
II.1 Floating particles with sharp edges
II.2 Gibbs inequality
II.3 Vertical force balance in equilibrium
II.4 Small particles, large particles and heavy particles

II.5 Experiments on floating disks pinned to the interface at the sharp edge (by J. Wang, R. Bai and
D.D. Joseph (WBJ) 2003)
III. Motions due to the capillarity of floating particles on liquid surfaces
IV. Governing Equations and dimensionless groups
V. Numerical Method
VI. Results
VI.1 Initial value problems for floating spheres
VI.2 Initial value problems for floating disks
VII. Conclusions
VIII. References
IX. Appendix

I. Introduction
In the work which follows, we will be considering the motions of particles which float in the
interfaces between two fluids. We shall sometimes describe the wettability properties of the particles as
hydrophobic or hydrophilic. The mathematical description of our problem in terms of air and water is
only a convention for the general problem of motion of particles in the interfaces between any two fluids.
It is well known that small tea leaves floating on the tea surface collect near the cup wall due to the
formation of a meniscus that rises near the wall and results in a net capillary force towards the wall. The
meniscus rises near the wall because the water wets the cup. If, on the other hand, the liquid does not wet
the cup, i.e., the meniscus falls near the cup wall, small floating particles tend to move away from the wall
and toward the center of the cup. Similarly, the deformation of liquid-liquid interfaces due to floating
light particles, or due to trapped heavy particles, gives rise to capillary forces on the particles which cause
them to cluster, as can be seen in figure I.1. The clustering of particles on interfaces is important because
it modifies the interfacial properties of the two-phase system and is used in many flotation based
extraction and separation processes (Gerson, Zaijc and Ouchi 1979). More recently, this effect has been
used for the self-assembly of submicron sized particles on two-liquid interfaces (see Bowden, et al.
1997,1999, Grzybowski, et al. 2001, and references therein).
(a)

(b)
(c)
Figure I.1. (WBJ 2003) Capillary attraction of floating particles. (a) Neutrally buoyant copolymer particles of
nominal diameter 0.1 cm cluster in water/air interface. (b) Heavy aluminum disks (short cylinders with circular
cross sections) hanging in water/air interface at the sharp rim. The distributions of 14 particles at 0 second (left),
after 60 seconds (middle) and after 200 seconds (right) are shown. The diameter of the disks is 0.3175 cm and its
height is 0.15875 cm. (c) Heavy aluminum bricks with square cross sections hanging in water/air interface at the
sharp corners. The distributions of 14 particles at 0 second (left), after 142 seconds (middle) and after 220 seconds
(right) are shown. The dimension of the bricks is 0.3175 cm × 0.3175 cm × 0.15875 cm. The attractive power of
capillarity on floating particles is very long range and the accelerations in the final stage of clustering are
exceedingly large. Movies of these experiments can be viewed at
http://www.aem.umn.edu/research/particles/floating/.
The motion of tea leaves towards or away from the wall, in the above example, is entirely due to the
deformation of the meniscus near the cup wall. The clustering of particles, on the other hand, is a
consequence of the interface deformation caused by neighboring particles. Specifically, when two heavy
hydrophobic spheres are close to each other the deformed interface around the spheres is not symmetric
because the interface height between the spheres is lowered by the capillary force; on the other hand,
lighter-than-water hydrophilic spheres will rise as shown in figure I.2. In both of these cases, the lateral
component of interfacial tension is attractive and the spheres tend to cluster. But, when one sphere is
hydrophilic and the other is hydrophobic, the lateral force at short range is repulsive and tight clusters
cannot form.

Figure I.2. Spherical particles in water. (a) heavier-than-water hydrophobic spheres. The meniscus between the
spheres is below the undisturbed level. Assuming that the contact angle remains fixed, the horizontal component of
capillary force moves them toward each other. (b) Lighter-than-water hydrophilic spheres will rise into the elevated
section of the meniscus and come together.
The literature on capillary attraction is cited by Kralchevsky and Nagayama 2000 and Saif 2002, here
in section III, and in the paper on capillary attraction of particles embedded in a thin film rimming the
inside of a rotating cylinder by Joseph, Wang, Bai, Yang and Hu 2003. These works do not treat the case
of capillary attraction of particles pinned to the interface at a sharp edge which is one of the main subjects
in this paper.
Problems of evolution to equilibrium of heavier-than-liquid floating particles may be studied by direct

numerical simulation (DNS); this simulation method fills a gap identified by Gifford and Scriven 1971
who note that
“casual observations… show that floating needles and many other sorts of particles do indeed come
together with astonishing acceleration. The unsteady flow fields that are generated challenge analysis by
both experiment and theory. They will have to be understood before the common-place ‘capillary
attraction’ can be more than a mere label, so far as dynamic processes are concerned.”
The basic facts about the equilibrium of single particles are discussed in section II and new
experiments on the equilibrium depth of disks pinned at their edges are presented. The prior literature on
capillary attraction is briefly reviewed in section III. In section IV we set out the equations which govern
the motions of floating particles and introduce the basic dimensionless groups which characterize these
motions. In section V, we outline the numerical method stressing only those details which are new.
Readers interested in constructing or improving the numerical algorithm used in this study can find a
detailed description in the appendix. In section VI, we compute the solutions of the initial value problems,
starting from rest, for one, two and four spheres with contact angle prescribed. In section VII, we compute

the solutions of the initial value problems, starting from rest, for one and two disks pinned at their sharp
edges. The solutions are compared with experimental data.
II. Floating particles which should sink
In this section we consider the forces that determine the equilibrium depth of a floating particle.
Princen 1969 gave an excellent analysis of this problem for a sphere and prismatic particles with sharp
edges. Keller 1998 generalized this analysis for smooth bodies of arbitrary shape. Kotah, Fujita and Imazu
1992 used the floating ball to measure contact angles.
Floating particles which should sink are held up by capillary forces at the line of contact of the
three phases on the particle surface. The hanging depth between the contact line and the highest point on
the meniscus depends on whether the meniscus attaches to the particle on the smooth surface with
uniquely determined normal or at a corner or edge where the normal is undefined. Here we show that the
hanging depth is determined by the position of the contact line on a floating sphere when the contact
angle is fixed by the Young-Dupré law, and by the value of the contact angle which changes with the
weight of the particles when the contact line is pinned at a sharp edge.
II.1 Floating particles with sharp edges
It is well known, but not well understood, that liquid-air-solid interfaces tend to locate at sharp edges.
This mechanism allows a prismatic disk or cube to float with contact line pinned to its sharp rim. Even
when a downward vertical force is applied by adding weights onto the top surface of a floating disk, as
discussed below, the contact line remains pinned to the rim.
Obviously, a prismatic particle which is denser than the liquid below can float only if the vertical
component of interfacial tension is sufficiently large to balance its buoyant weight and will sink when this
is no longer true.
C D α
α
A B

Figure II.1. The vertical component of capillary force for the disk does not change when the contact line moves
from AB to CD, for two different floating heights, because the contact angle α is fixed. For a sphere, the vertical
component of the capillary force changes as the contact line moves on its surface.
The effects of the buoyant weight may be isolated in the case of a circular cylinder or disk, with axis
vertical, which is suspended with the contact line on a circle perpendicular to the cylinder generator (see
figure II.1). The contact angle is fixed by the Young-Dupré law and does not change even as the contact
line sinks due to change in the cylinder buoyant weight. The cylinder can be denser than the liquid
provided that the vertical capillary force is just large enough to balance its buoyant weight. If the
cylinder’s weight is increased, it will sink further and the contact line on the smooth surface will move
upwards. But, the vertical component of capillary force will not change because the angle between the
interface and the horizontal, which only depends on the contact angle, does not change when the cylinder
sinks (see figure II.1). Consequently, the maximum interfacial deformation, the vertical distance between
the contact line and the highest point on the meniscus, will also not change as the cylinder sinks. The
buoyancy force acting on the disk, however, increases, as it sinks into the liquid below. Disks of different
weight in air, with same contact angle and buoyant weight can be suspended as in figure II.1.
ψ1 ψ1 ψ 2 >ψ1 ψ3 ≈ 900
W1
W1
W2
W3
(a) (b) (c) (d)
Figure II.2. Effect of changing the buoyant weight on the contact angle at the rim of a cylinder. The contact
angle is the same for a sphere (a) and disk (b) when the buoyant weights are the same. Increasing the buoyant
weight leads to larger contact angles which have larger vertical components of the capillary force as shown in (b),
(c) and (d). In the experiments of (WBJ 2003) the cylinder would sink when ψ ≤ ~ 90 0 ; however theoretically the
cylinder can float with ψ > 90 0 (Hesla and Joseph 2003).

At a critical value of the disk weight, the contact line moves from the smooth surface to the sharp
edge. If the disk weight is increased further, the contact line remains pinned at the sharp corner for a
range of weight, even though the disk continues to sink further (see figures II.2-II.5). A heavier-than-
liquid disk can float with the interface pinned to the sharp edge, as in figure II.2, provided the vertical
component of the capillary force is large enough to balance its buoyant weight. In this paper, we will
study the dynamics of floating disks in this state.
(a) (b)
Figure II.3. (WBJ 2003). Two photos of floating Teflon disks of density ρs=1.4g/cc held at the contact line in water
of density ρf=1g/cc. Both disks have a diameter of 0.8 cm; the height from the bottom of the disk to the contact line
is 0.4cm in (a) and 0.8 cm in (b). The contact angle in (b) is larger than that in (a) in order to satisfy the force
balance. The image of the disk projecting above the contact line is a reflection in the surface of the water.
(a) (b)
(c)

Figure 4 (WBJ 2003). (a) The meniscus for a Teflon cylinder of density ρs=1.4g/cc hanging from a flat edge in
water. (b) An aluminum plate can float in water hanging from the sharp edge; when weighted by a Teflon ball, the
plate still floats but the hanging depth increases. (c) A floating glass plate is held at the sharp edge in water.
Spheres of aluminum and glass will sink in water, provided that the spheres are not so small that the surface tension
will dominate the buoyant weight. The contact angle on the hydrophilic glass plate and the hydrophobic Teflon plate
is determined by their buoyant weight and not by wettability.
(a) (b)
Figure II.5 (WBJ 2003). (a) A cartoon for the experiment determining the critical contact angle at the sharp edge.
See section II.5 for details of the experiment. (b) A photo from the video showing that the contact angle reaches 90°
at a moment just before the disk sinks. The square, solid black part in the photo, is the disk and the bright part is
water.
II.2 Gibbs Inequality

This pinning of the contact line at the disk edge appears to be conflict with the Young-Dupré law
which states that the equilibrium contact angle between a liquid, a gas and a solid wall is constant
γLG cos α = γSG – γSL,
where α is the contact angle and γLG, γSG and γSL are the interfacial energy between liquid and gas,
solid and gas, and solid and liquid, respectively. To ensure that the equilibrium contact angle is fixed,
when the interface at a small distance away from the contact line moves the contact line must also move.
But, since the normal at the corner is not defined, Young-Dupré law is not violated provided the contact
angle α at the corner, as shown in figure II.6, stays within the range specified by the Gibbs extension to
Young-Dupré law:
α0 < α < 180-φ + α0
where φ is the wedge angle and α0 is the equilibrium contact angle for the vertical face (see Gibbs
1906 and Princen 1969).

180-φ+α0
α0
φ
Figure II.6. Two limiting angles for the Gibbs extension to Young-Dupré law which states that the contact angle
α at the sharp edge can take any value between α0 and 180-φ + α0.
II.3 Vertical force balance in equilibrium

The analysis of the forces which keep a sphere suspended in the interface between fluids was given
first by Princen 1969, then by Rapacchietta and Neumann 1977, and Kotah, Fujita and Imazu 1992, who
used the floating ball to measure contact angles. A detailed discussion of the vertical balance of a ball in
equilibrium can be found in Joseph et al. 2003. The analysis of the forces which keep a heavy disk
suspended in the interface at the sharp upper rim of the disk was given by Hesla and Joseph 2003,
following an earlier analysis of Princen 1969 for a prismatic particle.
For equilibrium, the buoyant weight of particle must be equal to the vertical component of the
capillary force. If the particle density is larger than that of both fluids, equilibrium is possible only when
the particle is hydrophobic and the vertical component of capillary force is large enough to balance its
buoyant weight. The interface shape in this case is concave down and the net capillary force acts against
gravity.
Force balance for a sphere

The conditions for equilibrium of a floating sphere can be framed with the help of the cartoon in
figure II.7. The vertical component of capillary force FC depends on the particle radius R, the surface
tension coefficient γ , the filling angle θ c and the contact angle α , and is given by
FC = 2π(R sin θc )γ sin[θc − (π − α )] = −2π(R sin θc )γ sin (θc + α ) . (II.1)
The above expression holds for both the hydrophobic and hydrophilic cases.

Hydrophobic particle Tangent to
point of
contact
Air Va Tangent to
meniscus
R
θc
Water θ1 h2
Vw α
g Contact
angle
Hydrophilic particle Tangent to point of contact
Air Va
R h2
Vw θc α
Water Tangent to
θ1
meniscus
Figure II.7 (Joseph et al. 2003). Hydrophobic and hydrophilic particles in equilibrium. The position of the contact
ring determines the angle θC. The point of extension of the flat meniscus on the sphere determines the angle θ1 .
h2=R (cosθc – cosθ1).
The weight mg of a heavy particle in equilibrium is balanced by a capillary force Fc and net
pressure force Fp satisfying:
FC + Fp = mg , (II.2)
where Fc is given by II.1. Fp is the pressure force given by
θc
Fp = ∫ p cos θ (2πR sin θ) Rdθ (II.3)
0

2 1 2 1
= ρ L gπR 3 ( − cos θ c + cos 3 θ c ) + ρ a gπR 3 ( + cos θ c − cos 3 θ c ) −
3 3 3 3
2 2
(ρ L − ρ a )gh 2 πR sin θ c
where h2 is the meniscus height, and ρL is the density of the lower liquid and ρa is the air density.
Substituting into (II.2) we get
2 1
2πγ (R sin θ c )sin(θ c + α ) − ρ L gπR 3 ( − cos θ c + cos 3 θ c ) −
3 3
(II.4)
2 1
ρ a gπR 3 ( + cos θ c − cos 3 θ c ) + (ρ L − ρ a )gh 2 πR 2 sin 2 θ c = mg
3 3
4
with m = πR 3 gρ p , the above expression may be expressed in a dimensionless form as
3
1 4 2 1
sin θ c sin(θ c + α) = − B[ l1 − ( − cos θ c + cos 3 θ c ) −
2 3 3 3
(II.5)
2 1
l 2 ( + cos θ c − cos 3 θ c ) + (1 − l 2 )(cos θ1 − cos θ c ) sin 2 θ c ]
3 3
where B = ρLR2g/γ is the Bond number and l1 = ρp /ρL and l2 = ρa /ρL are the density ratios.
The capillary force acts against gravity only when θ C + α − π is positive, in which case
sin (θ C + α ) <0, otherwise it acts in the same direction as gravity. For example, if α = 3π 4 , a heavy
sphere will float with θ C > π / 4 . For α = 3π / 4 and θ C = π / 4 , the force FC is zero and there is no
interface deformation (see figure II.8). Fc increases when θ C is increased from π 4 and reaches its
maximum value at θ C ≈ 1.9 and then decreases with increasing θC . On the other hand, when the contact
angle is π, FC is always non-negative and its maximum value is for θ C = π / 2 , i.e., the sphere half
immersed in the lower liquid. The buoyant weight of the particle, of course, also changes with θC .

0.2
0
0 1 2 3 4
-0.2
-0.4
-0.6
-0.8
-1
Figure II.8. For the contact angle α = 3π / 4 , sin (θc )sin(θc + α ) is plotted as a function of θC . Notice that
the vertical component of capillary force FC given by (II.1) is negative for θC < π / 4 and maximum for θ C ≈ 1.9
radians=~109º.
Force balance for a disk

The force balance for the disk is given by (II.2). From figure II.9 it is clear that
Fc=2πRγ sinψ (II.6)
Fp = (P0 – Pa)πR2 = ρ L g(h + h 2 ) πR2 (II.7)
Pa
ψ
z
h2
h
P0
R
Figure II.9. Heavier than liquid disk hanging from a flat edge. The capillary force is given by Fc=2π Rγ sinψ, where
γ is the interfacial tension. The meniscus is z = h(r); h(∞)=h2 is the highest value of z on the meniscus. Pa is air
pressure and P0 is the pressure at the bottom of the disk z = -h. The disk may be weighted by heavy balls in the cone
shaped cavity, increasing h2 and ψ without sinking.
Substituting into (II.2), we get
2πRγ sin ψ + ρLgV + πR2h2ρLg = mg, (II.8)

where V=πR2h is the volume of the disk. The angle ψ=α-90, where α is the contact angle, is measured
from the horizontal. The dimensionless form of (II.8) is given by
1⎡ h h ⎤
sin ψ = B ⎢(l1 − 1) −
2
. (II.9)
2⎣ R R ⎥⎦
The meniscus height h2 is determined from the solution of the meniscus equation
′
γ ⎡ ⎤
rh ′( r )
ρgh[h( r ) − h2 ] = ⎢ ⎥ , (II.10)
r ⎢ 1 + h ′( r ) 2 ⎥
⎣ ⎦
where the origin (z, r)=(0,0) is in the plane at the center of the circle of radius R defined by the
contact line. The integration starts at (z, r)=(0,R) where
h’(r)=tan ψ. (II.11)
Far from the particle, the meniscus is flat and
lim {rh ′(r ), h (r )} = {0, h 2 } . (II.12)

r →∞
For a cylinder, the values of ψ and h2 can be determined from the solution z=h(r), using (II.8)
together with
∞
ρg ∫ [h 2 − h (r )] r dr = γR sin ψ (II.13)
R
which follows from (II.9), (II.11) and (II.12).

Hesla and Joseph 2003 worked an exact numerical solution of the problem just considered; they
gave a simple mathematical argument that as the weight of the floating disk is gradually increased (figure
II.2), the maximum contact angle at the sharp rim which is attained before the disk sinks is greater than
90º. They presented numerical results which support this conclusion. Though such solutions are allowed
by the equilibrium analysis, they have not been observed. It may be that configurations with contact angle
greater than 90º are unstable (see figure II.5).
II.4 Small particles, large particles and heavy particles

The left side of equation (II.5), and thus also the right side, lies in the range
− 1 ≤ sin θ C sin (θ C + α ) ≤ 1 . Obviously, (II.5) cannot be solved if B is too large which may be the case
when the sphere is too heavy or too large. Similarly, for a floating disk if B is too large, (II.9) cannot be
solved; the disk will sink when the capillary force is not large enough to balance its buoyant weight.

θc
θc
Small hydrophobic sphere Small hydrophilic sphere
Small hydrophobic disk Small hydrophilic disk
Figure II.10. The deformation of the interface due to floating sufficiently small spheres or disks is negligible. A
small hydrophobic sphere will float with θC ≈ π − α so that interfacial deformation is negligible even if it is denser
than the liquid below. A small less dense hydrophilic sphere also does not deform the interface. Similarly, a small
dense hydrophobic disk floats on the surface with negligible penetration into the liquid.. A small hydrophilic disk
which is less dense than the lower liquid does not deform the interface and it is kept inside the lower liquid by the
capillary force which acts downwards.
As R approaches zero, the capillary force, which varies linearly with R, dominates the buoyant
weight of the sphere which varies with R3. In this limit since the Bond number B= ρ L R 2 g γ → 0 , the
right hand side of (II.5) is zero and thus sin (α + θC) ≈ 0 or θC ≈ π − α (see figure II.10). We may
therefore conclude that heavy small particles can be suspended without causing significant interfacial
deformation when B is small. Krahshesky, et al. 1992, 1993 noted that for particles floating on water this
limit is approximately reached when their diameter is 10 µm. Hence, the lateral capillary forces, which
arise from interfacial deformation, are also insignificant when the particle diameter is smaller than 10 µm.
Similarly, if volume of the disk approaches zero, the capillary force, which varies linearly with R, will
dominate its buoyant weight which varies as hR2. In this limit, the right hand side of (II.8) is zero and
sinψ ≈ 0 or ψ = 0 . The disk therefore does not deform the interface, and floats with its top surface in the
plane of the interface (see figure II.10).
The vertical component of the capillary force for the two positions in figure II.1 can be zero only if
the contact angle α is 90º. This implies that a small hydrophobic (α>90º) disk must float on its bottom
edge, as shown in figure II.10; it cannot be suspended as in figure II.1. When the contact line is pinned to
an edge the contact angle can take any value, between the two values specified by Gibbs inequality. The

argument just given applied to all cases in which the Bond number is small, to particles with other shapes,
like cubes and to lighter and larger particles.
When particles are partially immersed in a thin liquid film and their weight is supported by the
substrate below, the arguments just given are not applicable and the interface deformation can be
significant even for small particles. Kralchevsky and Nagayama 2000 have shown that in thin films the
particle-particle attraction force increases with decreasing particle size.
II.5 Experiments on floating disks pinned to the interface at the sharp edge (WBJ 2003)
WBJ used a 3.38g Teflon cylinder with a cone cut in the center. 0.25 g steel beads were put in the
cone to change the weight (see figure II.9). The radius, height and volume of the disk are [1.27 cm, 0.495
cm, 2.51cc]. The angle ψ and the depression h2 were measured using a video camera. Measurements were
taken at several azimuthal positions and the average value ψ and h2 recorded. After inserting the
measured parameters into the force balance equation (II.8), the difference between the measured vertical
force and the particle weight, the residual e was computed:
e = mg − 2πRγ sin ψ − ρ l gπR 2 (h + h 2 ) (II.16)
m(gram) 3.38 3.63 3.88 4.13 4.38

ψ (degree) 28.4 37.8 43.0 51.7 71.1
H2(cm) 0.130 0.176 0.206 0.255 0.302
2πRγ sinψ/mg(%) 5.27 6.31 6.58 7.11 8.08
ρ l gh2πR2/mg(%) 19.44 24.54 26.94 31.34 34.98
|e|/mg(%) 1.03 0.01 1.8 0.77 0.36
Table II.1 (WSJ 2003). Quantities entering into the force balance equation (II.8). The residual e is computed from
(II.16). The values of e are small.
Table II.1 shows that the contact angle at the rim increases when the weight of the particle is
increased. A maximum weight can be held in this manner; beyond this weight the particle will sink. WBJ
did experiments to determine the critical contact angle corresponding to this maximum weight. The 3.38g
Teflon disk with a cone cut in the center was used. The contact angle was gradually increased by pushing
the disk down into the water with a needle (see figure II.5). A video camera was used to record the whole
process and the critical contact angle was determined using the video replay. The contact angle increased
up to 90° while the contact line was pinned at the rim (see figure II.2); when the needle was pushed
further down, the contact line moved away from the sharp edge to the flat top of the disk, and the disk

sank instantaneously. They concluded that the critical contact angle corresponding to the maximum
weight which could be held at the sharp edge is 90°. Hesla and Joseph 2003 have shown that that the
equilibrium solution for this problem allows contact angles larger than 90º; the vertical component of the
capillary force decreases while the buoyant force increases maintaining the balance. These larger contact
angles may be unstable.
III. Motion due to the capillarity of floating particles on liquid surfaces
The deformation of a fluid-fluid interface due to floating or trapped particles gives rise to lateral
capillary forces. A simple explanation is given in figure I.2. A heavier-than-liquid particle will fall down
a downward sloping meniscus while an upwardly buoyant particle will rise.
There are several ways to isolate the effects of capillarity uninfluenced by gravity (see Joseph et al.
20003). Poynting and Thompson 1913 investigated the capillary effect by considering two vertical plates
immersed in a liquid, the space between the plates is a two dimensional capillary tube. If the plates are
hydrophobic, the level in the capillary gap sinks below the liquid outside; if the plates are hydrophilic the
levels will rise. Another way to take away the effects of gravity is to support the particles on a substrate.
In this case, the horizontal forces are due to capillary effects alone. Katoh, Fujita and Imazu 1992 studied
the motion of a particle floating on a liquid meniscus surface which could be interpreted as motion on a
substrate because the foaming phlystyrol particles used by them are an order of magnitude lighter than
water, and minimize the effects of gravity compared to capillarity. Their experimental results are
completely consistent with the predictions of Poynting and Thompson: when the sphere and the wall are
alike with respect to wetting, say both are hydrophobic or hydrophilic, the wall and sphere attracts; when
they are unlike the sphere and wall repel.
There are only a few theoretical studies of capillary attraction. Nicolson 1949 was the first to derive
an analytical expression for the capillary force between two floating bubbles by using the superposition
approximation to solve the Laplace equation of capillarity. A similar approximate method was applied by
Chan et al. 1981 to floating spheres and horizontal cylinders. For horizontal cylinders the alternative
approaches were proposed by Gifford and Scriven 1971 and by Fortes 1982. The theoretical works are
based on solutions of the Laplace equations for capillary menisci with translational or rotational
symmetry, where the Laplace equation reduces to an ordinary differential equation. Saif 2002 constructed
an interesting analysis of the capillary interaction of long plates with round ends at prescribed heights
which do not float.

For the case where the meniscus slope and the particle size are small, the Laplace equation for the
interface shape was solved using bipolar coordinates by Krahshesky, et al. 1992, 1993. This solution
provides expressions for calculating the capillary meniscus force between two vertical disks, between two
spheres partially immersed in a liquid layer and between a vertical disk and a sphere. Specifically,
Kralchevsky and Nagayama 2000 have shown that the lateral force Fl acting on particles of radii R1 and
R2 separated by distance L is equal in magnitude and opposite in sign and is given by
[ ]
Fl = −2πQ1Q 2 qK 1 (qL) 1 + O(q 2 R 2k ) when L >> rk. (II.17)
Here rk = R k sin(θ c ) , k=1, 2 are the radii of the two contact lines as shown in figure II.7 (where the
particle radius is assumed to be R), Q k = rk sin ψ k , where ψ k is the interface slope with the horizontal
plane at the point of contact, q = (ρ l − ρ p )g γ is the inverse of the capillary length, K1(x) is the
modified Bessel function of the first order. (II.17) is valid for particles much smaller than the capillary
length. The force acting between two floating particles decreases with increasing distance them.
The analysis just given is useful for determining the parameter values for which the particles can
remain trapped on two-fluid interfaces, as well as the sign and magnitude of forces that act between two
suspended particles, but to understand the actual motion of particles on the interface we must solve the
governing equations of motion. Since the governing equations are complex, the dynamic behavior of fluid
and particles is not well understood.
A small number of theoretical studies have looked at the drag and diffusion coefficient of a spherical
particle attached to a fluid interface (Brenner and Leal 1978, 1982; Goldman, Cox and Brenner 1967;
Schneider, O’Neill and Brenner 1973; Majumdar, O’Neill and Brenner 1974 and Wakiga 1957; Redoev,
Nedjalkov and Djakovich 1992; Danov, Aust, Durst and Lange 1995). Brenner and Leal have shown that
the drag FD acting on a floating sphere in the zero Reynolds number limit is FD = 3πηDU x f D , where Ux is
the lateral velocity of the sphere, D is the diameter, fD is the drag coefficient which is O(1) and depends
on the ratio of viscosities of the upper and lower fluids.
The only experimental study for determining drag coefficients of floating particles is by Petkov,
Denkov, Danov, Velev, Aust and Durst 1995. They calculated the drag coefficients for particles of sub-
millimeter radius by measuring the particle velocity under the action of a well defined external force.
They showed that the capillary interactions are quite strong and very long range. Danov, Aust, Durst and
Lange 1995 performed numerical simulations to obtain the drag coefficients for floating spheres, but they
assumed that the interface between the two fluids stays flat and the particle translates with a constant
velocity along the interface.

To understand the dynamics of clustering and self assembly of particles due to capillarity, we
have developed a numerical package which treats the problem by direct numerical simulation. The
method is as exact as numerical methods allow; in particular, the changing shape of the meniscus and the
hydrodynamic forces which move particles are computed and not modeled. At each time step, we solve
the governing mass and momentum conservation equations for the two fluids, compute the forces acting
on the particles and then move them using Newton’s equations for rigid solids. The interface shape
changes in response to the fluid motion while satisfying the contact angle or contact line requirement on
the particle surface. In addition, across the interface the fluid properties change suddenly and a capillary
force acts between the two fluids.
We have performed dynamic simulations of spherical particles for which the contact angle is
maintained at the equilibrium value and the position of the contact line changes, as well as for floating
disks with sharp edges. For floating disks, the meniscus remains pinned at the rim even when the disk
moves relative to the interface, but the contact angle at the rim changes. In our numerical study it is
assumed that the interface is initially flat and the top surface of the disk is in the plane of the interface. As
the disk is denser than the liquid, it sinks but the contact line remains at the rim. Consequently, the
interface near the rim becomes more vertical increasing the vertical component of the capillary force. In
our code, the contact line is kept at the sharp edge of a floating cylindrical particle by making the level set
function vanish on the rim.
In the next section we will state the governing equations for the fluids and the particles, briefly
describe the level set and distributed Lagrange multiplier approaches and present our finite element
method. A detailed description of the numerical method is included in the appendix. In section VI, we
will discuss the convergence study that shows that the numerical results are independent of the mesh size
as well as the time step size and present results for the transient motion of particles along two-fluid
interfaces.
IV. Governing Equations and dimensionless groups

In our numerical studies of particle motion in two-fluid interfaces we will assume that the fluids
are immiscible and Newtonian. The particles are assumed to be rigid. Let us denote the domain containing
the two liquids and N particles by Ω, the domain boundary by Γd, and the interior of the ith particle by
Pi(t). The governing mass and momentum conservation equations for the fluid phases can be written as
⎡ ∂u ⎤
ρ ⎢ + u. ∇ u ⎥ = ρ g - ∇p + ∇.σ +γ κ δ(φ) n; ∇. u = 0 in Ω\ P(t) (IV.1)
⎣ ∂t ⎦
u = uL on Γ d (IV.2)

u=U+ ωxr on ∂P( t ) (IV.3)
with the extra stress tensor σ =2ηD, ρ is the fluid density which is different for the two fluids, p is the
pressure, D is the symmetric part of the velocity gradient tensor, δ(.) is the Dirac delta function, n is the
outer normal at the interface, γ is the surface tension, κ is the mean surface curvature, φ is the distance
from interface, η is the viscosity which is different for the two fluids and uL is the prescribed velocity on
Γd. The surface tension force acts along the interface between the two fluids.
The particle velocity U and angular velocity ω are governed by
dU
M = Mg + F (IV.4)
dt
d(I P ω )
=T (IV.5)
dt
U | t =0 = U 0
ω | t =0 = ω 0 (IV.6)
where M and IP are the mass and moment of inertia of the particle. The particle density is denoted by ρ P .
The force F acting on a particle in the above equations is
F = ∫ (−pI + σ).n dA + ∫ Γ ds (IV.7)
C. L .
The first term on the right of (IV.7) is the force on the particle due to stresses generated by fluid motion;
the second term
∫ Γ ds = γ ∫ n c ds (IV.8)
C.L. C.L.
is the capillary force, Γ = γn is a line stress on the contact line (C.L.) and nc is the capillarity unit
c
vector which lies in the interface and is normal to the contact line. This unit vector gives the direction of
the action of the capillary force. A numerical algorithm for constructing nc is given in section 5.
Similarly, the torque T acting on the particle is given by
T = ∫ (x − X) × [(−pI + σ).n ] dA + ∫ (x − X) × Γ ds . (IV.9)
C.L.
Here X is the center of particle, the first term gives the torque due to the fluid stress and the second due to
the capillary force acting on the contact line. For a spherical particle, which is one of the cases considered
in this paper, if the interfacial tension γ is constant, the torque due to the interfacial tension is zero (see
Singh and Hesla 2003).
The shape of the meniscus must be compatible with conditions which are prescribed at the
contact line on every particle and at remote boundaries; for spherical particles the contact angle α is
prescribed (see figure II.7), but the contact line evolves during motion. For disks hanging at the sharp

edge, the position of the contact line is prescribed and the contact angle changes. At remote boundaries
different conditions could be considered, but in our simulations we have required the interface to be flat
there. The motion of particles in fluid interfaces is very complex because the prescribed value of the
contact angle is to be applied at the contact lines whose positions cannot be prescribed a priori and at the
sharp edges the contact angle is not known a priori.
A particle placed in a two-fluid interface can be in a state of equilibrium provided its buoyant
weight is equal to the z-component of capillary force. The capillary force changes when the particle sinks
or rises or the interface deforms to satisfy the contact angle requirement. Clearly, for a particle moving
laterally along the interface, the vertical acceleration is small, and thus the z-component of (IV.4) is
0 = − Mg + k ⋅ ∫ ( − pI + σ ).n dA + γ ∫ k ⋅ n c ds , (IV.10)
C. L.
where k is the unit vector in the z-direction. The last term of (IV.14) is the vertical projection of the
capillary force which depends on the contact angles. For isolated spheres or disks in equilibrium (IV.10)
and the vertical projection of (IV.1) with u=0 reduce to equation (II.5) or (II.8).
The x-component of particle momentum equation, which governs its lateral motion, can be
written as
dU x
M = i ⋅ ∫ ( − pI + σ ).n dA + γ ∫ i ⋅ n c ds . (IV.11)
dt CL.
where i is the unit vector in the x-direction. The first term on the right hand side is the x-component of
fluid stress and the second is the x-projection of the integral of n c around the contact line.
If we assume that a particle is accelerating slowly, which is the case, for example, when the two
attracting particles are far from each other, then the two terms on the right hand side of (IV.11) balance
each other. In the low Reynolds number limit, Brenner and Leal expressed the drag FD acting on a sphere
moving along the interface as
FD = 3πη L DU x f D (IV.12)
where fD is the drag coefficient which is of order one and depends on the viscosity ratio of the two fluids,
the contact angle and the deformation of the interface which in turn depends on the density of the particle.
Under these approximations, equations (IV.11) and (IV.12) give
0 = 3πηL DU x f D + γ ∫ i ⋅ n c ds . (IV.13)
CL.
Equation (IV.13) can be solved to obtain the lateral velocity U x of the particles; Ux is proportional
γ
to , the “capillary velocity” scale.
ηL

Petkov, Denkov, Danov, Velev, Aust and Durst 1995 used (IV.13) for estimating the drag
coefficient for floating spherical particles attracted by a plate. They measured U x in an experiment and
used the analytical expression for the horizontal force obtained by Kralchevsky, Paunov, Denkov, and
Nagayama 1994 which is related to the integral term in the above expression. They found that the drag
coefficient depends on the viscosities of the upper and lower fluid, as was shown by Brenner and Leal.
The experimental values of the drag coefficient for several fluid-particle combinations were found to be
of O(1). The drag coefficient was greater than one for heavy particles, since they cause a greater
deformation of the interface. They estimated the drag coefficients when the distance between the particle
and the plate was greater than 35R, where R is particle radius; for smaller distances (IV.17) is not
accurate because the inertial effects are not negligible. The estimate of the lateral capillary force they used
is accurate only when the distance between the particle and the plate is large.
Danov, Aust, Durst and Lange 1995 performed numerical simulations to study the dependence of
the drag on a spherical particle translating in the interface on the ratio of viscosities. In their simulations,
it is assumed that the interface between the two fluids is flat and the particle velocity is constant. They
found that the agreement with experiments deteriorates with increasing particle density because interfacial
deformation is not negligible.
In this paper we study problems for which inertial effects and time dependent changes in the
interface shape in response to particle motion are important. This happens to be the case when the
distance between two floating particles is of the order of the particle radius because the interface shape
changes continuously and the particles accelerate as they move toward each other.
The buoyant weight of particles is an important quantity in the description of the dynamics of
capillary attraction. To see how it enters, we first express gravity as a potential
g = −kg = −g∇z (IV.14)
and write
ρg − ∇p = −∇p̂ (IV.15)
where
p̂ = p + ρgz (IV.16)
In (IV.1), the interface is given by
z=h(x,y,t) (IV.17)
The contact line can be specified by zc, where h intersects the particle surface. Using (IV.16), we find the
pressure force acting on the particle
∫ pndA = ∫ ( p̂ − ρgz )ndA
= ∫ p̂ndA - ρ U VU g − ρ L VL g (IV.17)

where VU is the volume of the particle above and VL is the volume below the contact line and
M= ρ P g (VU + VL ) . We may now write (IV.4) as
dU
M = [(ρ P − ρ U ) VU + (ρ P − ρ L )VL ]g k + ∫ (−p̂I + σ).n dA + γ ∫ n ds . (IV.18)
dt C. L .
c
The first term on the right hand side of (IV.18) is only a portion the buoyant weight (see equation (II.3)).
For isolated spheres, with a prescribed contact angle, the contact line will be a circle on the sphere, so that
the unknowns are VU, VL and zc=h(x,y,t). For the disk hanging on the sharp rim, VU=0, VL=V and the
contact angle are ψ unknown. Equation (IV.5) can be written as
d (I P ω )
dt
= ∫ (x − X ) × [(( −p̂ + ρgz)I + σ).n] dA + γ ∫ ( x − X ) × n c ds (IV.19)
C.L.
The scaling parameters for equations (IV.1)-( IV.1), (IV.18) and (IV.19) are
[D, U, D/U, ηLU/D, U/D, ρL] = [diameter, velocity, time, stress, angular velocity, density]. (IV.20)
Here ηL and ρL are the viscosity and density of the lower liquid and D=2R is the diameter of the sphere or
disk. The dimensionless equations are then in the form
⎡ ∂u ⎤ 1 1
l ⎢ + u. ∇ u ⎥ = - ∇ p̂ + ∇.σ + κ δ(φ) n; ∇. u = 0 in Ω\ P(t) (IV.21)
⎣ ∂t ⎦ Re We
Re l p dU 1
= −G Re((l P − l U )VU + (l P − 1)VL ) k + ∫ (−p̂I + σ ).n dS + ∫ n c ds (IV.22)
β dt Ca C.L.
Re l p dI ' P ω ⎡ ρgzD ⎤ 1
= ∫ ( x − X) × ⎢((−p̂ + )I + σ ).n ⎥ ds + ∫ ( x − X) × n c ds (IV.23)
β dt ⎣ ηL U ⎦ Ca C.L.
ρP D3 6
where k is the unit vector along the z-direction. The particle mass M= , where β = for a sphere
β π
4D
and β = for a disk with h being the disk height. The particle moment of inertia IP=M I 'P D 2 , where I 'P
πh
is the dimensionless moment of inertia. It can be shown that the term proportional to ρgz in (IV.23)
vanishes when the particle is a sphere, but does not vanish when the particle is a disk.
The dimensionless parameters which define the motion of particles are
⎡ ρ UD g D ηL U ⎤
[Re, G, Ca] = ⎢ L , 2 ,
⎣ ηL U γ ⎥⎦
= [Reynolds, gravity, capillary] numbers (IV.24)

and the property ratios

ρ ρP ρU ηU
[l, lp, lU, m]=[ , , , ] (IV.25)
ρ L ρ L ρ L ηL
where the subscript ‘L’ refers to the lower liquid and ‘U’ to the upper liquid. The density parameter l is
ρU
equal to one in the lower liquid and in the upper fluid it is and the Weber number We =Re Ca. In our
ρL
numerical we use the dimensional equations (IV.1)-(IV.4) where the hydrostatic pressure variation is not
removed from the pressure.
The selection of a characteristic velocity U for the definition of the dimensionless parameters in
(IV.20) is ambiguous since a characteristic velocity is not prescribed in data. A natural choice for the
γ
velocity is the capillary velocity U= , which is suggested by other problems of motion driven by
ηL
surface tension. With this choice we may compute

⎡ ρ γD gη2 D ⎤
[Re, G, Ca] = ⎢ L 2 , L2 , 1⎥ (IV.26)
⎣⎢ ηL γ ⎦⎥
from the prescribed data.
V. Numerical Method
In this section we will briefly describe the key features of our numerical scheme. A detailed
description of the numerical algorithm is included as an appendix.
To perform direct numerical simulation of the motion of rigid particles trapped in a two-fluid
interface, we must solve the governing mass and momentum conservation equations for the two fluids,
compute the forces acting on the particles and then move them using Newton’s equations (IV.4). This is a
difficult task because the interface shape changes as the particles move and the capillary force between
the two fluids must be computed subject to the constraint that the contact angle is prescribed on a smooth
surface and the contact line is prescribed on edges.
In this study we will assume that the dynamic contact angle is the same as the static contact angle.
This enforcement of the contact angle on the particle surface causes the contact line to move which may
be described as a capillary induced motion of the contact line due to a prescribed contact angle (see
Friedrichs and Guceri 1993 and Sussman 2001 and references therein). At sharp edges, the motion of the
particles is computed under the constraint that the interface remains pinned to the sharp edges of particles
so that the contact angle changes as the motion proceeds. The contact angle can vary within the limits
specified by the Gibbs extension of the Young-Dupré law.

In this work the level set method is used to track the interface (see Osher and Sethian 1988, Sussman,
Smereka and Osher 1994, Pillapakkam and Singh 2001, Sussman 2001). The level set method works
efficiently on a regular fixed grid and is compatible with the distributed Lagrange multiplier method
(DLM) which will be used to track the motion of rigid particles (see Glowinski, Pan, Hesla and Joseph
1999 and Singh, Joseph, Hesla, Glowinski and Pan 2000). The DLM method also works efficiently on
regular fixed grids. There are several other numerical approaches available for tracking the interface
between two immiscible liquids, e.g., the surface tracking method (Unverdi and Tryggvason 1992), the
volume of fluid method (Hirt and Nichlos 1981), the moving grid methods (Glowinski, Tallec, Ravachol
and Tsikkinis 1992) and the mapping method (Ryskin and Leal 1984) that can be used with the DLM
method to study dynamics of floating particles.
In the level set method, the interface position is not explicitly tracked, but is defined to be the zero
level set of a smooth function φ, which is assumed to be the signed distance from the interface. In order
to track the interface, the level set function is advected according to the velocity field. One of the
attractive features of this approach is that the method does not require any special treatment when a front
splits into two or when two fronts merge.
The key idea in the level-set method is to advect φ with the local velocity, i.e.,
∂φ
+ u.∇φ = 0 . (V.1)
∂t
As φ is a smooth function, it is relatively easy to numerically solve the above equation to update the
interface position. In our implementation, it is assumed to be negative for the upper fluid, positive for the
lower fluid and zero along the interface. The method also allows us to enforce the contact angle on the
rigid particle surfaces and it is relatively easy to implement it in both two and three dimensions.
The motion of particles is tracked using a distributed Lagrange multiplier method (DLM). One of
the key features of the DLM method is that the fluid-particle system is treated implicitly by using a
combined weak formulation where the forces and moments between the particles and fluid cancel, as they
are internal to the combined system. The flow inside the particles is forced to be a rigid body motion
using the distributed Lagrange multiplier method. This multiplier represents the additional body force per
unit volume needed to maintain rigid-body motion inside the particle boundary, and is analogous to the
pressure in incompressible fluid flow, whose gradient is the force needed to maintain the constraint of
incompressibility.
In our numerical scheme the Marchuk-Yanenko operator splitting technique is used to decouple the
difficulties associated with the incompressibility constraint, the nonlinear convection term, the rigid body
motion constraint and the interface motion. The operator-splitting gives rise to the following four sub-

problems: a L2 projection problem for the velocity and the pressure; a nonlinear advection-diffusion
problem for the velocity; a distributed Lagrange multiplier problem that forces rigid body motion within
the particles; and an advection problem for the interface. Details of this method are set down in the
appendix.
V.1 Reinitialization of φ
The level set function φ is reinitialized to be a distance function after each time step by solving
the following equation obtained in Sussman, et al. 1994 to the steady state
∂φ
+ w.∇φ = S(φ 0 ) (V.2)
∂t
where φ0 is the distribution to be reinitialized and
∇φ
w = S(φ 0 ) .
∇φ
Here S(φ0) is the sign function, i.e., S(φ0) = 1 if φ0 > 0 and S(φ0) = -1 if φ0 < 0. In order to avoid
discontinuities, in our code we use the following smoothed sign function
φ0
S(φ 0 ) = ,
φ 02 + h e2
where he is equal to one and half times the element size. Equation (V.2) is a first order hyperbolic partial
differential equation which is solved using a positive only upwinding scheme described in Singh and Leal
1993. Clearly, the characteristics of (V.2) point in the direction of w. Therefore, for the points inside the
upper fluid w points upwards away from the interface and for the points inside the lower fluid it points
downwards. Thus, (V.2) can be solved by specifying the boundary condition φ = φ0 at the two-fluid
interface φ = 0.
V.2 Variation of fluid properties across the interface

In our finite element scheme the fluid viscosity is assumed to jump across the interface, i.e.,
ηL if φ > 0
η= 0.5(ηL + ηU) if φ = 0 (V.3)
ηU if φ < 0.
Here ηL and ηU are the viscosities of the lower and upper fluids, respectively. The fluid density, on
the other hand, is assumed to vary smoothly across the interface

ρL if φ > he
ρ= ρU if φ < -he (V.4)
⎛ πφ ⎞
0.5(ρL + ρU )+ 0.5(ρU - ρL) sin ⎜⎜ ⎟⎟ otherwise
⎝ 2h e ⎠
where he is equal to one and half times the element size, and ρL and ρU are the densities of the two
fluids, respectively. This smoothing of the density is similar to that used by Sussman, et al. 1994, and is
needed for avoiding numerical instabilities when the density ratio ρL/ρU is large.
The surface tension force is smoothed and acts only on the elements for which φ is smaller than
he. This is done by approximating δ(φ) in (IV.1) by a mollified delta function δ h e (φ) using the approach
described in Sussman, et al. 1994:
1 + cos( πφ / h e )
δ h e (φ) = for |φ| < he (V.5)
2h e
0 otherwise
The error introduced by smoothing of the surface tension force is O(he). Equations (V.4) and
(V.5) require that φ be maintained as a distance function which we do in our implementation by
reinitializing φ after each time step.
V.3 Contact angle and contact line conditions
The contact angle boundary condition on the particle surface, n.nφ =cos α, where n is the unit
∇φ
outer normal on the particle surface and n φ = is normal to the interface, is enforced using the
∇φ
approach described in Sussman 2001. Sussman used this approach to prescribe the contact angle on a
stationary flat wall by extending φ to the "outside" of the fluid domain. In this article the same approach is
used to prescribe the contact angle of the two-fluid interface on the surface of a moving sphere. Let us
define t and n2 as
nφ × n t×n
t= , n2 = .
nφ × n t×n
Notice that t is tangent to the contact line, and thus n2 is orthogonal to the contact line and lies in the
tangent plane of the particle surface (see figure V.1a). The next step is to construct a unit vector uex which
is tangent to the interface with contact angle α, points inwards, and lies in the plane formed by n and n2;

nc = -uex is the unit vector which gives the direction of the action of the capillary force. It is easy to verify
that uex depends on c = nφ.n2 and is given by
⎧ n − cot( π − α) n 2
⎪ n − cot( π − α) n if c < 0
⎪ 2
⎪ n + cot( π − α) n 2
u ex =⎨ if c > 0 (V.6)
⎪ n + cot( π − α) n 2
⎪n if c = 0
⎪
⎩
n̂ φ Interface
t̂
1350 n̂
Contact line n2
Tangent plane at the point of

contact
V.1(a)

nc=- uex
V.1(b)
Figure V.1. (a) The unit normal to the particle surface n, the tangent to the contact line t and the normal to the
interface nφ are shown. (b) A schematic of the interface shape and the contact line are shown for the initial and
steady states. In simulations the contact angle on the particle surface is prescribed to be 135º which is done by
extending the level set function to the inside of the particle. The contact line moves downwards because of the
interface deformation near the particle and this decreases the vertical component of capillary force.
To enforce the prescribed contact angle, φ is extended inside particles and on their surfaces by
solving
∂φ
+ u ex .∇φ = 0 . (V.7)
∂t
In other words, for all nodes inside and on the particle surface (V.7) is used to modify φ. The resulting
extended level set function satisfies the contact angle on the particle surface.
The contact line on the particle surface moves when the contact angle is enforced using (V.7).
This could be called the capillary induced motion of the contact line due to a prescribed contact angle.
This approach has been used in many past numerical studies of problems involving moving contact lines
(see Friedrichs and Guceri 1993 and Sussman 2001 and references therein). For example, in the injection
molding problems this approach has been used to track the motion of liquid front advancing into empty
molds.
Clearly, this motion of the contact line on the particle surface is in conflict with the no slip
condition for viscous fluids (see Dussan and Davis 1974, Dussan 1976 and Kistler and Scriven 1993 and
references therein). However, if the contact line position on the particle surface is not updated, the contact

line cannot move. In the capillary induced motion approach the no slip condition is satisfied before and
after the contact line moves; this pragmatic procedure could be called an effective numerical slip. This
method of moving the contact line when the contact angle is prescribed has been used by Friedrichs and
Guceri 1993 and Sussman 2001. An alternative approach used in some studies is to use a slip condition in
a small neighborhood of the contact line to ensure that it moves (see Kamal, Goyal and Chu 1988 and
references therein). The slip velocity of the contact line is assumed to be proportional to the shear stress
on the wall. This approach however does not ensure that the contact angle remains constant. Another
aspect of the floating particle problem not treated here is that the contact angle for advancing and
receding contact lines is different which can change the dynamical behavior of floating particles.
VI. Initial value problems for the evolution of floating particles to equilibrium
Here we report results of simulations of initial value problems for sphere and disks which are
initially motionless, but not in equilibrium, to an equilibrium in which they are again motionless. The
particles are heavier than the heavy liquid below and they float. Initially, the particles are motionless and
imbedded in a flat interface; the spheres are centered with their midplane in the interface and the contact
angle fixed and held at 135º throughout the simulation. The assumed value of the contact angle is likely to
be insensitive to the contact line speed in real experiments, as it is relatively large. Disks are pinned at the
sharp edge of the upper rim throughout the simulation.
We do simulations for one sphere, one disk, two and four spheres and two disks. Initially,
particles are not in equilibrium because they are heavy and must sink to equilibrium. For all cases, the
particles reach an equilibrium in which they are motionless and in a balance between capillary forces and
the buoyant weight; for single particles, spheres and disks, the computed values at equilibrium can be
computed with the analytical expressions (II.5) and (II.8) and the agreement is satisfactory. The evolution
to equilibrium for more than one particle takes place by sinking and capillary attraction; at the end the
particles have self assembled.
The conditions under which spheres and disks evolve to equilibrium are different. The interface
near the spheres adjusts to meet the contact angle requirement and they sink until the buoyant weight
becomes equal to the vertical component of the capillary force. The disks, on the other hand, sink causing
the interface to deform and increasing the contact angle and the vertical component of the capillary force.
The disks stop sinking when the vertical component of the capillary force becomes equal to the buoyant
weight.
An attractive force between floating particles arises because the meniscus between drops in much
the same way as a water meniscus will sink in a hydrophobic capillary tube. This dropping of the
meniscus inside relative to the outside produces an asymmetry which generates attractive capillary forces.

For spheres, since the contact angle is fixed, the contact line between the spheres drops. For disks, since
the contact lines are fixed at the rim, the contact angles between the two disks decrease. In both cases, the
asymmetry results in an attractive lateral capillary force to act on the particles.
The domains used in our simulations are box shaped with rectangular cross sections. The
coordinate system used throughout this paper is shown in figure VI.1. The x-, y- and z-components of
particle velocity will be denoted by u, v and w, respectively.
We will also assume that all dimensional quantities, unless otherwise noted, are in the CGS units.
The lower fluid density ρL=1.0 gm/cm3. The viscosity of the lower fluid, and the density and viscosity of
the upper fluid are varied. The particle density is assumed to be greater than one. The values of the
interfacial tension are selected to ensure that the particle remains suspended in the interface. The
acceleration due to gravity g=981.0 cm/s2 and acts along the negative z-direction. The initial velocities are
assumed to be zero everywhere.
The no slip boundary condition is applied on the surface of the box shaped computational
domains. The contact angle between the interface and the box boundaries is assumed to be 90º, the
interface near the walls is flat.
We next present the results for floating spheres and disks, to demonstrate that the scheme works
correctly, and that it reproduces the expected dynamical behavior and the equilibrium state.
VI.1 Initial value problems for floating spheres

In this subsection, we compute the motion of spheres released in the interface; the contact line
intersects the sphere1 at a place different than one required for equilibrium. The sphere diameter is
assumed to be 0.2 cm. The initial interface shape is flat, except near the sphere surface where a contact
angle of 135º is prescribed (see figure V.1). The parameters are in the range for which a sphere trapped on
the interface can be in equilibrium. The equilibrium interface shape and the floating height depend on the
problem parameters.
VI.1.1 Motion of a single sphere

When a sphere is suddenly released in the interface, the meniscus shape evolves to equilibrium.
During this time, the velocity field in the two fluids is non-zero and the capillary force acting on the
particle varies; the sphere velocity and its position in the interface change with time. The final equilibrium
1
If the initial particle position was such that the interface did not touch the particle surface, then we
would also need to address the problem of an interface coming in contact with a solid surface. This would
require us to include additional physics to specify the conditions under which an interface can touch a
solid surface. This physics is not included in the current version of our code.

position described by analytical expression (II.5), however, is independent of these transients and can be
used to verify the accuracy of numerical results.
We first present results that show that the trajectory of a sphere released in the two-fluid interface
is independent of the mesh resolution and the time step. We have used two regular tetrahedral meshes to
show that the results converge with mesh refinement. In a tetrahedral element there are seven velocity and
four pressure nodes. The rigid body constraint inside particles is enforced using uniformly distributed
collocation points. The number of velocity nodes and elements in the first mesh are 117,649 and 13,824,
respectively. In the second mesh, referred to as mesh B, there are 274,625 velocity nodes and 32,768
elements. The time step for these simulations is 0.0001, 0.00005 or 0.000025.
Oblique view Frontal view Velocity field

0.44
0.33
0.22
0.11
0
0 0.1 0.2 0.3 0.4
VI.1a
Oblique view Frontal view Velocity field

0.44
0.33
0.22
0.11
0
0 0.1 0.2 0.3 0.4
VI.1b
Figure VI.1. The particle position and the interface shape and the velocity field in the domain midsection are shown.
The length of velocity vectors is magnified 30 times. The length of the velocity vectors in (b) is smaller indicating
that the fluid velocity decreases with time. The oblique and side views are shown. (a) t=0.003. The fluid velocity is

largest near the contact line where the interface curvature is large, (b) t=0.08. The dimensionless parameters based
on the maximum particle velocity are (Re=0.064, G=1916.0, Ca=0.02) and based on the capillary velocity are
(Re=3.2, G=0.766).
0.4 dt=0.0001
dt=0.00005
w 0.3
dt=0.000025
0.2 dt=0.000025, mesh B
0.1
0
-0.1 0 0.05 0.1
t
-0.2
Figure VI.2. The vertical component of sphere velocity w released from rest on the interface is shown as a function
of time for three different values of the time step. The curve marked mesh B is for a more refined mesh. The density
and viscosity of the lower fluid are 1.0 gm/cm3 and 1.0 Poise, and those of the upper fluid are 0.1 gm/cm3 and 0.1
Poise. The interfacial tension is 16.0 dynes/cm and the particle density is 1.05 gm/cm3.
The sphere density is 1.05 gm/cm3 and the interfacial tension is 16.0 dynes/cm. For the upper
fluid density is 0.1 gm/cm3 and the viscosity is 0.1 Poise. The initial velocity distribution in the fluid, and
the sphere’s linear and angular velocities are assumed to be zero. The domain is assumed to cubical with
sides 0.4 cm. The sphere center is at a distance of 0.02 cm above the undeformed interface which passes
though the domain center.
In figure VI.2, w is plotted as a function of time for three time steps and two mesh resolutions.
When the time step is reduced or when the mesh is refined the variation of w with time remains
approximately the same. This allows us to conclude that the numerical results converge with both mesh
and time step refinements.
Figure VI.2 shows that the vertical component of the sphere velocity w increases for t<~0.005 s
and then it starts to decrease; it becomes negative for t=~0.019 s and then increases again and becomes
very small and fluctuates around zero for t>~0.06 s. The other components of velocity u and v remain
small for all times. We will assume that for t=0.06 s the sphere has reached a state of equilibrium with
h2=0.156R and θ c =65.26°. The computed values given in table VI.1 are in good agreement with the
equilibrium formula (II.5). We may therefore conclude that the state of equilibrium is captured correctly

by our code. The dimensionless parameters based on the maximum vertical velocity are: Re=0.064,
Ca=0.02, G=1916.0 and We=1.28x10-3.
Table VI.1. The interfacial deformation h2/R, the floating height R cos θ c from numerical computation are used to
compute the sum of the pressure and vertical component of capillary forces from (II.1) and (II.3) for 5 values of the
interfacial tension. The sphere density is 1.05 gm/cm3 and its weight is 4.315 g cm/s2. The density of lower fluid is
1.0 gm/cm3 and that of the upper fluid is 0.1 gm/cm3. For all five cases, Fp + Fc is approximately equal to the
particle weight; we get the correct value of the sphere weight from simulations. As expected, the sphere's floating
height increases and the interface deformation decreases with increasing surface tension. The interfacial
deformation for these calculations is restricted because the domain size is relatively small. But, we can still compare
these values as the same domain is used for all interfacial tension values.
γ (dynes/cm) h2/R R cos θ c (cm) θ c (degrees) Fp + Fc (g cm/s2)

10 0.237 0.257 75.00 4.33
14 0.173 0.376 67.95 4.35
16 0.156 0.419 65.26 4.35
20 0.130 0.466 62.28 4.36
25 0.114 0.514 59.07 4.33
To understand the initial increase in w, we notice that the angle θc giving the position of the
contact line in figure VI.1a is larger than that for the equilibrium state shown in figure VI.1b (also see
figure V.1b). Thus, the vertical component of capillary force is initially larger than the final value and as a
result the particle moves upwards. This is a consequence of the fact that initially the interface is
approximately flat everywhere except near the sphere (see figure VI.1a). The large curvature of the
interface near the sphere at early times is reduced by interfacial tension and the interface assumes its
equilibrium shape. The contact line moves downwards, reducing the vertical component of the capillary
force. The vertical component of the pressure force in figures VI.1a and b are different, but since in the
case shown in figure VI.1a the fluid velocity is not small, the pressure force cannot be determined using
hydrostatics.

Figure VI.3. The sphere position and the corresponding interface shape are shown on the domain midsection. For
the bottom sphere (shown as a circle) and the interface shape the surface tension is 10 dynes/cm and for the top
sphere and the interface shape it is 25 dynes/cm. The depth to which a sphere sinks into the lower fluid decreases
with increasing value of the surface tension.
To validate our code further, we performed calculations for five different values of interfacial
tension γ while keeping the other parameters fixed. Figure VI.3 shows that, as expected, when γ is smaller
the particle sinks to a greater depth. In table VI.1 we have listed the floating heights, defined to be the
vertical distance of the particle center from the contact line, and the sum of the pressure and vertical
component of capillary forces acting on the particle for these five values of the interfacial tension. For all
cases in equilibrium, as required, Fp + Fc is approximately equal to the particle’s weight. There are small
differences due to numerical errors.
Table VI.2. The interfacial deformation h2/R, the floating height R cos θ c and the point of contact θ c are listed as a
function of the upper fluid density. The interfacial tension is 16 dynes/cm. The sphere density is 1.05 g/cm3 and its
weight is 4.315 g cm/s2. The density of lower fluid is 1.0 g/cm3 and that of the upper fluid is varied. The floating
height increases and the interface deformation decreases with decreasing density of the upper fluid.
ρU (g/cm3) h2/R R cos θ c θ c (degrees)

0.1 0.156 0.419 65.26
0.01 0.159 0.417 65.34
0.0016 0.161 0.417 65.37
In table VI.2 we have listed the floating heights for two additional cases where the density of the
upper fluids are 0.01 gm/cm3 and 0.0016 gm/cm3, and the corresponding viscosities are 0.033 Poise and

0.0166 Poise. The interfacial tension is 16.0 dynes/cm. The time step used for these calculations was
2x10-5s. It is necessary to use a smaller value of the time step for these simulations because the ratio of
lower and upper fluids densities is larger. The time step used is smaller also when the ratio of the lower
and upper fluids viscosities is larger. The domain was discretized using mesh B described above. Table
VI.2 shows that the floating height slightly decreases when the density of the upper fluid is reduced.
The equilibrium analysis, presented in section II, assumes that the fluid extends to infinity in the
x-, y- and z-directions which is not the case for our simulations. This may explain some differences
between our simulations and the analytical results. These differences are expected to decrease with
increasing box size. We also wish to note that for our simulations the magnitude of fluid velocity
decreases as the state of equilibrium is approached, but it does not decrease beyond a certain value which
depends on the fluid viscosity, surface tension and the interface curvature. The flow develops steady
spurious circulation cells around the interface that are similar to those seen in simulations of drops
(Scardovelli and Zaleski (1999)). It has been noted by D.D. Joseph that these circulation cells arise in
simulations because the discretized equation for the vorticity, which can be obtained by taking the curl of
the momentum equation, contains a non zero contribution from the layer (V.4) representing the delta
function in the level set method. This creates vorticity along the discretized interface which diffuses into
the domain. The presence of these cells, however, does not seem to affect the overall force balance,
discussed in table VI.1, for equilibrium.
VI.1.2. Motion of two spheres

We next present results for the case where two spherical particles are released near each other on
the interface at the same vertical height. The initial interface position is assumed to be flat, except near the
particle surfaces where a contact angle of 135º is prescribed. The initial vertical height of the spheres is
higher than for a single sphere in equilibrium for the same parameter values. The parameters are assumed
to be in the range for which a single sphere can be in equilibrium.
For these calculations, the particle density is 1.05 gm/cm3. The interfacial tension is 16.0
dynes/cm. The upper fluid density is 0.01 gm/cm3 and viscosity is 0.033 Poise. The initial velocities are
assumed to be zero. The domain height is 0.4 cm. The domain width in the x-direction is 0.4 cm and in
the y-direction is 0.8 cm. The undeformed interface passes though the domain center and the particle
centers are initially at a height of 0.02 cm above the interface. The initial distance between the spheres in
the y-direction is 2.6 R or 3.2 R. The mesh resolution is comparable to that for the coarse mesh in section
VI.1.1.
We have already noted that when two or more spheres released in the interface are close they
move towards each other due to the action of the lateral component of the capillary force associated with

the asymmetric deformation of the interface around the particles. Figure VI.4a shows that for t=0.0042
the interface shape is deformed in a small region around the spheres and farther away it is relatively flat;
hence there is no lowering of interface. Consequently, at this time, the spheres do not experience any
lateral attractive force. But, as for a single sphere, the interface height around the spheres decreases as
time increases and, as a result, the contact lines on both spheres move downward. The vertical component
of capillary force, which is initially larger than the final value, causes the spheres to move upward, but as
the contact lines move downward the vertical capillary force decreases and the vertical velocities become
negative. After this initial motion for t<~0.05 s, the vertical velocities become relatively small, but remain
negative as the spheres approach each other. These initial transients in the velocities of the spheres could
be diminished by setting their initial positions and interface shape closer to those for the equilibrium of an
isolated sphere. This would however complicate the problem of prescribing initial conditions and make
the problem less realistic.

0.44
0.33
0.22
0.11
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
(a)
0.44
0.33
0.22
0.11
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
(b)
0.4
4
0.3
3
0.2
2
0.1
1
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
(c)
.
0.4
4
0.3
3
0.2
2
0.1
1
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
(d)
Figure VI.4. The positions of two spheres suspended in the two-fluid interface and the velocity distribution at the
domain midsection are shown at t=0.0042, 0.175, 0.225 and 0.339. The length of velocity vectors is magnified 50
times. The length of the velocity vectors is the largest in (a) and the smallest in (d) which indicates that the
maximum velocity in the domain is decreasing with time. The particles are moving toward each other in the

interface. The particles are "supported" by the capillary force associated with the deformation of the interface. The
surface tension is 16.0 dynes/cm, the particle density is 1.05 g/cm3 and the density of the top fluid is 0.01 g/cm3 and
that on the bottom is 1.0 g/cm3. The initial distance between the spheres is 3.2 R. (a) Oblique views, (b) Side view of
the velocity field on the domain midsection. The dimensionless parameters based on the maximum particle velocity
are (Re=0.028, G=1.0x104, Ca=0.00875) and based on the capillary velocity are (Re=3.2, G=0.766).
Snapshots at t=0.0042, 0.175, 0.225 and 0.399 of the evolution to equilibrium are shown in figure
VI.4. At t=0.399 the spheres are close to equilibrium.
The magnitude of the lateral capillary force F ( γD) increases as the distance between the spheres
decreases, where F is the y-component of capillary force acting on a sphere (see figure VI.5). This is also
seen as an increase in the approach velocity v2-v1 plotted as a function of time in figure VI.5. However,
the approach velocity begins to decrease due to the lubrication forces when the gap between the spheres is
small enough. Simulations also show that when the distance between the spheres is larger, the initial
approach velocity is smaller, because the lateral component of capillary force is smaller, and the time
needed for the spheres to come together is larger. This is in agreement with (II.17) which implies that the
lateral force increases with decreasing distance between the spheres. The dimensionless parameters based
on the maximum lateral velocity are (Re=0.028, G=1.0x104, Ca=0.00875).
The contact line and the interface shape evolve as the spheres move toward each other. Therefore,
the lateral component of the capillary force and the contribution of the pressure to force change as the
distance between the spheres decreases. Petkov, et al. 1995 measured the approach velocity of particles
and found that when the gap between the particles is smaller than O(30R), which is the case for our
simulations, the non linearity of the interface curvature also influences the approach velocity. For the case
described in figure VI.4 the simulations were stopped at t=0.339 when the distance between the spheres’
surfaces reduced to 0.047 R. The interface shape at this time is shown in figure VI.4d. Notice that the
interface height between the spheres in this figure is even lower and the interface shape is flatter. For the
case where the initial distance between the spheres’ surfaces is 0.6 R the time taken to reach the same
separation is ~0.181. The approach velocity remains relatively small for t<~0.04 during which the initial
interfacial deformation takes place, which is approximately the time interval in which a single sphere
reached equilibrium in section VI.1.1.

0.7
0.4 dt=0.0001
0.6
dt=0.00005
0.5 0.3
dt=0.000025
(y2-y1)/R
0.4
0.2
0.3
0.2 0.1
0.1
0
0 0 0.05 0.1 0.15 0.2
0 0.05 0.1 0.15 0.2
-0.1 t
t
0.2
F/(γ D)
0.1
0
0 0.05 0.1 0.15 0.2
t
-0.1
Figure VI.5. The distance between the surfaces of spheres, the approach velocity v2-v1 and the lateral capillary force
are plotted as functions of time. The approach velocity v2-v1 is shown for three different values of the time step.
Notice that the approach velocity initially increases as the interface height between the spheres decreases and then
decreases as the gap between the spheres becomes small. The initial distance between the spheres is 2.6R.
In figure VI.6 the approach velocity and the dimensionless distance between the spheres are
shown for five additional parameter values. For these cases, the spheres centers are initially at a height of
0.04 cm above the interface and the initial distance between the spheres in the y-direction is 3.2 R. These
figures show that the approach velocity becomes positive after a short time interval and increases as the
spheres move toward each other. The approach velocity decreases when the gap between the spheres
becomes small due to the lubrication forces. The lateral velocity for small times is negative because when
the spheres sink into the lower liquid the hydrodynamic force is repulsive. From these figures we also
conclude that the approach velocity increases when the viscosity of the upper or lower liquid is decreased.
It also increases when the particle density or the surface tension coefficient is increased. The lateral
velocity contains oscillations when the Reynolds number is of order one or larger.

0.25 3.6
v2-v1 3.2
0.15 d/R
2.8
0.05
2.4
-0.05 0 0.5 1 2
t t
0 0.5 1
(a)
0.6 3.5
v2-v1
0.3 3
d/R
0 2.5
-0.1 0.1 0.3 0.5
t 2
-0.3
0 0.2 0.4
t
(b)
4
3.5
v2-v1
2 3
d/R
0 2.5
0 0.05 0.1
-2 t 2
0 0.05 t 0.1
(c)
0.8
3.5
v2-v1
0.4
3
0 d/R
0 0.1 0.2 0.3 2.5
-0.4
t
2
-0.8
0 0.1 0.2 0.3
t
(d)

0.8 3.5
v2-v1
0.4 3
d/R
0 2.5
0 0.1 0.2 0.3 0.4
-0.4
2
t
0 0.1 t 0.2 0.3
(e)
Figure VI.6. The distance between the spheres and the approach velocity v2-v1 are plotted as functions of time.
Notice that the approach velocity initially increases as the interface height between the spheres decreases and then
decreases as the gap between the spheres becomes small. The initial distance between the spheres is 3.2R. (a) The
parameters are: (lp=1.1, lU=1.0, m=1.0), η L =0.1 Poise and γ=2.0 dynes/cm. The dimensionless parameters based
on the maximum particle velocity are (Re=0.38, G=5.4x103, Ca=0.0095) and based on the capillary velocity are
(Re=40.0, G=0.491). (b) The parameters are: (lp=1.1, lU=1.0, m=1.0), η L =0.01 Poise and γ=2.0 dynes/cm. The
dimensionless parameters based on the maximum particle velocity are (Re=10.2, G=784.8, Ca=0.0025) and based
on the capillary velocity are (Re=4000, G=4.9x10-3). (c) The parameters are: (lp=1.01, lU=10.0, m=10.0), η L =0.1
Poise and γ=16.0 dynes/cm. The dimensionless parameters based on the maximum particle velocity are (Re=5.12,
G=29.9, Ca=0.016) and based on the capillary velocity are (Re=320, G=7.66x10-3). (d) The parameters are:
(lp=1.005, lU=2.0, m=2.0), ηL =0.1 Poise and γ=4.0 dynes/cm. The dimensionless parameters based on the
maximum particle velocity are (Re=1.38, G=412.1, Ca=0.017) and based on the capillary velocity are (Re=80,
G=0.12). (e) The parameters are: (lp=1.005, lU=2.0, m=1.0), ηL =0.1 Poise and γ=4.0 dynes/cm. The
on the capillary velocity are (Re=80, G=0.12).
Another interesting consequence of the lowering of interface height and contact lines between the
spheres is a decrease in the vertical component of capillary force. This component of the capillary force
acts against gravity and keeps them floating even though they are heavier than both liquids. This decrease
causes the floating height of the spheres to decrease slightly (see table VI.3). For example, the height for a
single sphere in section VI.1.1 for the same parameters was 0.419, but for the two spheres in figure VI.4d
it is 0.375, where they are almost touching. This decrease in the floating height raises the contact line
position everywhere except between the spheres so that the vertical component of the capillary force
returns to the value required for balancing their buoyant weights (see figure VI.7). When the parameters
are such that the sum of the vertical component of the capillary and pressure forces is near its maximum
value for an isolated sphere, a decrease in the floating height would actually decrease the vertical
component of total force. Therefore, if the spheres are barely floating, a decrease in the floating height
will cause them to sink. This suggests that the clusters of spheres are more likely to sink than an isolated
sphere.

R h2
θc1
θc2
Figure VI.7. The figure shows the contact line and the interface shape near a floating sphere. The contact line
position on the left is lower, i.e., θc2<θc1, as there is another sphere floating to the left (not shown). Figure II.8
shows that the vertical capillary force increases with increasing θ c only when θ c is less than 109º (which depends
on the contact angle). Thus, the vertical capillary force acting on the right side of the sphere is larger than on the
left side. But, if θ c1 is greater than 109º, this may not be the case.
Table VI.3. The floating height R cos θ c , the point of contact θ c and the interfacial deformation h2/R are shown as
a function of the distance d between the spheres. The interfacial tension is 16.0 dynes/cm. The sphere density is 1.05
gm/cm3. The density of the lower fluid is 1.0 gm/cm3 and that of the upper fluid is 0.1 gm/cm3. The floating height is
computed based on the contact line position on the sphere surface away from the gap. Notice that the floating height
of the spheres decreases and the interfacial deformation increases as they come closer. These values should
approach the values for an isolated particle when the distance d is large.
d (cm) R cos θ c (cm) θ c (degrees) h2/R

0.279 0.397 66.60 0.234
0.2493 0.388 67.13 0.289
0.2118 0.375 67.99 0.302
VI.1.3. Motion of four spheres

For these calculations the sphere density is 1.05 gm/cm3 and the interfacial tension is 16.0
dynes/cm. The upper fluid density is 0.1 gm/cm3 and the viscosity is 0.1 Poise. The motion starts from
rest. The domain height is 0.4 cm. The width along x-direction is 0.8 cm and along y-direction is 0.8 cm.
The undeformed interface passes though the domain center. The initial positions of the four spheres are
(0.4,0.22,0.22), (0.4,0.58,0.22), (0.23,0.4,0.22) and (0.57,0.4,0.22). The mesh resolution is the same as for
the coarse mesh in section VI.1.1.

Figure VI.8. Oblique views showing the positions of four rigid spheres suspended in the two-fluid interface at
t=0.002 and 0.096. After they are released, they move toward each other. The surface tension is 16.0 dynes/cm, the
particle density is 1.05 g/cm3 and the density of the top fluid is 0.1 g/cm3 and that on the bottom is 1.0 g/cm3. The
on the capillary velocity are (Re=3.2, G=0.766).
After initial transients have died out, the spheres move toward each other. Their vertical velocity
becomes small at t=~0.05 and around this time the interface height between the spheres begins to
decrease.
In figures VI.8a and b the spheres and interface shape are shown at t=0.002 and 0.096,
respectively. In the first figure, the interface shape between spheres is not significantly deformed and the
lateral velocities are small and in the second figure the interface between the spheres is lowered. The
interface shape and the contact lines for the spheres are no longer symmetric and thus the lateral
component of capillary force acting on the spheres is not zero which causes them to move laterally
towards each other.

As spheres approach each other, the vertical component of capillary force decreases due to the
lowering of interface height between the spheres and thus the floating heights of the spheres decrease
slightly (see figure VI.8b). For example, the height for a single sphere in section 3.1 for the same
parameters is 0.419, for the two sphere case it is 0.375, and for the four spheres in figure VI.8b, where
they are almost touching, it is 0.355.
VI.2 Initial value problems for floating disks

The disks in our simulations are released with the top plane of the disk in the plane of the
undeformed interface. The disk velocity, its position and the interfacial deformation change with time.
The final state, described by expression (II.8), is independent of these transients and can be used to verify
the accuracy of numerical results. In equilibrium, the interface shape is such that the capillary force is
exactly balanced by a jump in the pressure across the interface and therefore no fluid flow is induced. The
fluid velocity in simulations is small but nonzero.
VI.2.1. Motion of a single disk

We first discuss results that show that the trajectory of a disk released in the two-fluid interface is
independent of the mesh resolution and the time step. We have used two regular tetrahedral meshes to
show that the results converge with mesh refinement. The rigid body constraint inside particles is
enforced using uniformly distributed collocation points. The number of velocity nodes and elements in
the first mesh are 117,649 and 13,824, respectively. In the second mesh, referred to as mesh B, there are
274,625 velocity nodes and 32,768 elements. The time step for these simulations is 0.0001 s or 0.00005 s.
The disk radius is 0.1 cm and its density is 1.5 gm/cm3. The interfacial tension is 5.0 dynes/cm.
The density of the upper fluid is 0.1 gm/cm3 and its viscosity is 0.1 Poise. Initially, all velocities are zero.
The domain is cubical with sides 0.4 cm. The top surface of the disk is in the plane of the undeformed
interface which is at z=0.28 cm. The parameters are in the range for which a disk trapped in the interface
can be in equilibrium.

t
0 0.05 0.1 0.15
0
dt=0.0001
-0.5
dt=0.00005, Mesh B
-1
Figure VI.9. The vertical component of velocity w for a disk released in the interface is shown as a function of time
for two different values of the time step. The curve marked mesh B is for a more refined mesh. The density and
viscosity of the lower fluid are 1.0 gm/cm3 and 1.0 Poise and those of the upper fluid are 0.1 gm/cm3 and 0.1 Poise.
The interfacial tension is 5.0 dynes/cm and the particle density is 1.5 gm/cm3. The dimensionless parameters based
on the maximum vertical velocity of the disk are (Re=0.35, G=100.2, Ca=0.014) and based on the capillary velocity
are (Re=50, G=0.039).
Figure VI.9 shows that the numerical results for w vs. t converge with both mesh and time step
refinements. Figures VI.9 also shows that the vertical component w of the velocity of the disk decreases
for t<~0.02 s. It begins to increase slowly at t=~0.02 s and fluctuates around zero for t>~0.12 s (see figure
VI.10). The other components of velocity u and v are negligible for all times. We will assume that for
t=0.12 s the particle has reached a state of equilibrium with h2=0.28R and ψ=45.8º. Table VI.4 shows that
these values are in good agreement with the equilibrium requirement (II.8). We may therefore conclude
that the state of equilibrium is captured correctly by our code.
VI.10a

VI.10b
Figure VI.10. The disk position, the interface shape and the velocity field in the domain midsection are shown. The
length of velocity vectors is magnified 30 times. The oblique and side views are shown. (a) t=0.0125 s, (b) t=0.15 s.
The dimensionless parameters based on the maximum vertical velocity of the disk are (Re=0.35, G=100.2,
Ca=0.014) and based on the capillary velocity are (Re=50, G=0.039).
The vertical velocity w decreases initially with time because the disk is denser than the liquid
below and the vertical component of the capillary force is zero, as the interface is not deformed. The
buoyant weight and the contact angle increase simultaneously as the disk sinks.

Figure VI.11. The oblique and front views of the floating disks and the interface shape are shown. The depth to
which a disk sinks into the lower fluid increases with increasing disk density. (a) ρp=1.1, (b) ρp=1, and (c) ρp=1.5.
We performed calculations for two additional values of the disk density ρp while keeping the
other parameters fixed. Figure VI.11 shows that heavier disks sink to a greater depth and the contact angle
is larger. In table VI.4 we give the computed values of h2/R and the forces Fp + Fc for the three values of
ρp. For all cases, the computed values of the force are approximately equal to the weight of the disk.
Some small differences are due to numerical errors. The equilibrium analysis of section II assumes that
the fluid extends to infinity in the x-, y- and z-directions which is not the case for our simulations. This
may explain small differences between our simulations and the analytical results.
Table VI.4. The interfacial deformation h2/R and the sum of the pressure and vertical component of capillary forces
obtained using (II.6) and (II.7) are shown as a function of the disk density. The interfacial tension is 5.0 dynes/cm.
The density of lower fluid is 1.0 gm/cm3 and that of the upper fluid is 0.1 gm/cm3. For all three cases, Fp + Fc is

approximately equal to the disk weight. Also notice that, as expected, the interface deformation increases with
increasing particle density.
ρ ( gm/cm3) weight (g cm/s2) h2/R Fp + Fc (g cm/s2)

1.1 6.78 0.15 6.86
1.2 7.40 0.21 7.38
1.5 9.24 0.28 9.26
VI.2.2 Motion of two disks

We next present results for the case where two disks are released near each other in the interface
at the same vertical height. The domain height is 0.4 cm. The domain width along the x-direction is 0.8
cm and along the y-direction is 0.4 cm. The disk diameter is 0.1 cm and the height is 0.1 cm. The
undeformed interface is at a distance of 0.23 cm from the bottom. The initial velocities are assumed to be
zero. The mesh resolution is comparable to that for the coarse mesh in subsection VI.2.1.
We first describe the case for which the particle density is 1.1 gm/cm3, the interfacial tension is
3.0 dynes/cm, and the upper fluid density and viscosity are 0.1 gm/cm3 and 1.0 Poise, respectively. The
viscosity of the lower fluid is 10.0 Poise. The two disks are placed at (0.32, 0.2, 0.18) and (0.48, 0.2,
0.18). The initial distance between the disks in the x-direction is 3.2R. Figure VI.12a shows that at t=0.04
the interface is deformed only in a small region around the disks and the disks do not experience a lateral
attractive force. When the disks sink the contact angles increase. After this initial motion for t<~0.12 s,
the disks sink slowly, as they move toward each other.

0.4
4
0.3
3
0.2
2
0.1
1
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
0.44
0.33
0.22
0.11
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
0.4
4
0.3
3
0.2
2
0.1
1
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
0.4
4
0.3
3
0.2
2
0.1
1
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8

Figure VI.12. The positions of two disks suspended on the two-fluid interface and the velocity distribution at the
domain midsection are shown at t=0.04, 0.2, 0.5 and 0.8 s. The length of velocity vectors is magnified 50 times. The
disks move toward each other in the interface. The surface tension is 3.0 dynes/cm, the particle density is 1.1 g/cm3
and the density of the top fluid is 0.1 g/cm3 and that of the bottom fluid is 1.0 g/cm3. The initial distance between the
disks is 3.2 R. The dimensionless parameters based on the maximum velocity of the disk are (Re=0.0023, G=242.0,
Ca=1.5) and based on the capillary velocity are (Re=1.5x10-3, G=545.0).
In figure VI.12b the disks and the interface shape are shown at t=0.2 s. The interface height
between the disks in figure VI.12c at t=0.5 s is significantly lower than on the sides. Since the contact
angle between the disks is smaller, the lateral capillary force is larger.
Figure VI.13a shows that the approach velocity u2-u1 first increases and then decreases with time.
The approach velocity increases with time because the lateral component of the capillary force increases
as the distance between the disks decreases. This also causes the disks to slightly tilt in the xz-plane. For
all cases simulated in this paper it is smaller than 0.5º. The time taken by the disks to come together
increases with the initial distance not only because the distance is larger, but also because the approach
velocity decreases with increasing distance. Figure VI.13a also shows that the approach velocity starts to
decrease when the gap between disks becomes small compared to the disk radius due to the activation of
lubrication forces.

3 0.6
d/R
0.4
u2-u1
2.5
0.2
0
0 0.05 0.1 0.15
t
2
0 0.05 0.1 0.15 -0.2
t
VI.13. (a)
0.1
3.5
u2-u1
3
d/R 0.05
2.5
t
2 0
0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8
t
VI.13. (b)
0.4
d/R u2-u1
4
0.2
0
0 0.2 0.4 0.6 0.8
2 t
0 0.2 0.4 0.6 0.8
t
VI.13. (c)
Figure VI.13. The approach velocity u2-u1 and the dimensionless distance d/R between the disks are plotted as
functions of time. The approach velocity initially increases as the interface height between the disks decreases and
then decreases as the gap between the particles becomes small. (a) The initial distance between the disks is 2.8R and
the dimensionless parameters based on the maximum disk velocity are (Re=0.11, G=2224.5, Ca=0.0021), (b) The

initial d/R=3.2 and the dimensionless parameters are (Re=0.0023, G=242.0, Ca=1.5) (c) The initial d/R= 5.2 and
the dimensionless parameters are (Re=0.11, G=2224.5, Ca=0.0021).
In figure VI.13 we have plotted the separation distance d/R ≥ 2 and the approach velocity u2-u1 as
function of time for three initial values of d/R = (2.8, 3.2, 5.2) at t=0. In all cases the approach velocity
has a maximum value at a certain time and then decreases. The slow down is due to activation of
lubrication force at close approach. The approach velocity in the case of the smallest initial distance
d/R=2.8 shown in figure VI.13a is negative for a short period and the particles first disperse and then
attract.
The same type of plots are presented for the case in which the viscosity of the lower fluid is
reduced tenfold, from 0.1 to 0.01 Poise and d/R=6, larger even than d/R=5.2 in figure VI.13c. The
magnitudes of the velocities which develop are larger in the small viscosity fluid even though the initial
separation is larger. The approach velocity is not monotonic due to changes in the interface shape away
from the disk.
The tendency toward initial repulsion followed by attraction seen in figure IV.1a when the initial
separation d/R=6 is small has been observed in experiments. Joseph et al. 2003, say (page 143)
“We create such dispersions by pouring particles on the liquid, nothing complicated, just like a salt shaker.
As soon as the particles hit the liquid surface they disperse radially leading to dispersions like that 3 minutes in
figure 4. The dispersion, followed by attraction, is more or less universal and we have not seen it mentioned in the
literature.”
We expect the limiting value d/R=2 to be achieved asymptotically as t → ∞ ; this asymptotic
result cannot be achieved with the present numerical package which uses a security zone to prevent
collisions of particles. In the future we will implement the new scheme of Singh, Hesla and Joseph 2003
which does allow collision and close packing in equilibrium.
Since the computation domain size for our simulations is only 8 R, the interface evolution near
the particles is also influenced by the conditions imposed at the domain boundary. On the domain walls,
the contact line does not remain flat; its level is lower in the middle of the domain walls and higher in the
domain corners. This causes the particles approach velocity to vary, as the lowering on contact lines on
the domain walls influences the magnitude of the lateral attractive force, especially when the gap between
the disk and a domain wall is comparable to the distance between the disks. This effect, which is due to a
finite size of the computational domain, can be diminished by performing simulations in larger domains.

1.3
6
0.8
d/R
u2-u1
4
0.3
2 -0.2 0 0.1 0.2 0.3

0 0.1 0.2 0.3 t
t
Figure VI.14. The approach velocity u2-u1 and the dimensionless distance d/R between the disks are plotted as
functions of time. The approach velocity initially increases as the interface height between the disks decreases and
then decreases as the gap between the particles becomes small. The lower fluid viscosity is 0.01 Poise and the initial
d/R= 6. The dimensionless parameters based on the maximum disk velocity are (Re=2.4, G=425.8 Ca=0.0024).
An interesting consequence of the lowering of the interface height between the disks is that the
average value of the contact angle around the disk decreases. This decrease reduces the vertical
component of capillary force which acts against gravity and keeps the disks floating. A reduction in the
vertical capillary force causes the disks to sink slightly increasing the contact angle everywhere except
between the disks. After this additional sinking, the buoyant weight again becomes equal to the vertical
component of the capillary force. This additional sinking, which also happens when two floating spheres
come near each other, suggests that the clusters of disks are more likely to sink than an isolated disk.
VI.2.2.1 Role of the particle density

In table VI.5 we have given results for the approach velocity for different values of the particle
density. A denser disk sinks more causing a greater interfacial deformation. The lateral force which arises
from the asymmetry is then greater and the velocity with which the disks approach each other is larger.
Table VI.5. The maximum approach velocity for the disks is listed to show the dependence on their density. The
lower fluid viscosity is 0.1 Poise and the density is 1.0 gm/cm3. The initial distance between the disks is 2.8 R.
ρp ρ (upper fluid) η (upper fluid) γ (u2-u1)max
1.1 1 0.1 10 0.0036

1.5 1 0.1 10 0.0762
1.5 0.1 0.01 10 0.034
2.0 0.1 0.01 10 0.066

VI.2.2.2 Role of the fluid density ratio
ρU
In this subsection we show how u2-u1 depends on lU = . The case of matched densities
ρL
ρ U = ρ L is interesting because the lowering of the interface does not cause an increase in Fp. This case is
also interesting because after the motion stops an equilibrium is reached in which the variation of pressure
does not depend on the interface shape. This implies that the interface must deform so that there is no
pressure jump across it; hence the mean curvature of the deformed interface is zero.
When the density ratio lU is decreased the disks sink to a lower depth and their lateral approach
velocities are smaller. The maximum approach velocities for two cases are listed in table VI.6.
Table VI.6. The maximum approach velocities for the disks are listed to show dependence on the upper fluid density.
The lower fluid viscosity is 0.1Poise and the particle density is 1.5 gm/cm3. The initial distance between the disks is
2.8 R.
ρlower ρ (upper fluid) η (upper fluid) γ (u2-u1)max
1.0 1 0.1 10 0.0762

1.0 0.1 0.1 10 0.039
VI.2.2.3 Role of the viscosity ratio

We know from analysis of section II that the viscosity does not play a role in determining the
equilibrium state of floating particles, but it does alter the velocity. The lateral velocity of the disk
ηU
decreases with increasing viscosity, but is not so strongly dependent on the viscosity ratio m= .
ηL
6.5
6
5.5 viscosity= 0.001
5 viscosity= 0.01
4.5
d/R
4
3.5
3
2.5
2
0 0.1 0.2 0.3
t

Figure VI.14. The dimensionless gap between the disks is plotted for two different values of the upper fluid viscosity.
As the disks are almost completely immersed, the effect of the upper fluid viscosity on the lateral disk velocity is
negligible.
ηU
In figure VI.14 we have shown results for two different values of . All other parameters are
ηL
kept fixed. The approach velocity increases, but only slightly as the viscosity ratio is decreased. These
velocities at a fixed distance of 2.7 R for the disks released at the initial distance of 2.8 R. are listed in
table VI.7. The velocity increase is due to a decrease in the viscous resistance of the upper fluid. The
increase in the velocity is greater for the heavier particles and also for larger values of lU; in these cases
the interfacial deformation is greater.
Table VI.7. The maximum approach velocities for the disks are listed to show the dependence on the upper fluid
viscosity. The lower fluid viscosity is 0.1 Poise and the particle density is 2.0 gm/cm3. The initial distance between
the disks is 2.8 R.
ρlower ρ (upper fluid) η (upper fluid) γ (u2-u1)max
1.0 0.01 0.01 10 0.291

1.0 0.01 0.005 10 0.297
VII. Conclusions
This is the first paper on the direct numerical simulation of the motion of solid floating particles.
A floating smooth particle, embedded in the interface between two fluids, touches the interface at a fixed
contact angle, which at equilibrium is given by the Young-Dupré law. For a floating prismatic particle, on
the other hand, the contact line is assumed to be fixed at its sharp edge, provided the contact angle at the
edge is within the limits specified by the Gibbs extension to the Young-Dupré law. For example, in our
simulations of disks, floating with their axis normal to the interface, the contact line is assumed to be
pinned at the top sharp edge of the disk.
The constraint that the contact line remains pinned at a sharp edge of a prismatic particle is
different from that for a smooth particle, such as a sphere. For a smooth particle, the contact line can
move, but the contact angle is fixed and here is assumed to be that at equilibrium given by the Young-
Dupré law. The problem of the motion of the contact line is difficult and unsolved. The problem is that
the no slip condition implies that the contact point cannot move which contradicts the behavior seen in
experiments (see Dussan and Davis 1974, Dussan 1976 and Kistler and Scriven 1993 and references
therein). Several approaches to this problem have been proposed and could be implemented in direct
numerical simulations (see Sussman 2001, Friedrichs and Guceri 1993, Kamal, Goyal and Chu 1988, and

references therein). Here we look at one approach which could be called capillary induced motion of the
contact line due to a prescribed contact angle. In this approach, we let the contact line adjust to the
dynamics under the constraint that the contact angle is fixed. The no slip condition is satisfied before and
after the contact line moves. In another approach, not implemented here, the contact line is allowed to slip
by relaxing the no slip condition by putting the velocity of the contact line proportional to the shear stress
there. This approach requires us to relax the prescribed contact angle condition. Another feature, not
treated here, which impacts the floating particle problem is the existence of advancing and receding
contact angle. The subject is very rich with many possibilities that might be studied in direct simulations.
The Marchuk-Yenenko operator-splitting technique is used to decouple the four primary
numerical difficulties of the governing equations: incompressibility constraint, nonlinear convection term,
Lagrange multiplier problem that forces rigid body motion within the particles, and the interface
advection problem. The resulting four sub-problems are solved efficiently using matrix free approaches.
Specifically, the DLM approach is used for enforcing rigid body motion inside the particles, the level set
approach for tracking the interface and the conjugate gradient algorithms are used to enforce the
incompressibility constraint and to solve the nonlinear convection problem. The contact line is kept at a
sharp edge by making the level set function zero along the edge. On smooth surfaces the contact angle is
kept fixed by using the technique described above. The code is validated by simulating the time
dependent motion of a particle released on a two-fluid interface to steady state.
Our simulations show that the interface shape near a sphere adjusts quickly after it is released to
meet the contact angle requirement, but away from the sphere the interface shape takes a much longer
time interval to adjust during which the vertical position of particle changes. The computed results are
shown to be independent of the mesh resolution as well as the size of time step. The steady state limits of
our dynamic simulations agree with the equilibrium results. The time dependent simulations of two or
more floating particles lead to capillary attraction and the formation of clusters. The attractive capillary
force arises due to the asymmetric interface deformation around the particles. Specifically, the interface
height between the particles is lowered giving rise to a lateral attractive force due to capillarity. This
lowering of the interface also decreases the net vertical component of the capillary force which causes the
floating height of the particles to decrease. These results agree qualitatively with the experimental data as
well as the equilibrium analysis.
For floating disks, simulations are started by assuming that the interface is flat and that the top
surface of the disk is in the plane of the interface ensuring that the contact line is initially pinned at the
sharp rim. The disk assumed to be denser than the liquid sinks into the lower liquid. As it sinks, the
contact angle increases which in turn increases the vertical component of the capillary force. The disk
stops sinking when its buoyant weight is balanced by the capillary force. The code is validated by

simulating the time dependent motion of a disk released on a two-fluid interface to steady state.
Simulations show that as the disk sinks, the interface near the disk moves downward, as the contact line is
pinned to the edge, but away from the disk the interface shape takes a longer time interval to adjust during
which the vertical position of the disk changes. The computed results are independent of the mesh
resolution as well as the size of the time step. The steady state limits of our dynamic simulations agree
with the equilibrium results.
Simulations also show that the lowering of the interface between disks pinned to the interface at
the rim reduces the contact angle between the disks giving rise to the unbalanced attractive capillary
force. Two floating spheres also attract, but the mechanism by which the attractive force arises is
different. Specifically, the lateral force on two floating spheres arises due to the lowering of the contact
line on the surfaces between the two spheres which is a consequence of the constraint that the contact
angle is fixed. For floating disks, on the other hand, the contact line is fixed and the lateral force arises
due to a decrease in the contact angle between the disks due to the lowering of the interface. The lowering
of the interface also decreases the net vertical component of the capillary force which causes the floating
height of the particles to decrease.
Though our floating particle code has been optimized in various ways it still runs slowly.
Improvements can be made by parallelizing the code for machines with large clusters. The security zone
which has been used in previous direct numerical simulation based studies, and this one, does not allow
contact between particles and hence is not faithful to the crystal structures which can evolve in problems
of self assembly due to capillary attraction. Fortunately, the new collision scheme put forward by Singh,
Hesla and Joseph 2003 does allow particle-particle contact and close packing. We will implement this
new strategy in our future work.
Acknowledgements
This paper is dedicated to Professor Stan Osher on the occasion of his 60th birthday. This work
was partially supported by National Science Foundation KDI Grand Challenge grant (NSF/CTS-98-
73236) and a GOALI grant NSF/CTS-0109079, Engineering Research Program of the Office of Basic
Energy Science at DOE, a grant from Schlumberger foundation, from STIM-LAB Inc., New Jersey
Commission on Science and Technology through the New-Jersey Center for Micro-Flow Control under
(Award Number 01-2042-007-25) and the University of Minnesota Supercomputing Institute.

Appendix. Weak form of equations and finite-element discretization
The approach used for obtaining the weak form of the governing equations was described in Glowinski, et
al. 1999 and Singh, et al. 2000. In obtaining this weak form, the hydrodynamic forces and torques acting on the
particles can be completely eliminated by combining the fluid and particle equations of motion into a single weak
equation of motion for the combined fluid-particle system. The hydrodynamic stresses acting at the interface are also
completely eliminated. For simplicity, in this section we will assume that there is only one particle. The extension to
the many-particle case is straightforward.
The solution and variation are required to satisfy the strong form of the constraint of rigid body
motion throughout P(t). In the distributed Lagrange multiplier method this constraint is removed from the
velocity space and enforced weakly as a side constraint using a distributed Lagrange multiplier term. It
was shown in Glowinski, et al. 1999 that the following weak formulation of the problem holds in the
extended domain:
For a.e. t>0, find u ∈ W uΓ , p ∈ L20 (Ω) , λ ∈ Λ ( t ) , U ∈ R3, ω ∈ R3 and φ ∈ Wφ , satisfying
⎛ du ⎞
∫Ω ρ⎜ − g ⎟ ⋅ vdx − ∫Ω p∇.v dx + ∫Ω 2ηD[u] : D[ v ]dx
⎝ dt ⎠
⎛ ρ ⎞⎛ ⎛ dU ⎞ dω ⎞
+ ⎜⎜1 − ⎟⎟⎜⎜ M⎜ − g⎟⋅ V + I ξ ⎟⎟ − F ' .V − ∫Ω γκδ(φ)n ⋅ v dx = λ , v − (V + ξ × r )
⎝ ρd
P( t )
⎠⎝ ⎝ dt ⎠ dt ⎠
for all v ∈ W 0 , V ∈ R3, and ξ ∈ R3, (1)
∫Ω q∇.u dx = 0 for all q ∈ L2 (Ω) , (2)
µ, u − ( U + ω × r ) P( t )
=0 for all µ ∈ Λ ( t ) , (3)
u |t = 0 = u o in Ω, (4)
⎛ ∂φ ⎞
∫ ⎜⎜⎝
Ω ∂t
+ u .∇φ ⎟⎟ g dx = 0 for all g ∈ Wφ0,
⎠
(5)
φ | t =0 = φ 0 in Ω.
as well as the kinematic equations and the initial conditions for the particle linear and angular velocities.
Here F’ is the additional body force applied to the particles to limit the extent of overlap (see equation
(19) in Glowinski, et al. 1999, λ is the distributed Lagrange multiplier,
W uΓ = {v ∈ H1 (Ω) 3 | v = u Γ ( t ) on Γ},
W 0 = H10 (Ω) 3 ,
L20 (Ω) = { q ∈ L2 (Ω) | ∫ q dx = 0 } ,

Ω
Wφ = {φ ∈ H1 (Ω) | φ = φ 0 ( t ) on Γ − },

Wφ0 = {φ ∈ H1 (Ω) | φ = 0 on Γ − },
L20 (Ω) = { q ∈ L2 (Ω) | ∫Ω q dx = 0 } , (6)
where Γ − is the upstream part of Γ and Λ ( t ) is L2 (P( t )) 3 , with .,. P(t)

denoting the L2 inner product
over the particle. In our simulations, since the velocity and µ are in L2, we will use the following inner
product
µ, v P(t)
= ∫P ( t ) (µ.v ) dx . (7)
In order to solve the above problem numerically, we will discretize the domain using a regular
tetrahedral mesh Th for the velocity, where h is the mesh size, and a regular tetrahedral mesh T2h for the
pressure. The following finite dimensional spaces are defined for approximating W uΓ , W 0 , L2 (Ω) ,
L20 (Ω) , Wφ and Wφ0 :
WuΓ ,h = {v h ∈ C 0 ( Ω ) 3 | v h |T ∈ P1 × P1 × P1 for all T ∈ Th , v h = u Γ ,h on Γ} ,
W0,h = {v h ∈ C 0 ( Ω ) 3 | v h | T ∈ P1 × P1 × P1 for all T ∈ Th , v h = 0 on Γ} ,
L2h = {q h ∈ C 0 ( Ω ) | q h | T ∈ P1 for all T ∈ T2 h } ,
L20,h = {q h ∈ L2h | ∫Ω q h dx = 0} ,
W φn, h = {g h ∈ H1 (Ω) | g h |T ∈ P1 for all T ∈ Th , g h = φ n on Γ − }
Wφ0,h = {g h ∈ H1 (Ω) | g h |T ∈ P1 for all T ∈ Th , g h = 0 on Γ − }
WφR ,h = {g h ∈ H1 (Ω) | g h |T ∈ P1 for all T ∈ Th , g h = 0 on the interface} . (8)
The particle inner product terms in (1) and (3) are obtained using the discrete L2 inner product defined in
Glowinski, et al. (1999). Specifically, we choose M points, x1,…,xM that uniformly cover P ( t ) , and
define
⎧⎪ M ⎫⎪
Λ h ( t ) = ⎨µ h µ =
⎪⎩
∑µ
i =1
h ,i δ(x − x i ), µ h ,1 ,..., µ h ,M ∈ R 3 ⎬ .
⎪⎭
Using these finite dimensional spaces, it is straightforward to discretize equations (1) - (5). Notice that the
discrete space Wφ,h assumes that φ is known on the upstream portion of the boundary. This is not a
problem even when φ(t) is not known on the upstream boundary in advance because the imposed
boundary value can be corrected during the reinitialization step. Since only the zero level set of φ(t) is
physically relevant, we have a lot of freedom in treating φ(t) away from the interface. In our code, the

value from the previous time step is used as the boundary value. The reinitialization space WφR ,h
assumes that φ remains zero along the interface which is done by imposing Dirichlet boundary condition,
φ = 0, along the interface during reinitialization iterations.
A.1 Strong form

The strong form for the weak formulation can be obtained by integrating the stress term by parts. The
resulting equations inside the region occupied by the fluid Ω\ P( t ) are
⎡ ∂u ⎤
ρ ⎢ + u. ∇ u ⎥ = ρg - ∇p + ∇.σ +γ κ δ(φ) n in Ω\ P(t)
⎣ ∂t ⎦
∇. u = 0 in Ω\ P(t)
u = uL on Γd
u=U+ωxr on ∂P( t )
and the equations inside the region occupied by the particles P(t) are
⎡ ∂u ⎤
ρ ⎢ + u. ∇ u ⎥ = ρg - ∇p + ∇.(2ηD) + λ − R 2 ∇ 2 λ (9)
⎣ ∂t ⎦
u = Ui + ωi x ri on ∂Pi ( t ) , i=1,…,N,
Here we have used the fact that the rigid body motion satisfies the incompressibility constraint and that D(u)
inside the particles is zero. The boundary condition for λ on the interface between the fluid and particle regions
∂ P(t) is
n.(−σ L ) = n.∇λ (10)
where n is the normal at the fluid-particle interface, and σ L = -pI + 2ηD is the stress in the fluid phase, and
σ P = 0 is the stress inside the particles. For given U(t) and ω(t), and the positions Xi(t), i=1,…, N, equation (9)
can be written as:
⎛ dU dω ⎞
λ − R 2∇ 2 λ = ρL ⎜ + × r + ω × (ω × r ) − g ⎟ (11)
⎝ dt dt ⎠
It is shown in Joseph 2002, chapter IV, that it is possible to completely eliminate the Lagrange multiplier
λ from the strong form.
A.2 Time discretization using the Marchuk-Yanenko operator splitting scheme
The discretized initial value problem (1-5) is solved by using the Marchuk-Yanenko operator splitting
scheme. This allows us to decouple its four primary difficulties:
1. The incompressibility condition and the related unknown pressure ph,

2. The nonlinear convection term,
3. The rigid body motion problem inside the particles,
4. The interface problem and the related unknown level set distribution φh,
The Marchuk-Yanenko operator splitting scheme can be applied to an initial value problem of the
form
dφ
+ A1 (φ) + A 2 (φ) + A 3 (φ) + A 4 (φ) = f
dt
where the operators A1, A2, A3, and A4 can be multiple-valued.
Let ∆t be the time step, and α1 and α2 be two constants: 0≤α1, α2 ≤1 and α1 + α2 = 1. We use the
following version of the Marchuk-Yanenko operator splitting to simulate the motion of particles on two-
fluid interfaces:
Set u0 = u0,h, and φ 0 = φ 0,h .
For n = 0,1,2,… assuming un, and φn are known, find the values for n+1 using the following:
STEP 1:
Find un+1/4 ∈ Wu ,h and pn+1/4 ∈ L20, h , by solving
u n +1 / 4 − u n n +1 / 4
∫Ω ρ .v dx − ∫Ω p ∇.v dx = ∫Ω γκδ(φ)n ⋅ v dx
∆t
for all v ∈ W0, h ,
n +1 / 4
∫Ω q∇.u dx = 0 for all q ∈ L2h , (12)
STEP 2:
Find un+2/4 ∈ Wu ,h , by solving
u n + 2 / 4 − u n +1 / 4
∫Ω ρ .vdx + ∫Ω ρ(u n + 2 / 4 .∇u n + 2 / 4 ).v dx + α 1 ∫Ω 2ηD[u n + 2 / 4 ] : D[ v]dx = 0
∆t
for all v ∈ W0, h , (13)
STEP 3:
Compute Un+3/4 and Xn+3/4 using the prediction procedure

Set Un,0= Un, Xn,0 = Xn.
Do k=1,K
⎛ ⎛ ρ ⎞
−1
⎞ ∆t
U *n,k
=Un,k-1
+ ⎜ g + ⎜⎜1 − ⎟ M −1 F' (X n , k −1 ) ⎟
⎟
⎜ ⎟K
⎝ ⎝ ρd ⎠ ⎠
⎛ U n , k −1 + U *n , k ⎞ ∆t
X*n,k = Xn,k-1 + ⎜⎜ ⎟⎟
⎝ 2 ⎠K
⎛ ⎛ ρ ⎞
−1
F' (X n ,k −1 ) + F' (X *n ,k −1 ) ⎞⎟ ∆t
Un,k = Un,k-1 + ⎜ g + ⎜⎜1 − ⎟⎟ M −1
⎜ ⎝ ρd ⎠ 2 ⎟K
⎝ ⎠
⎛ U n , k −1 + U n , k ⎞ ∆t
Xn,k = Xn,k-1 + ⎜⎜ ⎟⎟
⎝ 2 ⎠K
end do
Set Un+3/4= Un,K, Xn+3/4 = Xn,K. (14)
Find un+1 ∈ WunΓ+,1h , λn+1 ∈ Λ h ((n + 2 / 4)∆t ) , Un+1 ∈ R3, and ωn+1 ∈ R3, satisfying
u n +1 − u n + 2 / 4 ⎛ ρ ⎞⎛ U n +1 − U n +3 / 4 ω n +1 − ω n +3 / 4 ⎞
∫Ω ρ .v dx + ⎜⎜1 − ⎟⎟⎜ M .V + I ξ ⎟⎟ +
∆t ⎜
⎝ ρd ⎠⎝ ∆t ∆t ⎠
γ ∫Ω 2ηD[u n +1 ] : D[ v]dx = λ n +1 , v − (V + ξ × r n +3 / 4 )
P (( n + 3 / 4 ) ∆t )
for all v ∈ W0, h , V ∈ R3, and ξ ∈ R3
µ h , u n +1 − ( U n +1 + ω n +1 × r ) =0 for all µ h ∈ Λ ((n + 3 / 4)∆t ) , (15)

P (( n + 3 / 4 ) ∆t )
where the center of particle P((n + 3 / 4)∆t ) is at X n +3 / 4 . For a disk the moment of inertia changes as the disk
rotates and so we also keep track of its orientation.
Set Xn+1,0 = Xn.
Do k=1,K
⎛ U n + U n +1 ⎞ ∆t
X*n+1,k = Xn+1,k-1 + ⎜⎜ ⎟⎟
⎝ 2 ⎠K
−1
⎛ ρ ⎞ ⎛ F' (X n +1, k −1 ) + F' (X *n +1, k ) ⎞ (∆t ) 2
Xn,k = X*n,k-1 + ⎜⎜1 − ⎟⎟ M −1 ⎜ ⎟
ρ ⎜ 2 ⎟ 2K
⎝ d ⎠ ⎝ ⎠
end do
Set Xn+1 = Xn+1,K.

STEP 4:
Find φn+4/4 ∈ Wφn,h , by solving
⎛ φ n+4 / 4 − φ n ⎞
∫ ⎜
Ω⎜
⎝ ∆t
+ u n +1 .∇φ n +1 ⎟⎟ g h dx = 0
⎠
for all gh ∈ Wφ0,h. (16)
Set pn+1= pn+1/4, φn+1 = φn+4/4.
STEP 4.1:
Reinitialize φ n+1
Set φ 0R =φn+1
For r = 0, 1, 2, …
∇φrR
w r = S(φn +1 )
∇φrR
Find φ rR+1 ∈ WφR , h , by solving
⎛ φ rR+1 − φrR ⎞
∫Ω ⎜⎜ + w r .∇φrR ⎟ g h dx = ∫ S(φn +1 ) g h dx
∆t ⎟
⎝ ⎠ Ω
for all gh ∈ WφR,h. (17)
go back to the above for loop.
STEP 4.2:
Modify φn+1 inside and on the particle surfaces to enforce contact angle
Set φ 0R =φn+1
For r = 0, 1, 2, …
r ⎧0 outside particles
u ex =⎨
⎩ u ex from (12) otherwise
Find φ rR+1 ∈ WφR , h , by solving

⎛ φ rR+1 − φ rR ⎞
∫Ω ⎜⎜ + u ex .∇φ rR ⎟⎟ g h dx = 0
r
⎝ ∆t ⎠
for all gh ∈ WφR,h. (18)
go back to the above for loop.
Set φn+1 = φ rR+1 and go back to the first step.
The decoupled sub-problems can be solved much more efficiently than the original problem (1-5). In our
code all of these sub-problems are solved using matrix-free algorithms, which reduces the memory
required.
Remarks:
1. The first step gives rise to a L2 projection problem for the velocity and pressure distributions which is
solved by using a conjugate gradient method (Glowinski, Tallec, Ravachol and Tsikkinis 1992 and
Glowinski, et al. 1999).
2. The second step is a nonlinear problem for the velocity, which is solved by using a least square
conjugate gradient algorithm (Glowinski and Pironneau 1992).
3. In this paper, we will assume that α1 =1, α2 = 0.
4. The fourth step is a hyperbolic problem for the scalar level set function φ. This problem is solved by
using a upwinding scheme where the advection term is discretized using a third order scheme (Glowinski
and Pironneau 1992 and Pillapakkam and Singh 2001).
5. After advecting φ according to (16), we reinitialize φ to be a distance function near the interface by
performing one or two iterations of (17) using a fast algorithm developed in Pillaipakkam and Singh
2001. The interface position φ(t) = 0 does not change during these reinitialization iterations. It is
necessary to re-initialize φ to ensure that the scheme accurately conserves mass.
6. The level set function is extended inside and on the surface of the particles such that the contact angle
is 135º by performing one or two iterations of (18).
5. References
1. Bowden, N., I.S. Choi, B.A. Grzybowski, G.M. Whitesides, Mesoscale self-assembly of hexagonal
plates using lateral capillary forces: synthesis using the “capillary bond”, J. Am. Chem. Soc. 121,
5373-5391 (1999).

2. Bowden, N., A. Terfort, J. Carbeck, G.M. Whitesides, Self-assembly of mesoscale objects into
ordered two-dimensional arrays, Science, 276, 233-235 (1997).
3. Brenner, H. and L.G. Leal. A micromechanical derivation of Fick’s law for interfacial diffusion of
surfactant molecules, J. Colloid Interface Sci. 65, 191 (1978).
4. Brenner, H. and L.G. Leal. Conservation and constitutive equations for adsorbed species undergoing
surface diffusion and convection at a fluid-fluid interface, J. Colloid Interface Sci., 88, 136 (1982).
5. Bristeau, M.O., R. Glowinski & J. Periaux. Numerical methods for Navier-Stokes equations.
Application to the simulation of compressible and incompressible flows, Computer Physics Reports,
6, 73 (1987).
6. Chan, D.Y.C., J.D. Henry Jr. and L.R. White. The interaction of colloidal particles collected at the
fluid interface, J. Colloid Interface Sci. 79, 410 (1981).
7. Danov, K.D., R. Aust, F. Durst and U. Lange. Influence of the surface viscosity on the hydrodynamic
resistance and surface diffusivity of a large Brownian particle, J. Colloid and Interface Science, 175,
36-45 (1995).
8. Dussan V., E.B. and S.H. Davis. On the motion of a fluif-fluid interface along a solid surface. J. Fluid
Mech. 65, 71 (1974).
9. Dussan V., E.B. The moving contact line: the slip boundary condition. J. Fluid Mech. 77, 665-684
(1976).
10. Fortes, M.A. Attraction and repulsion of floating particles, Can. J. Chem. 60, 2889 (1982).
11. Friedrichs, B. and S.I.Guceri. A novel hybrid numerical technique to model 3-D fountain flow in
injection molding processes, J. Non-Newtonian Fluid Mech. 49, 141-173 (1993).
12. Gerson, D.F., Zaijc, J.E. and Ouchi, M.D. Chemistry for energy, Ed. M. Tomlinson, ACS symposium
series, 90, American Chemical Society, Washington DC, 66 (1979)
13. Gifford, W.A. and L.E. Scriven. On the attraction of floating particles, Chem. Engrg. Sci. 26, 287-297
(1971).
14. Glowinski, R., T.W. Pan, T.I. Hesla & D.D. Joseph. A distributed Lagrange multiplier/fictitious
domain method for particulate flows, Int. J. Multiphase Flows 25(5), 755 (1999).
15. Glowinski, R. and O. Pironneau. Finite element methods for Navier-Stokes equations, Annu. Rev.
Fluid Mech. 24, 167 (1992).
16. Glowinski, R., P. Tallec, M. Ravachol & V. Tsikkinis. Chap. 7 in Finite Elements in Fluids, Vol 8,
Ed. T.J. Chung, Hemisphere Publishing Corp., Washington DC (1992).
17. Goldman, A.J., R.G. Cox and H. Brenner. Slow viscous motion of a sphere parallel to a plane wall—I
Motion through a quiescent fluid, Chem. Eng. Sci. 22, 637-653 (1967).
18. Grzybowski, B.A., N. Bowden, F. Arias, H. Yang, G.M. Whitesides, Modeling of menisci and
capillary forces from the millimeter to the micrometer size range, J. Phys. Chem. B 105, 404-412.
(2001).
19. Hesla, T.I. and D.D. Joseph, The maximum contact angle at the rim of a heavy floating disk,
submitted to J. Colloid and Interface Science (2003).
20. Hirt, C.W. and B.D. Nichols, Volume of fluid (VOF) methods for the dynamics of free boundaries, J.
Comput. Phys. 39, 201 (1981).
21. Joseph, D.D., Interrogations of Direct Numerical Simulation of Solid-Liquid Flows, Web based book:
http://www.efluids.com/efluids/books/efluids_books.htm, (2002).
22. Joseph, D.D., J. Wang, R. Bai, B.H. Yang and H.H. Hu, Particle motion in a liquid film rimming the
inside of a rotating cylinder. J. Fluid Mech., 496, 139-163 (2003).
23. Kamal, M.R., S.K. Goyal and E. Chu. Simulation of injection mold filling of viscoelastic polymer
with fountain flow, AIChE Journal, 34, 94-106 (1988).

24. Katoh, K., H. Fujita and E. Imazu. Motion of a particle floating on a liquid meniscus surface, J.
Fluids Engrg., 114, 411 (1992).
25. Keller, J.B., Surface tension force on a partly submerged body, Physics of Fluids, 10, 3009-3010
(1998).
26. Kistler, S.F. and L.E. Scriven. The teaplot effect: sheet-forming flows with deflection, wetting and
hysteresis, J. Fluid Mech. 263, 19-62 (1994).
27. Kralchevsky, P.A., V.N. Paunov, N.D. Denkov, I.B. Ivanov and K. Nagayama. Energetical and force
approaches to the capillary interactions between particles attached to a liquid-fluid interface, J.
Colloid and Interface Sci. 155, 420-437 (1993).
28. Kralchevsky, P.A., V.N. Paunov, N.D. Denkov, and K. Nagayama., J. Colloid and Interface Sci. 167,
47 (1994).
29. Kralchevsky, P.A., V.N.Paunov, I.B.Ivanov and K.Nagayama, "Capillary Meniscus Interactions
between Colloidal Particles Attached to a Liquid - Fluid Interface", J. Colloid Interface Sci. 151 79 -
94 (1992).
30. Kralchevsky, P. A. and K. Nagayama, "Capillary interactions between particles bound to interfaces,
liquid films and biomembranes", Advances in Colloid and Interface Science, 85, 145-192 (2000).
31. Manon, V.B. and D.T. Wasan. Particle-fluid interactions with application to solid-stabilized
emulsions. Part 1, The effect of asphaltene adsorption. Colloids Surf. 19, 89-105 (1986).
32. Marchuk, G.I. Splitting and alternate direction methods. Handbook of Numerical Analysis, P.G.
Ciarlet and J.L. Lions (Eds.), Volume I, 197-462. Amsterdam: North-Holland (1990).
33. Nicolson, M.M. The interaction between floating particles, Proc. Cambridge Philosophical Soc., 45,
288 (1949).
34. Osher, S. and J.A. Sethian. Fronts propagating with curvature-dependent speed: Algorithms based on
Hamilton-Jacobi formulations, J. Comput. Phys. 83, 12 (1988).
35. Petkov, J.T., N.D Denkov, K.D Danov, O.D. Velev, R. Aust and F Durst. Measurement of the drag
coefficient of spherical particles attached to fluid interfaces, J. Colloid and Interface Science, 172,
147-154 (1995).
36. Pillapakkam, S.B. and P. Singh. A Level Set Method for computing solutions to viscoelastic two-
phase flow, J. Computational Physics, 174, 552-578 (2001).
37. Poynting, J.H. and J.J. Thompson. A Text-book of Physics: Vol. 1, Properties of Matter, C. Griffith &
Co. Ltd, London, 153-155 (1913).
38. Princen, H.M. Equilibrium shape of interfaces, drops and bubbles. Rigid and deformable particles at
interfaces, Surface and Colloid Science, E. Matijevie, ed., Interscience, New York, Vol. 2, p.1 (1969).
39. Rapacchietta A.V. and A.W. Neumann. Force and free-energy analyses of small particles at fluid
interfaces: II. Spheres, J. Colloid and Interface Sci., 59(3), 555-567 (1977).
40. Redoev, B., M. Nedjalkov and V. Djakovich. Brownian motion at liquid-gas interfaces. I. Diffusion
coefficients of maroparticles at pure interfaces, Langmuir, 8, 2962 (1992).
41. Ryskin, G. & L.G. Leal. Numerical solution of free-boundary problems in fluid mechanics. Part 1:
The finite-difference technique, J. Fluid Mech. 148, 1 (1984).
42. Saif, T.A. On the capillary interaction between solid plates forming menisci on the surface of a liquid,
J. Fluid Mech., 453, 321-347 (2002).
43. Scardovelli, R. and Zaleski, S., Direct Simulation of Free-surface and interfacial flow, Annu. Re. Of
Fluid Mech. 1999 31: 567-603.
44. Schneider, Y.C., M.E. O’Neill, and H. Brenner. On the slow viscous rotation of a body straddling the
interface between two immiscible semi-infinite fluids, Mathematika, 20, 175 (1973).

45. Singh, P. and T.I. Hesla, Torque due to interfacial tension on a floating sphere,
http://web.njit.edu/~singhp/ research/publication.html (2003).
46. Singh, P., T.I. Hesla and D.D. Joseph, A Modified Distributed Lagrange Multiplier/Fictitious Domain
Method for Particulate Flows with Collisions, Int. J. of Multiphase Flows, 29, 495-509 (2003).
47. Singh, P., D.D. Joseph, T.I. Hesla, R. Glowinski & T.W. Pan. A distributed Lagrange
multiplier/fictitious domain method for viscoelastic particulate flows, J. Non-Newtonian Fluid Mech.
91, 165 (2000).
48. Singh, P. and L.G. Leal. Finite element simulation of the start-up problem for a viscoelastic problem
in an eccentric cylinder geometry using third-order upwind scheme, Theoret. Comput. Fluid
Dynamics, 5, 107-137 (1993).
49. Sussman, M. An Adaptive Mesh Algorithm for Free Surface Flows in General Geometries, Adaptive
Method of lines, Editors, A. Vande Wouwer, Ph. Saucez and W.E. Scheisser, Chapman and
Hall/CRC, 207-227 (2001).
50. Sussman, M., P. Smereka & S. Osher. A level set approach for computing solutions to incompressible
two-phase flow, J. Comput. Phys. 114, 146 (1994).
51. Unverdi, S.O. & G. Tryggvason. A front-tracking method for viscous, incompressible, multi-fluid
flows, J. Comput. Phys. 100, 25 (1992).
52. Wakiya, S. Niigata Univ. College of Engng. Res. Rept. 6, Nagoaka, Japan (March 30, 1957).
53. Wang, J., R. Bai, and D.D. Joseph, to be submitted (2004).
54. Yan N. and J.H. Masliyah. Absorption and desorption of clay particles at the oil-water interface. J.
Colloid and Interface Science, 168, 386-392 (1994).

XVII Epilogue
In this document I have explored some of the ways in which direct numerical simulation
(DNS) of solid-liquid flow can be interrogated for useful results. The value added by DNS is that
initial value problems for particulate flows can be solved as exactly as numerical methods will
allow. The signal feature of DNS is that the particles are moved by forces computed from the
fluid motion, as they should be; the modeling of forces needed by other approaches to particulate
flow is not needed or done by DNS. In our work the calculation of forces is implicit because the
mutual forces disappear in the variational formulation for the total solid-liquid momentum. In
this we avoid the explicit computation of forces, on the one hand, and the modeling of forces on
the other.
We developed two types of finite element packages: an ALE particle mover in which
particles move on a body-fitted unstructured grid and a DLM particle mover in which particles
move over a fixed triangular grid. The DLM calculation can be said to be over a fluid which is
partitioned into particle and particle-free parts; the field of multipliers is defined on particles and
is chosen so that the fluid there moves as a rigid body. The multiplier field is coupled with the
rigid motion constraint in a manner analogous to the way in which the pressure is associated with
the constraint of incompressibility.
The DLM method is very efficient; it runs matrix free, uses fast solvers and runs in parallel
using multilevel preconditioners. With this method we can track thousands of particles in 3D.
The number 1204 in the DLM simulation of the fluidization of 1204 spheres was chosen to
simulate the motion of 1204 real spheres which we had previously used in experiments. This
number is not near the limiting and unknown number for which such simulations may fail. It
takes days to compute seconds of bed expansion, but at least we can compute and the computing
will only get faster and better. It is probable that motions of tens of thousands of particles in 3D
is an achievable goal short term.
Unstructured grids are convenient for complex geometries and unlike fixed grid methods
they can be programmed for local mesh adaptivity in regions of close approach of particles. The
ALE method carries the overhead of remeshing from time to time due to mesh distortion. Parallel
algorithms based on domain decomposition are presently under development for the ALE
methods.
There are many ways in which direct simulations may be interrogated for useful results. It
has to be understood that DNS does not replace theory even though it gives rise to exact
numerical solutions of the initial value problem for particulate flow. It is more appropriate to
think of DNS as a surrogate for experiments. The simulations have some great advantages; you
can suppress physical effects one at a time in simulations which is something that cannot be done
in experiments. Virtual experiments also have the potential to replace real experiments in
generating data which form the basis for correlations of the type used in engineering practice; the
generation of Richardson-Zaki correlations in our study of fluidization of 1204 spheres is an
example. Quantities needed for theory, like slip velocities are ever so much easier to determine
in simulations than in experiments.
Direct simulations lead to better understanding of flow fundamentals in situations otherwise
opaque. The fluidization by lift of slurries in horizontal conduits is a concept generated by direct
234 • Interog-Epilogue.doc
simulations. Liquid-solid flows are a nonlinear dynamical system which give rise to typical
bifurcations like that discussed in Chapter X, to periodic and even chaotic solutions. It is not
possible to study such bifurcation by analytical methods or two-fluid models. DNS may be the
only method to study bifurcations of particulate flow.
There are many technologies which depend critically on solid-liquid flows. Industries which
utilize such technologies typically control operations with simplified models of particulate flow
which can be run on PCs. It is widely believed and presently true that DNS is too slow and
expensive to guide field operations. Such beliefs should always be revisited because the rapid
expansion of software and hardware has a proven capacity for upward revision.
We have been trying to find the structure of data arising from DNS which can be used to
form useful models. There are different kinds of models; effective media and models which
require the modeling of forces on particles. Modeling forces is a big problem; for example, there
are no good models of lift forces in slurries. At the risk of being tiresome I want again to call
attention to the huge difference between the modeling of forces as is done in modeling and the
computing of forces as is done by DNS.
There are two kinds of models which require the modeling forces; models in which the fluid
motions are resolved by direct methods and the particles are moved by Newton’s laws using
modeled forces and two-fluid models. Two-fluid models may be regarded as arising from
averaging as was done in Chapter VI.
Effective media, two-fluid models do not require the modeling of forces. Our study of the
Rayleigh-Taylor instability arising in the direct numerical simulation of 6400 circular particles in
Chapter VIII is a good example of an effective media two-fluid model. In that study we regarded
the sedimenting suspension and the entrained fluid as another effective fluid. To realize the
model it was necessary to come up with an effective density and viscosity; modeling forces was
not required. It seems likely to me that the fluidization of 300 particles by lift studied in Chapter
X can be modeled as an effective fluid, with an effective viscosity, density and zero surface
tension. The waves which propagate on the top of the fluidized suspension look like waves
which develop in two-fluid situations.
Maybe effective media two-fluid models in which model assumptions can be tested by DNS
ought to be restricted to special situations like those mentioned in the prior paragraph. One two-
fluid model of particulate flow which covers all situations is not likely to be an achievable goal
since closures which work for some situations will not work for others.
In the literature one finds formulas for single particle lift which apply in special
circumstances mainly for low Reynolds numbers. Rigorously derived mathematical formulas for
the lift and drag on particles moving at finite Reynolds numbers do not exist and such formulas
are put forward only as empirical results following out of experimental data, from real
experiments and now from numerical experiments.
It can be said that two-fluid models and perturbations of Stokes flow have not worked all that
well. An alternative to these methods is the method of correlations used in this book. This
method leads from data to formulas. Data from experiments and numerical experiments are
processed in the same way. This method makes maximum use of computers and storage tapping
opportunities provided by new technology. I think that curve fitting plus computers and storage
gives rise to new and great opportunities for particulate flow and multiphase science. The secrets
are in the data and with digital technology we can interrogate this data.
Generating correlations from experiments is an old method on which many industrial
applications are based but it has come to have a bad name, viewed as empirical and not
fundamental. The great example is the Richardson-Zaki correlation which is the cornerstone of
fluidized bed practice. My enthusiasm for correlations has to do with the surprising emergence of
correlations from the simplest kind of post-processing of our numerical experiments. We have
done lift correlations for single particles and for the bed expansion of many particles in slurries.
The procedure we follow is to plot the results of our simulations in log-log plots of the relevant
variables. The surprise for us is that these plots frequently come up as straight lines giving rise to
power laws. For example, a single particle will lift-off in a Poiseuille flow at a certain Reynolds
number R = Ud/v for a given settling Reynolds number RG = ρf (ρp - ρc)gd3/η2. When we plotted
the lift off criterion from about 20 points we found that R = aRGn with an intercept a and slope n
in the log-log plot. The straight lines are impressively straight and we generated such
correlations for lift to equilibrium, for the bed expansion of many particles and in
non-Newtonian fluids. The existence of such power laws is an expression of self-similarity,
which has not been predicted from analysis or physics. The flow of dispersed matter appears to
obey those self-similar rules to a large degree.
We can get power laws when only two variables are at play; when there are three variables or
more, it would appear that we get different power laws separated by transition regions. This is
certainly the case for the Richardson-Zaki correlation; it has one power law relating the
fluidization velocity to the solids fraction at low Reynolds number, and another at high Reynolds
with a Reynolds number-dependent transition between. We got such correlations between three
variables for slurries, and from experiments (see Chapter XV).
The direction of our work is to develop simulations to get efficient computation leading to
3D correlations. This will happen. Then we will get real engineering correlations from numerical
experiments. I like this approach since it uses numerical simulations in a natural way evolving
from their intrinsic properties rather than trying to fit them into a more familiar frame using
models. I think that processing of data for correlations, from experiments, field data or
simulations is a great new opportunity of the computer age and ought to be vigorously pursued.
The problem faced by models is how to get the various interaction terms right. Much of the
time the guesses made for these interaction terms are poor and the predictive power of the model
is not there. Better models must also make use of correlations for the interaction terms. For
example, the Richardson-Zaki correlation gives an excellent correlation for bed expansion, but
leaves the modeling of the drag force needed for a mechanist’s model to imagination.
Let it be said that the active pursuit of correlations is an excellent direction for future
research using computers in a new way with direct applications to both engineering practice and
model construction.
§ Acknowledgment
This work was supported mainly by successive National Science Foundation grants, first as a
Grand Challenge HPCC Grant (1995-1999) and then as a KDI/NCC Grant (NSF/CTS-98-
73236). Our goals for these grants was to do state of the art high performance computing
directed strongly to applications and to form working collaborations between computational fluid
dynamics and computer science. The work was also supported by the Engineering Research
Program of the Office of Basic Energy Sciences DOE, and the ARO (Mathematics), by a grant
from the Schlumberger Foundation, from StimLab Inc. and by the Minnesota Supercomputer
Institute. Dave Vogel, my documentation specialist, did wonderful work in processing this
document, and I am indebted to Toshio Funada for finding ever so many corrections and for the
preparation of an excellent index. I am delighted to acknowledge my collaborators who did the
work reported here: B. Barree, H. Choi, M. Conway, R. Glowinski, T. Hesla, H. Hu, P. Huang,
M. Knepley, T. Ko, D. Ocando, T.-W. Pan, N.A. Patankar, Y. Saad, A. Sameh, V. Sarin, P.
Singh and J. Wang. I am grateful to M. Roco for his encouragement and support over the years.
XVI Epilogue ....................................................................................................................................234
§ Acknowledgment ......................................................................................................................237
[ ignore this page, only used to generate table of contents ]
Interrogations of Direct Numerical Simulation of Solid-Liquid Flow
References
C.K. Aidun, Y. Lu and E. Ding. Direct analysis of particulate suspensions with inertia using the
discrete Boltzmann equation, J. Fluid Mech. 373, 287, (1998).
T.B. Anderson and R. Jackson. A fluid mechanical description of fluidized beds. Equations of
motion, I&EC Fundamentals, 6, 527 (1967).
M.J. Andrews and P.J. O’Rourke. The multiphase particle-in-cell (MP-PIC) method for dense
particulate flows. Int. J. Multiphase Flow, 22, 379-402, (1996).
G.S. Arnold, D.A. Drew and R.T. Lahey. Derivation of constitutive equations for interfacial
force and Reynolds stress for a suspension of spheres using ensemble averaging. Chem Eng.
Comm. 86, 43 (1989).
E.S. Asmolov. The inertial lift on a spherical particle in a plane Poiseuille flow at large channel
Reynolds number, J. Fluid Mech. 381, 63-87 (1999).
E.S. Asmolov. Dynamics of s spherical particle in a laminar boundary layer, Fluid Dyn. 25, 886-
890 (1990).
T.R. Auton. The lift force on a spherical body in a rotational flow, J. Fluid Mech. 183, 199-218
(1987).
R. A. Bagnold. Fluid forces on a body in shear-flow; experimental use of ‘stationary flow’. Proc.
R. Soc. Lond. A. 340, 147-171 (1974).
G.J. Barenblatt. Scaling, self-similarity, and intermediate asymptotics. Cambridge Texts in
Applied Math. Cambridge University Press, UK ( 1996).
E. Barnea and J. Mizrachi. A generalized approach to the fluid dynamics of particulate systems,
Chem. Eng. Sci, 5, 171-189 (1973).
G.K. Batchelor. A new theory of the instability of a uniform fluidized bed, J. Fluid Mech. 193,
75-110 (1988).
T. B. Benjamin. Note on formulas for the drag of a sphere, J. Fluid Mech., 246, 335 (1993).
J.F. Brady. Stokesian Dynamics Simulation of Particulate Flows in Particulate Two-Phase Flow,
(M. Roco ed.) 971-998, Butterworth-Heinemann (1993).
H. Brenner. Hydrodynamic resistance of particles at small Reynolds numbers. Adv. Chem.
Engng, 6, 287, (1966).
F.P. Bretherton. Slow viscous motion round a cylinder in simple shear, J. Fluid Mech. 12, 591,
(1962).
M.O. Bristeau, R. Glowinski, J. Periaux. Numerical methods for the Navier-Stokes equations.
Applications to the simulation of compressible and incompressible viscous flow, Computer
Physics Reports, 6, 73-187 (1987).
S. Chen and G.D. Doolen. Lattice Boltzmann method for fluid flows, Annual Rev. Fluid Mech.,
30, 329-364 (1998).
printed 03/04/02 241 • Interog-ref.doc

P. Cherukat and J. McLaughlin. The inertial lift on a rigid sphere in a linear shear flow field near
a flat wall, J. Fluid Mech. 263, 1–18 (1994).
P. Cherukat, J. B. McLaughlin, and A. L. Graham. The inertial lift on a rigid sphere translating in
a linear shear flow, Int. J. Multiphase Flow, 20, 339--353 (1994).
H. Choi, Splitting method for the combined formulation of the fluid-particle problem, Computer
Methods in Applied Mechanics and Engineering, 190(11-12), 1367-1378 (2000).
H. Choi and D.D. Joseph. Fluidization by lift of 300 circular particles in plane Poiseuille flow by
direct numerical simulation. J. Fluid Mech., 438, 101-128 (2001).
A.J. Chorin. A numerical method for solving incompressible viscous flow problems, J. Comp.
Phys. 2, 12–26 (1967).
A.J. Chorin. Numerical solution of the Navier-Stokes equations. Math. Comput. 22, 745-762
(1968)a.
A.J. Chorin. On the convergence and approximation of discrete approximation to the Navier-
Stokes equations, Math. Comp. 23, 341–353 (1968)b.
A.J. Chorin. Numerical study of slightly viscous flow, J. Fluid Mech. 57, 785–796 (1973).
R.G. Cox. The steady motion of a particle of arbitrary shape at small Reynolds numbers, J. Fluid
Mech. 23, 625-642 (1965).
R.G. Cox and H. Brenner. The lateral migration of solid particles in Poiseuille flow. Part 1
Theory. Chem. Eng. Sci. 23, 147-163 (1968).
R.G. Cox and S.K. Hsu. The lateral migration of solid particles in a laminar flow near a plane.
Intl. J. Multiphase Flow, 3, 201 (1977)
R.G. Cox and S.G. Mason. Suspended particles in fluid flow through tubes, Ann. Rev. Fluid
Mech. 3, 291, (1971).
R.G. Cox, I.Y. Zia and S.G. Mason. Particle motions in sheared suspensions. XXV. Streamlines
around cylinders and spheres, J. Colloid Interface Sci. 27, 7 (1968).
C.T. Crowe, J.N. Chung, T.R. Troutt. Numerical models for two-phase turbulent flows. Annual
Review of Fluid Mechanics, 28, 11-43 (1996).
J.M. Dallavalle. Micrometrics, 2nd Edition, Pitman, London (1948)
J.M. Delhaye. General equations of two-phase systems and their applications to air-water bubble
flow and to steam-water flashing flow. ASME-AIChE Heat Transfer Conference, Minneapolis
(1969).
D.A. Drew. Averaged field equations for two-phase media, Stud. Appl. Math. 50, 133 (1971).
D.A. Drew. Mathematical modeling of two-phase flow. Ann. Rev. Fluid Mech. 15, 261 (1983)
D.A Drew and Passman. Theory of Multicomponent Fluids, Springer-Verlag, New York (1999).
D.A. Drew and L.A. Segel. Averaged equations for two-phase flows. Stud. Appl. Math. 50, 205
(1971)
D.A. Drew and R.T. Lahey. Analytical modeling of multiphase flow, in Particulate Two-Phase
Flow, ed. M.C. Roco, p. 510, Butterworth-Heinemann (1993)
R. Eichhorn and S. Small. Experiments on the lift and drag of spheres suspended in a Poiseuille
flow, J. Fluid Mech. 20, 513-527 (1964).
J. Feng, H.H. Hu, and D.D. Joseph. Direct simulation of initial values problems for the motion of
solid bodies in a Newtonian fluid. Part 2: Couette and Poiseuille flows. J. Fluid Mech. 277,
271-301 (1994).
J. Feng, P.Y. Huang, and D.D. Joseph. Dynamic simulation of the motion of capsules in
pipelines. J. Fluid Mech. 286, 201-207 (1995).
J. Feng, P.Y. Huang, and D.D. Joseph. Dynamic simulation of the sedimentation of solid
particles in an Oldroyd B fluid, J. Non-Newt. Fluid Mech. 63, 63–88 (1996).
J. Feng, D.D. Joseph, and P.Y. Huang. The motion and interaction of solid particles in
viscoelastic liquids, Rheol. Fl. Mech. Nonlin. Mat. 217, 123–133 (1996).
F. Feuillebois. Some theoretical results for the motion of solid shperical particles in a viscous
fluid, in Multiphase Science and Technology (ed. G.F. Hewitt et al.), 4, 583 Hemisphere,
(1989).
A. Fortes, D.D. Joseph, and T.S. Lundgren. Nonlinear mechanics of fluidization of beds of
spherical particles, J. Fluid Mech. 177, 467–483 (1987).
P.V. Foscolo and L.G. Gibilaro. A fully predictive criterion for transition between particulate and
aggregate fluidization, Chem. Engng. Sci. 39, 1667-1675 (1984).
P.V. Foscolo and L.G. Gibilaro. Fluid dynamic stability of fluidized suspensions. The particle
bed model. Chem. Engng. Sci. 42, 1489-1500 (1987).
P.V. Foscolo, L.G. Gibilaro and S.P. Waldram. A Unified model for particulate expansion of
fluidized bed and flow in porous media, Chem Eng. Sci. 38, pp 125-1260 (1983).
A.W. Francis. Wall effect in falling ball method for viscosity, Physics 4, 403-406 (1933).
R.C. Givler. An interpretation of solid phase pressure in slow fluid particle flows, Int. J
Multiphase Flow, 13, 717-722 (1977).
R. Glowinski. Numerical Methods for Nonlinear Variational Problems, Springer-Verlag, New
York (1984).
R. Glowinski, T. Hesla, D.D. Joseph, T.-W. Pan, J. Periaux. Distributed Lagrange multiplier
methods for particulate flows, in Computational Science for the 21st Century, M.O. Bristeau,
G. Etgen, W. Fitzgibbon, J.L. Lions, J. Periaux, M.F. Wheeler eds., J. Wiley, Chichester, U.K.,
270-279 (1997).
R. Glowinski, T.-W. Pan, T. I. Hesla, and D.D. Joseph. A distributed Lagrange
multiplier/fictitious domain method for particulate flows, Int. J. Multiphase Flow 25, 755-794
(1999).
R. Glowinski, T.-W. Pan, T.I. Hesla, D.D. Joseph, and J. Périaux. A distributed Lagrange
multiplier/fictitious domain method for flows around moving rigid bodies: Application to
particulate flow, Int. J. Numer. Methods Fluids, 30, 1943-1066 (1999).

R. Glowinski, T.-W. Pan, T.I. Hesla, D.D. Joseph, and J. Périaux. A fictitious domain method
with distributed Lagrange multipliers for numerical simulation of particulate flows, in Domain
Decomposition Methods, (J. Mandel, C. Farhat, X.C. Cai eds.), American Mathematical
Society, Providence, R.I., 10, 121-137, (1998).
R. Glowinski, T.-W. Pan, T. I. Hesla, D. D. Joseph, J. Periaux. A distributed Lagrange
multiplier/fictitious domain method for the simulation of flows around moving rigid bodies:
Application to particulate flow, Comp. Meth. Appl. Mech. Eng., 184, 241-268 (2000).
R. Glowinski, T.W. Pan, T.I. Hesla, D.D. Joseph, and J. Periaux. A fictitious domain method
approach to the direct numerical simulation of incompressible viscous flow past moving rigid
bodies: Application to particulate flow, Journal of Computational Physics, 169, 363-426
(2001).
A.J. Goldman, R.G. Cox and H. Brenner. Slow viscous motion sphere parallel to a plane wall—
Couette Flow, Chem. Eng. Sci. 22, (1967).
R. A. Gore and C. T. Crowe. Effect of particle size on modulating turbulent intensity, Int. J.
Multiphase Flow 16, 279 (1989).
P. Hansbo. The characteristic streamline diffusion method for the time-dependent incompressible
Navier-Stokes equations, Comp. Meth. Appl. Mech. Eng. 99, 171–186 (1992).
T. I. Hesla. The dynamical simulation of two-dimensional fluid/particle systems, unpublished
notes (1991).
E.J. Hinch. An averaged equation approach to particle interactions in a fluid suspension, J. Fluid
Mech. 83, 695 (1977).
T. I. Hesla. Collisions in viscous fluids, in preparation (1998).
B.P. Ho and L.G. Leal. Inertial migration of rigid spheres in two-dimensional unidirectional
flows, J. Fluid Mech. 65, 365 (1974).
K. Hofler, M. Muller, S. Schwarzer and B. Wachmann. Interacting particle-liquid systems, in
High Performance Computing in Science and Engineering '98, E. Krause and W. Jager eds.,
54-64, Springer-Verlag, Berlin (1999).
S. Hou, Q. Zou, S. Chen, G. Doolen and A. Cogley. Simulation of cavity flow by the Lattice
Boltzmann method, J. Comp. Physics,. 118, 329-347, (1995)
H. H. Hu. Direct simulation of flows of solid-liquid mixtures, Int. J. Multiphase Flow 22, 335–
352 (1996)a.
H. H. Hu. Numerical simulation of channel Poiseuille flow of solid-liquid mixtures, in
Proceedings of the Fluids Engineering Division Summer Meeting, American Society of
Mechanical Engineers, 236, 97–103 (1996)b.
H.H. Hu. Direct numerical simulations of fluid solid systems. See
http://www.lrsn.upenn.edu/~howard/homepage.html (to appear 2001).
H.H. Hu and D.D. Joseph. Lift on a sphere near a plane wall in a second-order fluid. J. Non-
Newtonian Fluid Mech. 88, 173-184 (1999).

H.H. Hu, N.A. Patankar and M.Y. Zhu, Direct numerical simulation of fluid-solid systems using
the arbitrary Lagrangian-Eulerian technique, J. of Computational Physics, 169 (427-462).
(2001).
H.H. Hu. D.D. Joseph, and M. J. Crochet. Direct simulation of fluid particle motions, Theor.
Comp. Fluid Dyn. 3, 285 (1992)a.
H.H. Hu. D.D. Joseph, and A. Fortes. Experiments and direct simulations of fluid particle
motion, Int. Vid. J. Eng. Res. 2, 17 (1992)b.
H.H. Hu, N.A. Patankar and M.-Y. Zhu. Direct numerical simulations of fluid-solid systems
using the Arbitrary-Lagrangian-Eulerian technique, J. of Computational Physics, 169, 427-
462. (2000).
P.Y. Huang, J. Feng, H. H. Hu, and D.D. Joseph. Direct simulation of the motion of solid
particles in Couette and Poiseuille flows of Viscoelastic Fluids, J. Fluid Mech. 343, 73–94
(1997).
P.Y. Huang, J. Feng, and D.D. Joseph. The turning couples on an elliptic particle settling in a
vertical channel, J. Fluid Mech. 271, 1–16 (1994).
P.Y. Huang, H. H. Hu, and D.D. Joseph. Direct simulation of the sedimentation of elliptic
particles in Oldroyd B fluids, J. Fluid Mech. 362, 297–326 (1998).
P.Y. Huang and D.D. Joseph. Effects of shear thinning on migration of neutrally buoyant
particles in pressure driven flow of Newtonian and viscoelastic fluid, J. Non-Newtonian Fluid
Mech., 90, 159-185 (2000).
N.A. Patankar, P.Y. Huang, T. Ko and D. D. Joseph. Lift-off of a single particle in Newtonian
and viscoelastic fluids by direct numerical simulation, J. Fluid Mech., 438, 67-100, 2001.
A. Huerta and W. K. Liu. Viscous flow with large free surface motion, Comp. Meth. Appl. Mech.
Eng. 69, 227–324 (1988).
G. Iooss and D.D. Joseph. Elementary Stability and Bifurcation Theory. Springer, 2nd edition
(1990).
R. Jackson. Locally averaged equations of motion for a mixture of identical spherical particles
and a Newtonian fluid. Chem. Eng. Sci. 52, 2457 (1997)
A. Johnson and T. E. Tezduyar. Simulation of multiple spheres falling in a liquid-filled tube,
Comp. Meth. Appl. Mech. Eng. 134, 351–373 (1996).
A. Johnson and T.E. Tezduyar. Advanced mesh generation and update methods for 3D flow
simulations, Computational Mech. 23, 130-143 (1999).
D.D. Joseph. Fluid Dynamics of Viscoelastic Liquids, Springer-Verlag, New York (1990)a.
D.D. Joseph. Generalization of the Foscolo-Gibilaro analysis of dynamic waves, Chem. Eng. Sci.
45(2), 411-414 (1990)b.
D.D. Joseph. Finite size effects in fluidized suspension experiments, in Particulate Two-Phase
Flow, (M. C. Roco, ed.), pp. 300–324. Butterworth-Heinemann (1993).

D.D. Joseph. Flow induced microstructure in Newtonian and viscoelastic fluids, in Proceedings
of the Fifth World Congress of Chemical Engineering, Particle Technology Track 6, 3–16,
American Institute of Chemical Engineers, San Diego Keynote presentation (Paper no. 95a,
Second Particle Technology Forum). San Diego, California (1996).
D.D. Joseph. Flow induced microstructure in Newtonian and viscoelastic fluids. In Proc. 5th
World Congress of Chem. Engng, Particle Technology Track, San Diego, July 14-18. AIChE,
6, 3-16 (1996).
D.D. Joseph. Power law correlations for lift from direct numerical simulation of solid-liquid
flow, Int. J. Multiphase Flow, accepted (2001).
D.D. Joseph, J. Belanger and G.S. Beavers. Breakup of a liquid drop suddenly exposed to a high
speed airstream. Int. J. Multiphase Flow, 25, 1263-1303 (1999).
D.D. Joseph and J. Feng. A note on the forces that move particles in a second order fluid, J. Non-
Newt. Fluid Mech. 64, 299–302 (1996).
D.D. Joseph, A. Fortes, T.S. Lundgren, and P. Singh. Nonlinear mechanics of fluidization of
beds of spheres, cylinders and disks in water, in Advances in Multiphase Flow and Related
Problems, (G. Papanicolau, ed.), pp. 101–122. SIAM (1987).
D.D. Joseph and T. Liao. Potential flows of viscous and viscoelastic liquids, J. Fluid Mech., 265,
1-23, (1994).
D.D. Joseph and T.S. Lundgren. Ensemble averaged and mixture theory equations for
incompressible fluid-particle suspensions, Int. J. Multiphase Flow 16, 35 (1990).
D.D. Joseph, Y. J. Liu, M. Poletto, and J. Feng. Aggregation and dispersion of spheres falling in
viscoelastic liquids, J. Non-Newt. Fluid Mech. 54, 45–86 (1994).
D.D. Joseph and D. Ocando, 2002. Slip velocity and lift, 454, J. Fluid Mech., 263-286.
D.D. Joseph and J.C. Saut. Short-wave instabilities and ill-posed initial-value problems,
Theoretical Comput. Fluid Dynamics, 1, 191-227 (1990).
A. Karnis and S.G. Mason. Particle motions in sheared suspensions. XIX: Viscoelastic media,
Trans. Soc. Rheol. 10, 571–592 (1966).
T. Kataoka, T. Kitano, M. Sasahara and K. Nishijima. Viscosity of particle filled polymer melts,
Rheol. Acta, 17, 149-155 (1978).
L.R. Kern, T.K. Perkins, and R.E. Wyant. The mechanics of sand movement in fracturing,
Petroleum Transactions, AIME 216, 403–405 (1959).
S. Kim and S.J. Karrila. Microhydrodynamics: Principles and Selected Applications,
Butterworth-Heinemann, Boston (1991).
M. R. King and D. T. Leighton, Jr. Measurement of the inertial lift on a moving sphere in contact
with a plane wall in shear flow, Phys. Fluids, 9, 1248--1255 (1997).
M.G. Knepley, V. Sarin, and A.H. Sameh. Parallel simulation of particulate flows, in
Proceedings of the Fifth International Symposium on Solving Irregularly Structured Problems
in Parallel: IRREGULAR '98, Lecture Notes in Computer Science, Springer-Verlag, No. 1457,
pp. 226-237, Aug. 1998.

G.P. Krishnan and D. T. Leighton. Inertial lift on a moving sphere in contact with a plane in a
shear flow, Phys Fluids 7(11), 2538–2545 (1995).
A. Kuethe and C.Y. Chow. Foundations of Aerodynamics, 5th Ed. John Wiley & Sons (1998).
A.J.C. Ladd. Numerical simulations of particulate suspensions via a discretized Boltzmann
equation, Part 1, Theoretical foundation, J. Fluid Mech. 271, 285 (1994a).
A.J.C. Ladd. Numerical simulations of particulate suspensions via a discretized Boltzmann
equation, Part 2, Numerical results, J. Fluid Mech. 271, 311 (1994b).
A.J.C. Ladd. Sedimentation of homogeneous suspensions of non-Brownian spheres, Phys.
Fluids. 9, 491 (1997).
L.G. Leal. Particle motions in a viscous fluid, Ann. Rev. Fluid Mech. 12, 435 (1980).
D.T. Leighton and A. Acrivos. The lift on a small sphere touching a plane in the presence of a
simple shear flow, Z. Angew. Math. Phys 36, 174–178 (1985).
E.W. Lewis, and E.W. Bowerman. Fluidization of solid particles in liquids, Chem. Engng.
Progr., 48, 603 (1952).
W.K. Lewis, E.R. Gilliand and W.C. Bauer. Indust. Engrg. Chem., 41, 1104 (1949).
C.-J. Lin, J.H. Peery and W.R. Schowalter. Simple shear flow round a rigid sphere: inertial
effects and suspension rheology, J. Fluid Mech., 44, part 1, 1-17 (1970).
Y.J. Liu and D.D. Joseph. Sedimentation of particles in polymer solutions. J. Fluid Mech. 255,
565-595 (1993).
G.I. Marchuk. Splitting and alternating direction methods, in Handbook of Numerical Analysis,
Vol. I, P.G. Ciarlet and J.L. Lions eds., North-Holland, Amsterdam, 197-462 (1990).
B.A. Maury and R. Glowinski. Fluid-particle flow: a symmetric formulation, C. R. Acad. Sci.
Paris T. S´erie I, 324, 1079–1084 (1997).
M.R. Maxey, and B.K. Patel. Force-coupled simulations of particle suspensions at zero and finite
Reynolds numbers, ASME FEDSM, 97-3188 (1997).
M.R. Maxey, B.K. Patel, E.J. Chang, L.P. Wang. Simulations of dispersed turbulent multiphase
flow, Fluid Dynamics Res., 20, 143-156 (1997).
M.R. Maxey and J.J. Riley. Equation of motion for a small rigid sphere in a nonuniform flow,
Phys. Fluids A 26, 883 (1983).
J.B. McLaughlin. Inertial migration of a small sphere in linear shear flows, J. Fluid Mech., 224,
261-274 (1991).
J.B. McLaughlin. Numerical computation of particle-turbulent interaction, Int. J. Multiphase
Flow, 20, 211-132 (1994).
A.B. Metzner. Rheology of suspensions in polymeric liquids, J. Rheol. 29, 739-776, (1985).
A.M. Mollinger and F.T.M. Nieuwstadt. Measurement of the lift force on a particle fixed to the
wall in the viscous sublayer of a fully developed boundary layer, J. Fluid Mech. 316, 285-306
(1996).

R.I. Nigmatulin. Spatial averaging in the mechanics of heterogeneous and dispersed systems, Int.
J. Multiphase Flow 5, 353 (1979).
K. G. Nolte. Fluid flow considerations in hydraulic fracturing, SPE 18537 (1988)a. See also
Errata submitted July 1992 for SPE 18537, available from SPE Book Order Department.
K. G. Nolte. Principles for fracture design based on pressure analysis, in SPEPE, 22–30 (Feb
1998)b.
T. Nomura and T. J. R. Hughes. An arbitrary Lagrangian-Eulerian finite element method for
interaction of fluid and a rigid body, Comp. Meth. Appl. Mech. Eng. 95, 115–138 (1992).
J. Nunziato, S. Passman, R.C. Givler, D. MacTigire and J Brady. Continuum theories for
suspensions, Advancements in Aerodynamics, Fluid Mechanics and Hydraulics (Proc ASCE
S.N.) 465-472 (1986).
M. Oettli and A. Sameh. A Partitioning Scheme for the Parallel Solution of Banded Linear
Systems, Technical Report, Computer Science Dept., Univ. of Minnesota, Minneapolis (1997).
E. Ory, H.N. Oguz and A. Prosperetti. Physalis: a new O(N) method for the numerical simulation
of particle flows, Proceedings of ASME FEDSM'00, ASME 2000 Fluids Engineering
Divisions Summer Meeting, June 11-15, Boston, MA (2000).
T.-W. Pan, D.D. Joseph and R. Glowinski. Modeling Rayleigh-Taylor instability of a
sedimenting suspension of several thousand circular particles in direct numerical simulation. J.
Fluid Mech. 434, 23-37 (2001).
T.-W. Pan, V. Sarin, R. Glowinski, A. H. Sameh, and J. Périaux. A fictitious domain method
with distributed Lagrange multipliers for the numerical simulation of particulate flow and its
parallel implementation, in Proceedings of the 10th Parallel CFD Conference (1998), and in
Parallel Computational Fluid Dynamics, Development and Applications of Parallel
Technology, C.A. Lin, A. Ecer, N. Satofuka, P. Fox, and J. Periaux eds. North-Holland,
Amsterdam, 467-474 (1999).
T.W. Pan, D.D. Joseph, R. Bai, R. Glowinski, V. Sarin. Fluidization of 1204 spheres: simulation
and experiment. J. Fluid Mech., 451, 169-191 (2002)
T.-W. Pan. Numerical simulation of the motion of a ball falling in an incompressible viscous
fluid. Compte Rendus Acad. Sci. Paris, 327, 1035-1038 (1999)b.
Y. Pan and S. Banerjee. Numerical investigation of the effects of large particles on wall-
turbulence, Phys Fluids, 9(12), 3786 – 3806 (1997).
N.A. Patankar. Numerical simulation of particulate two-phase flow. Ph.D. thesis. University of
Pennsylvania (1997).
N.A. Patankar, Y. Huang, T. Ko, and D.D. Joseph. Lift-off of a single particle in Newtonian and
viscoelastic fluids by direct numerical simulation. J. Fluid Mech., 438, 67-100 (2001).
N.A. Patankar and D.D. Joseph. Lagrangian numerical simulation of particulate flows. Int. J.
Multiphase Flow, 27(10), 1685-1706 (2001).
N.A. Patankar and D.D. Joseph. Modeling and numerical simulation of particulate flows by the
Eulerian-Lagranian approach, Int. J. Multiphase Flow, 27(10), 1659-1684 (2001).

N.A. Patankar, D.D. Joseph, J. Wang, R.D. Barree, M. Conway and M. Asadi. Power law
correlations for sediment transport in pressure driven channel flows, Int. J. Multiphase Flow,
to appear (2002).
N.A. Patankar, T. Ko, H.G. Choi and D.D. Joseph. A correlation for the lift-off of many particles
in plane Poiseuille flows of Newtonian fluids, J. Fluid Mech., 445, 55-76 (2001).
N.A. Patankar, P. Singh, D.D. Joseph, R. Glowinski and T.-W. Pan. A new formulation of the
distributed Lagrange multiplier/fictitious domain method for particulate flows, Int. J.
Multiphase Flow, 26(9), 1509-1524 (2000).
C. S. Peskin. Numerical analysis of blood flow in the heart, J. Comp. Phys. 25, 220–252 (1977).
C. S. Peskin. Lectures on mathematical aspects of physiology, in Lectures in Applied
Mathematics. Volume 19: Mathematical Aspects of Physiology, (Frank C. Hoppensteadt, ed.),
pp. 69–107. American Mathematical Society, Providence, Rhode Island (1981).
C.S. Peskin and D.M. McQueen. Modeling prosthetic heart valves for numerical analysis of
blood flow in the heart, J. Comp. Phys. 37, 113–132 (1980).
M. Poletto and D.D. Joseph. Effective density and viscosity of a suspension, J. Rheol., 39(2),
323-343 (1995).
I. Proudman and J.R.A. Pearson. Expansions at small Reynolds numbers for the flow past a
sphere and a circular cylinder, J. Fluid Mech. 2, 237-262 (1957).
J. F. Richardson and W. N. Zaki. Sedimentation and fluidization: Part I, Trans. Instn. Chem.
Engrs. 32, 35–53 (1954).
P.N. Rowe. A convenient empirical equation for estimation of the Richardson-Zaki exponent,
Chem. Engng. Sci. 42, 2795-2796 (1987).
S. I. Rubinow and J.B. Keller. The transverse force on a spinning sphere moving in a viscous
fluid. J. Fluid Mech. 11, 447-459 (1961).
D. Qi. Lattice-Boltzmann simulations of particles in non-zero-Reynolds-number flows, J. of
Fluid Mech. 385, 41 (25 April 1999)
Y. Qian, D. d’Humieres and P. Lallemand. Lattice BGK models for Navier-Stokes equation,
Europhys. Lett. 17, 479 (1992).
M. J. Riddle, C. Narvaez, and R. B. Bird. Interactions between two spheres falling along their
line of centers in a viscoelastic liquid, J. Non-Newt. Fluid Mech. 2, 23–35 (1977).
M.C. Roco. One-equation turbulence modeling of incompressible mixtures. Encyclopedia of
Fluid Mechanics, Vol 10, chapter 1, 1-68, Gulf. Pub. Co, (1990).
M.C Roco and N. Balakrishnam. Multi-dimensional flow analysis of solid-liquid mixtures. J.
Rheology, 29 (4), 431-456 (1985).
S.I. Rubinow and J.B. Keller. The transverse force on a spinning sphere moving in a viscous
fluid. J. Fluid Mech., 11, 447 (1961).
Y. Saad and M. Sosonkina. Distributed Schur complement techniques for general sparse linear
systems, UMSI 97/159, University of Minnesota Supercomputer Institute (1997).

P.G. Saffman. The lift on a small sphere in a slow shear flow, J. Fluid Mech. 22, 385 (1965); and
Corrigendum J. Fluid Mech. 31, 624 (1968).
P.G. Saffman. On the motion of small spheroidal particles in a viscous liquid, J. Fluid Mech. 1,
pp. 540-543 (1956).
A.H. Sameh and V. Sarin. Hybrid parallel linear system solvers, in Proceedings of the Fourth
Japan-US Symposium on Finite Element Methods in Large-Scale Computational Fluid
Dynamics (1998).
P. Saramito. A new scheme algorithm and incompressible FEM for viscoelastic fluid flows,
Math. Modeling and Num. Anal. 28, 1–35 (1994).
V. Sarin. Spectral Analysis of Balanced CG Methods, Technical Report, Computer Science
Dept., Univ. of Minnesota, Minneapolis (1993).
V. Sarin and A.H. Sameh. An efficient iterative method for the generalized Stokes problem,
SIAM J. Sci. Stat. Comp. 19(1), 206–226 (1998)a.
V. Sarin and A. H. Sameh. Multilevel algorithms for elliptic PDEs, in Proceedings of the Fifth
Copper Mountain Conference on Iterative Methods (1998)b.
J.A Schonberg and E.J. Hinch. Inertial migration of a sphere in Poiseuille flow, J. Fluid Mech.
203, 517 (1989).
G. Segré and A. Silberberg. Radial Poiseulle flow of suspensions, Nature, 189, 209 (1961).
G. Segré and A. Silberberg. Behaviour of macroscopic rigid spheres in Poiseuille flow, Part I, J.
Fluid Mech., 14, 115 (1962).
A. Shields. Anwendung der aenlichkeitsmechanik und der turbulenzforschung auf die
geschiebebewegung. Mitteilungen der Preussischen Versuchsanstalt fur Wasserbau und
Schiffbau, Berlin, Germany, translated to English by W.P. Ott and J.C. Uchelen, California
Institute of Technology, California (1936).
D. Silvester and A. Wathen. Fast iterative solution of stabilized Stokes systems. Part II: Using
general block preconditioners, SIAM J. Num. Anal. 31(5), 1352–1367 (1994).
P. Singh, P. H. Caussignac, A. Fortes, D.D. Joseph, and T. S. Lundgren. Stability of periodic
arrays of cylinders across the stream by direct simulation, J. Fluid Mech. 205, 553–571 (1989).
P. Singh and D.D. Joseph Sedimentation of a sphere near a wall in an Oldroyd-B fluid. J. Non-
Newtonian Fluid Mech., 94(2-3), 179-203 (2000).
P. Singh, D.D. Joseph, T. Hesla, R. Glowinski and T.-W. Pan. A distributed Lagrange
multiplier/fictitious domain model for viscoelastic particulate flows, J. Non-Newtonian Fluid
Mech., 91, 165-188 (2000).
P. Singh and L.G. Leal. Finite-element simulation of the start-up problem for a viscoelastic fluid
in an eccentric cylinder geometry using a third order upwind scheme, Theor. Comp. Fluid Dyn.
5, 107–137 (1993).
P. Singh and L.G. Leal. Computational studies of the FENE dumbbell model in a co-rotating
two-roll mill, J. Rheology 38(3), 485–517 (1994).

D.M. Snider, P.J. O’Rourke, and M.J. Andrews. Sediment flow in inclined vessels calculated
using a multiphase particle-in-cell model for dense particle flow, Int. J. Multiphase Flow 24,
1359-1382 (1998).
A. Tehrani, P.S. Hammond, and A.T. Unwin. Experimental study of particle migration in
suspensions undergoing Poiseuille flow, presentation given at the Society of Rheology 66th
Annual Meeting (1994).
D.G. Thomas. Transport characteristics of suspension: VIII. A note on the viscosity of
Newtonian suspension of uniform spherical particles. J. Colloid Sci. 20, 267 (1965).
S. Turek. A comparative study of time-stepping techniques for the incompressible Navier-Stokes
equations: From fully implicit non-linear schemes to semi-implicit projection methods, Int. J.
Num. Meth. Fluids 22, 987–1011 (1996).
A.T. Unwin and P.S. Hammond. Computer simulations of proppant transport in a hydraulic
fracture, SPE 29649, Society of Petroleum Engineers Presented at the Society of Petroleum
Engineers Western Region Meeting (1995).
P. Vasseur and R.G. Cox. The lateral migration of a spherical particle in two-dimensional shear
flows, J. Fluid Mech. 78, 385 (1976).
G.B. Wallis. One Dimensional Two-phase Flow, McGraw-Hill, New York (1969).
G.B. Wallis. The averaged Bernoulli equation and macroscopic equations of motion for the
potential flow of a two-phase dispersion, Int. J. Multiphase Flow 17, 683 (1991).
J. Wang, D.D. Joseph and M. Conway. Carpet traction correlations for proppant transport, in
preparation (2002).
S. Whitaker. Advances in the theory of fluid motion in porous media, Ind. Eng. Chem. 61, 14
(1969).
R.H. Wilhelm, and M. Kwauk. Fluidization of solid particles, Chem. Engng. Progr. 44, 201
(1948).
J. Ye, and M.C. Roco. Particle rotation in a Couette flow, Phys. Fluids, 4(2), 220-224 (1992).
D.Z. Zhang and A. Prosperetti. Averaged equations for inviscid disperse two-phase flow, J. Fluid
Mech. 267, 185-219 (1994).
D.Z. Zhang and A. Prosperetti. Momentum and energy equations for disperse two-phase flows
and their closure for dilute suspensions, Int. J. Multiphase Flow 23, 425 (1997).
M.-Y. Zhu. Direct numerical simulation of the solid-liquid flows of Newtonian and viscoelastic
fluids. Ph.D. thesis. University of Pennsylvania (2000).

List of symbols
English Symbols
1 Identity tensor
A Area of cross section
A Configuration tensor (= τE + (ηE / λ1)1
1)
a Radius of the particles
b Body force
D Diameter of cylindrical tube
D Drag force
d Diameter of the particles
d Average diameter of the particles
di,j Distance between the center of the i-th and j-th particles
ex Unit vector in the x-direction
ey Unit vector in the y-direction
F Force
Fi′ Short-range repulsive force exerted on the i-th particle (Eq. VII.9)
Fi ,pj Particle-particle repulsive force (Eq. VII.10)
Fi ,pj Particle-wall repulsive force (Eq. VII.11)
g Acceleration due to gravity
H Indicator function
H Height of the channel
He Equilibrium height of the fluidized bed
hv Velocity mesh size
hp Pressure mesh size
I Moment of inertia tensor
L Channel length
L Lift force
L′ Lift force per unit length
Le Elastic component of lift force
Lp Pressure component lift force
Ls Shear component lift force
M Particle mass
N Number of particles
n Unit outward normal vector on the particle surface
nl Growth rate of the wave amplitude
p Pressure
p Average pressure gradient
pa Atmospheric pressure
r Position vector of the point w.r.t. the center of mass of the particle
T Torque
T Dimensionless stress
T* Stress tensor (Chap. 6)
Tf Stress for the fluid phase (Chap. 6)
t Time
t Traction on the particle surface
U Translational velocity of the particles
Um Mean velocity of the fluid (Chap. 9)
Us Slip velocity (= Uf – Up)
u Velocity of the fluid
Vc Composite velocity (= εVf + φVs) (Chap. 6)
Vf Average velocity of the fluid phase (Chap. 6)
Vm Mass averaged velocity (Chap. 6)
Vp Volume per unit length of the particles (Chaps. 9 and 11)
Vs Average velocity of the solid phase (Chap. 6)
W Width of domain
X Coordinate of the center of mass of the particle
Greek Symbols
ε Volume fraction of the fluid phase

εp and εw Stiffness parameter for the collision scheme
Φe Elastic lift fraction
Φp Pressure lift fraction
Φs Shear lift fraction
φ Volume fraction of the solid phase
Γ Circulation
γ Surface tension
γ Shear rate
γ w Shear rate at the wall
η Viscosity of the fluid
η2 Effective viscosity of the composite (Eq. VIII.8)
ηE Elastic viscosity
ηm Effective viscosity of the composite (Chap. 14)
λ Lagrange multiplier field
λ Wavelength (Chap. 8)
λ1 Relaxation time
λ2 Retardation time
ν Kinematic viscosity
Π Pressure in a periodic domain
π 3.141592
θ Angular position of the particle
ρf Density of the fluid
ρ2 Effective density of the composite (Chap. 8)
σ Stress tensor
τ Shear stress
τE Elastic stress
υ composite velocity
υm Migration velocity (Chap. 10)
Ω Computational domain, angular velocity
Ωf Fluid domain
Ωs Particle domain
Ωf Angular velocity of the fluid
Ωp Angular velocity of the particle
Ωs Slip angular velocity (=Ωf – Ωp)
ω Angular velocity of the particle
Dimensionless Groups
CD Drag coefficient
De Deborah number
E Elasticity number De/R
G Gravity parameter
R Shear Reynolds number
RG Gravity Reynolds number
Rs Slip Reynolds number
Subscript
2D Two dimensional
3D Three dimensional
0 Initial Value t = 0 (Chap. 14)
c Composite
E Elastic
f Fluid
i and j i-th and j-th particle
p particle
s solid
sed Sedimentation
T Total
∞ t→∞
Index
A
Advection-diffusion problem ............................................................................................................... 9
ALE particle mover.............................................................................................. 6-7, 10, 11, 20-22, 24
Angular velocity .........................................................13, 24, 53, 87, 92-93, 95, 97-98, 100, 102-104,
108-109, 114-115, 118-120, 125, 126, 133-138, 140, 164-165, 171-172, 178, 186, 195, 202
Animations ............................................................................................................................. 1, 21, 164
Anisotropic............................................................................................................................. 17, 19, 74
Approximate methods .......................................................................................................................... 2
Arbitrary Lagrangian-Eulerian (ALE).................................................................................................. 6
B
Batchelor’s theory .............................................................................................................................. 40
Bed load transport case ..................................................................................... 213, 217, 222-225, 229
Bifurcation diagram ......................................................................................................................... 141
Bifurcation, double turning point .................................................................................................... 141
Bifurcation, Hopf .............................................................................................................................. 86
Bifurcation, turning point ........................................................................................................ 118, 136
Bi-power law..............................................................................................207, 216-218, 221-223, 229
Bird-Carreau model.......................................................................................................................... 150
Birds in flight ................................................................................................................................ 18-19
Buoyant weight ......................................................... 24, 33-34, 36, 38, 57, 59, 64, 85, 87-89, 95, 101,
105, 112, 115, 117-118, 121, 133-134, 140, 143, 149, 163, 168-169, 200, 202, 205, 208-209
C
Chain, chaining ................................................................................................................ 17, 18, 19, 20
Circular particle..................................................... 21-22, 43, 65, 74, 86, 89, 94, 96-97, 105, 107, 119,
120-122, 125, 127, 129, 130, 134-136, 140, 145-147, 149, 151, 153, 156, 158-159, 163, 165-166, 172, 178,
186-187, 194, 199, 202, 206, 210
Collision strategies ....................................................................................................................... 11, 20
Collision strategy ................................................................................................................... 11, 20, 54
Couette flow....................................................................................................................... 98, 103, 105
Courant number................................................................................................................................ 206
Critical shear Reynolds number .................................................... 90, 92, 136, 149, 152, 159, 161-162
Cross-stream migration .............................................................................................................. 21, 150
Cubic crystal array .................................................................................................... 163-164, 195, 202
D
Dallavalle formula.............................................................................................................................. 37
Darcy's law................................................................................................................................. 32, 212
Data structure .................................................................................................. 3, 53, 93, 95, 98-99, 217
Deborah number.................................................................................................................. 10, 150-152
Dirac's delta function.......................................................................................................................... 28
Direct numerical simulation (DNS) ...................... 1-3, 17, 19, 21, 23-26, 29-30, 37, 40-41, 43, 55-56,
58, 60, 63-65, 85, 92-93, 95, 98, 100, 102, 105, 107, 111-112, 118-119, 124-126, 134, 149, 163, 186, 199,
206-208, 210, 212-213
Dirichlet problem ......................................................................................................................... 14, 16
Distributed Lagrange multiplier (DLM)...........................................7-16, 21, 29, 41-43, 46, 52, 63, 65
DLM particle mover....................................................................................................................... 7, 10
Interog-index.doc
Double turning point bifurcation...................................................................................................... 141

Draft, drafting................................................................................................................................ 17-19
Drag......................................................................... 1, 17, 23, 30, 31, 33-38, 47, 57, 59, 61-62, 64, 66,
85, 97, 99-100, 102-105, 107-108, 112, 119, 140, 149, 163, 203-204, 207-209, 211
Drag force .................................................................................30, 33, 38, 97, 104, 108, 149, 207, 208
E
Eddy viscosity .................................................................................................................................. 206
Effective fluid............................................................................................................................... 21, 68
Effective viscosity .................................................21, 34, 35, 56 ,68 , 71, 73-76, 82-84, 163, 204-206
Elastic stress ....................................................................................................................................... 10
Elastic viscosity.................................................................................................................................. 10
Elasticity....................................................................................................19, 20-21, 149-152, 159-160
Elasticity number ....................................................................................................................... 20, 159
Engineering correlation ........................................................................................ 1, 199, 202, 207, 229
Ensemble ........................................................................................................25-27, 29-30, 53, 74, 205
Equilibrium height .................................................24, 89-93, 95, 97-98, 115, 131, 133-134, 136, 138,
141-142, 148-150, 152, 159, 199-200
Equilibrium position...................................................................110, 118, 124, 127-129, 132, 134, 152
Equilibrium radius.....................................................................................................110, 118, 134, 150
Erosion case ...............................................................................................213-214, 216, 222-224, 229
Eulerian-Lagrangian......................................................................................................................... 2, 6
F
Fictitious domain.......................................................................................................... 7, 15, 41, 42, 65
Film rupture........................................................................................................................................ 12
Fluid elasticity.......................................................................................................................... 151, 159
Fluid fraction............................................................................................26, 32, 61, 199-200, 209-210
Fluidization ....................................................1, 20-21, 23, 30-33, 35, 37, 39-41, 44-46, 48-49, 53-54,
57-61, 63, 85, 89, 95, 100, 163-169, 171-174, 177, 179-197, 199, 202, 206-212, 229, 233-234
Fluidized beds ............................................................................................... 1, 18, 30, 38, 40, 163-164
Fluidized suspension ................................................. 30, 33, 45, 54, 163, 165, 172, 178, 186, 194-195
Fluidizing velocity.....................................................................................21, 30, 45, 48, 56, 58, 61, 63
Foscolo and Gibilaro model ............................................................................................................... 38
Foscolo-Gibilaro theory ..................................................................................................................... 40
Fractured reservoirs............................................................................................................................ 23
Froude number, generalized ............................................................................................................... 88
Free rotation ............................................................................................................................... 95, 136
G
Generalized Froude number ............................................................................................................... 88
Generalized Newtonian fluid ............................................................................................................. 22
Gravity Reynolds number .................................................................................................. 88, 150, 210
H
Hadamard instability .......................................................................................................................... 10
Hexagonal ...........................................................................65-66, 68, 73-74, 76, 80, 83, 194-195, 205
Hindered settling ..................................................................................................................... 30, 34-35
Hits .................................................................................................................................................. 53
Homogeneous shear flow apparatus (HFA) ....................................................................................... 98
Hopf bifurcation ................................................................................................................................. 86
Interog-index.doc
Hysteresis ..........................................................................................................................118, 136, 141
I
Immerse boundary method................................................................................................................. 42
Inelastic collisions.............................................................................................................................. 11
Inertial lift............................................................................................................... 97-98, 110, 114-115
Initial lattice ........................................................................................................... 65-69, 72-76, 78-84
Integrated method................................................................................................................................. 6
Inter-particle stress ............................................................................................................................... 2
J
Jump .......................................................................................................................... 136-137, 139, 143
K
Kiss, kissing .................................................................................................................................. 17-19
L
Lagrange multiplier field......................................................................................................... 13-15, 65
Lagrange multipliers ........................................................................................................ 1, 6, 9, 42, 65
Lagrangian............................................................................................................................ 2, 6, 25, 41
Lagrangian numerical simulation (LNS)............................................................................................ 25
Lattice Boltzmann method (LBM) ....................................................................................................... 3
Leapfrogging vortex rings.................................................................................................................. 20
Levitation (levitating).................................................................................85-86, 97, (100), (148), 207
Lift................................................................................... 85, 89, 91, 100, 104, 108-109, 128, 136, 144
Lift balance......................................................................................................................... 85, 133, 143
Lift force...................................................................... 24, 30, 85, 97-98, 101, 104, 108, 110, 115-116,
120, 130, 133-134, 139, 140-141, 146-147, 149, 154-155, 157-159, 167, 172, 202, 207
Lift fraction .............................................................................................................................. 143, 154
Lift-off.................................... 1, 23-24, 85-86, 89, 90-92, 94-95, 97-98, 112, 136, 141, 151, 207, 211
Logistic dose response curve....................................................................... 62, 223, 225, 229, 231-234
Lorenz, reciprocal theorem of ......................................................................................................... 113
Lubrication ........................................................................................... 11-12, 19, 22-23, 109, 115, 149
M
Mach number, Viscoelastic ............................................................................................................... 20
Method of correlations .................................................................................................. 60, 64, 229-230
Microstructure .............................................................................................................................. 17, 19
Multiphase flows................................................................................................................................ 29
Multiphase turbulent models.............................................................................................................. 25
Multiple steady states ............................................................................................................... 136, 141
N
Navier-Stokes equation ............................................................................... 1, 3, 8, 12, 16, 42, 102-103
Neutrally buoyant...............................................22, 95, 96, 98, 104-106, 110-111, 115, 118, 120, 127,
128, 130, 134, 138, 139, 150
Newton’s law........................................................................................................................................ 1
Newtonian fluid ............................................... 4-5, 12, 17-24, 28, 40, 55, 89, 136, 139-141, 145-147,
149, 152, 154-157, 159-162, 199, 201-202, 211
Newtonian suspending fluid............................................................................................................. 140
Interog-index.doc
O
Oldroyd-B constitutive model............................................................................................................ 10
Oldroyd-B fluid..................................................................................5, 20, 22, 149-155, 158-160, 162
Oldroyd-B rate equation................................................................................................................... 149
Oldroyd-B rate law........................................................................................................................... 150
Oseen flow ........................................................................................................... 99-100, 104-105, 118
P
Particle mover .......................................................................................................6, 7, 9, 11, 17, 21, 23
Particle mover, DLM ..................................................................................................................... 7, 10
Particle Reynolds number ................................................................................................... 66, 104-105
Particle tracking methods..................................................................................................................... 2
Particle-in-cell method ......................................................................................................................... 2
Particle-laden flow ............................................................................................................................. 21
Particle-laden region ........................................................................................................................ 204
Particulate.........................................................................................................1-3, 7, 15, 17-19, 41, 76
Plane-Poiseuille flow ....................................................................................................................... 149
Plateau value ............................................................................................................................ 150, 231
Poiseuille (flow) ................................22, 24, 41, 57, 86-87, 89, 90, 94-95, 97, 102-104, 110-112, 118,
121, 125, 134, 136, 140, 145-149, 151-153, 155-156, 158, 163-164, 186, 199, 201-206, 210
Porous media...................................................................................................................................... 32
Positioning property ..................................................................................................................... 58, 64
Power law................................................................22, 41, 60-62, 64, 92, 159, 207-208, 212, 229-230
Pressure lift........................................................................................................................ 143-144, 154
Pressure wave.................................................................................... 163, 166, 173-175, 183, 194, 197
Projected particle mover....................................................................................................................... 7
R
Rayleigh-Taylor............................................................................................................................ 31, 73
Rayleigh-Taylor instability................................................................................... 31, 65, 66, 68, 74, 76
Reciprocal theorem of Lorenz.......................................................................................................... 113
Rectangular ..................................................................................... 65-66, 69-70, 73-74, 76, 80-81 ,84
Regular perturbation..........................................................................................................................111
Relaxation time ...................................................................................................................... 5, 10, 150
Resuspension................................................................................................................................. 23-24
Retardation time ........................................................................................................................... 5, 150
Return flow ...................................................................................................................................... 148
Reynolds number, critical shear ................................................... 90, 92, 136, 149, 152, 159, 161-162
Reynolds number, gravity ................................................................................................. 88, 150, 210
Reynolds number, particle .................................................................................................. 66, 104-105
Reynolds number, slip ..................................................................................................................... 106
Reynolds number, shear ............. 24, 90-91, 98, 118, 136-138, 140-145, 149, 152-155, 159-160, 164,
201-202, 210
Richardson-Zaki (RZ) correlation .........................................35-37, 60-61, 64, 163, 207-208, 231, 233
Richardson-Zaki (RZ) exponent......................................................................................... 37, 208, 236
Richardson-Zaki formula ................................................................................................................... 45
Richardson-Zaki theory.................................................................................................................... 100
Richardson-Zaki type ....................................................................................................................... 211
Rigid-body motion ................................................................................................................. 1, 6, 8, 42
Interog-index.doc
S
Security zone free............................................................................................................................... 12
Sedimentation...................................................... 20-21, 23, 30, 31, 33, 35, 41, 57, 58, 64, 85, 88, 163
Sedimenting suspension .................................................................................................. 34-35, 68, 204
Segré-Siberberg position .............................................................................................................59, 111
Segré-Silberberg effect......................................................................................................110, 118, 151
Segré-Silberberg radius ............................................................................................................ 134, 138
Separation distance..................................................................................................................... 18, 149
Shear lift........................................................................................................................................... 143
Shear Reynolds number .....................................24, 90, 91, 98, 118, 136-138, 140-145, 149, 152-155,
159-160, 164, 201-202, 210
Shear thinning ....................................................................................................................... 20-22, 150
Shear-wave speed............................................................................................................................... 20
Slip ...................................... 75, 85, 86, 92, 95-96, 98-100, 102-106, 108-109, 111, 115-123, 125-138
Slip angular velocity.................................................................................... 92, 103, 120, 133, 136-138
Slip coefficients.................................................................................................................................. 23
Slip Reynolds number ...................................................................................................................... 106
Slip velocity ............................................... 21, 24-25, 38, 54, 64, 85, 92, 95-96, 98-99, 102, 104-106,
108-109, 111, 115-121, 123, 125-135, 137, 203, 205
Slurry................................................................................. 1, 10, 17, 19, 23, 30, 74, 100, 163-164, 205
Solid-liquid flow .......................................................................1, 16, 25, 30, 37, 43, 63, 120, 210, 229
Solids fraction .................................................................................................... 26, 32, 38, 46, 65, 174
Splitting method .......................................................................................................................... 6-7, 41
Square.........................................19, 24, 39, 46, 47-49, 65-67, 72, 74, 76, 78-79, 81-84, 107, 114, 152
Stagnation point ......................................................................................................................... 19, 148
Stokes flow................................................................. 2, 24, 34, 35, 36, 41, 88, 100, 102-105, 109-115
Strong solutions............................................................................................................................ 14, 16
T
Total momentum ...................................................................................................................... 4, 15, 29
Tumbling ....................................................................................................................................... 17-18
Turbulent flow models ......................................................................................................................... 2
Turning point bifurcation ......................................................................................................... 118, 136
Two-fluid model.......................................................................................... 2, 16, 30, 37, 39, 68, 74-76
Two-phase flow.............................................................................................................................. 2, 19
Two-phase models.............................................................................................................................. 39
U
Upper convected derivative.................................................................................................................. 5
V
Viscoelastic ..............................................................................................5, 12, 17-21, 23-24, 149, 154
Viscoelastic liquid ......................................................................................................................... 17-18
Viscoelastic Mach number ................................................................................................................. 20
Viscoelastic pressure .................................................................................................................... 17, 19
Viscous potential flow......................................................................................... 65, 68, 70, 74-75, 100
Void fraction.............................................................................................. 12, 30-31, 40, 173, 177, 198
Volume fraction ............................................................................ 1, 2, 33, 53, 65, 73-75, 163, 204-205
W
Weak solution..................................................................................................................... 4, 13, 15, 29
Interog-index.doc

Interrogations of DNS of Solid-Liquid Flows Daniel D

Hochgeladen von

Dokumentinformationen

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

Interrogations of DNS of Solid-Liquid Flows Daniel D

Hochgeladen von

Copyright:

Verfügbare Formate

Interrogations of Direct Numerical Simulation of

© 2002 Daniel D. Joseph i

© 2002 Daniel D. Joseph ii

© 2002 Daniel D. Joseph iii

© 2002 Daniel D. Joseph 1 Updated 02/21/03 • Interog-1.doc

© 2002 Daniel D. Joseph 2 Updated 02/21/03 • Interog-1.doc

 Data Structure for Applications of DNS

© 2002 Daniel D. Joseph 3 Updated 02/21/03 • Interog-1.doc

II. Equations of Motion

 Weak Solution for the Total Momentum

© 2002 Daniel D. Joseph 4 Updated 02/21/03 • Interog-1.doc

III. Numerical Packages for Moving Particles in Direct Numerical

 ALE Particle Mover

© 2002 Daniel D. Joseph 6 Updated 02/21/03 • Interog-1.doc

 A Projected Particle Mover

 DLM Particle Mover

The scheme uses an operator-splitting technique for discretization in time. (Operator-splitting

DLM2 D�u� � 0, x � P(t )

where D[u] is the symmetric part of �u.

© 2002 Daniel D. Joseph 8 Updated 02/21/03 • Interog-1.doc

© 2002 Daniel D. Joseph 9 Updated 02/21/03 • Interog-1.doc

 Viscoelastic DLM Particle Mover

where D is the rate-of-strain tensor, A � J E � (D E / �1 )1 is the configuration tensor, JE is the

 Computation domains for pipe flows of slurries

© 2002 Daniel D. Joseph 10 Updated 02/21/03 • Interog-1.doc

Simulations of single particle lift to equilibrium were also performed in a computational

© 2002 Daniel D. Joseph 11 Updated 02/21/03 • Interog-1.doc

© 2002 Daniel D. Joseph 12 Updated 02/21/03 • Interog-1.doc

The physics of collisions is thought to correlate with the Stokes number

• G. G. JOSEPH, R. ZENIT, M. L. HUNT AND A. M. ROSENWINKEL. PARTICLE-WALL COLLISIONS IN A

IV. Weak and Strong Formulations of the DLM Method

 Body Force Formulation of DLM Method; Weak Solution

 Body Force Formulation of DLM: Strong Solutions

© 2002 Daniel D. Joseph 13 Updated 02/21/03 • Interog-1.doc

λ satisfies a natural boundary condition

evaluating (IV.4) in 9 f , we find

in 9 f and from (IV.5), using no-slip, we get

Using (IV.5) to evaluate (IV.4) on the solids 9 s we get

© 2002 Daniel D. Joseph 14 Updated 02/21/03 • Interog-1.doc

 Stress Formulation of DLM: Weak Solutions

and a partial differential equation equivalent to (IV.11) is

� � D�u� � 0 in 9 s and D�u�  n � 0 on �� s . (IV.12)

The combined form of the total momentum equation is formulated as follows:

Find u, p, λ, U and M satisfying

© 2002 Daniel D. Joseph 15 Updated 02/21/03 • Interog-1.doc

 Stress Formulation of DLM: Strong Solutions

In 9 s , div u=0, D[u]=0 and

The velocity and stress traction vectors are continuous across �� s � �� f

n  D�λ �s � n  � p1 � 2DD�u � f . (IV.19)

© 2002 Daniel D. Joseph 16 Updated 02/21/03 • Interog-1.doc

Noting now that Ω = Ω / Ω s ∪ Ω s , where Ω / Ω s = Ω f is the part of Ω occupied by the

is the moment of inertia tensor.

© 2002 Daniel D. Joseph B-1 Updated 01/25/05 • AddendumA B.doc

where r = x − X(t ) and X is the center of mass.

Noting next that M = ρ s ∫ dx , the terms proportional to V in (IV.21) are

and the terms proportional to ξ become

which, after applying well known vector identities, becomes

We note next that

© 2002 Daniel D. Joseph B-2 Updated 01/25/05 • AddendumA B.doc

and the condition

guarantees that the fluid moves as a rigid body.

© 2002 Daniel D. Joseph B-3 Updated 01/25/05 • AddendumA B.doc

(1) Extensions of DLM to different element bases

• H. HU, N. A. PATANKAR AND M. Y. ZHU. DIRECT NUMERICAL SIMULATIONS OF FLUID-SOLID SYSTEMS

Data Structure for Applications of DNS

Weak Solution for the Total Momentum

ALE Particle Mover

A Projected Particle Mover

DLM Particle Mover

Viscoelastic DLM Particle Mover

Computation domains for pipe flows of slurries

Body Force Formulation of DLM Method; Weak Solution

Body Force Formulation of DLM: Strong Solutions

Stress Formulation of DLM: Weak Solutions

� � D�u� � 0 in 9 s and D�u� n � 0 on �� s . (IV.12)

Stress Formulation of DLM: Strong Solutions

n D�λ �s � n � p1 � 2DD�u � f . (IV.19)

<H div T> = div <HT> – <∇H • T*> (VI.15)

<HT> = ε Tf (x, t) (VI.17)