Sie sind auf Seite 1von 10

Numerical Simulations using Gerris of Two and Three-Dimensional Patterns

in the Rayleigh-Taylor Instability of a Thin Viscous Layer:


Research Summary Report

Minh-Thi Nguyen, Princeton University


Mentor: Dr. Stephan Weiss, Max Planck Institute for Dynamics and Self-organization
Max Planck Institute for Dynamics and Self-organization
Göttingen, Germany
June 3, 2018 – August 3, 2018

I. Introduction and Motivation:


The Rayleigh-Taylor instability characterizes the instability of the interface between two
superimposed incompressible, immiscible fluids of densities 𝜌" and 𝜌# , subjected to the
influence of gravitational acceleration. The instability occurs when the heavier fluid is
suspended over the lighter fluid (𝜌" > 𝜌# , where 𝜌" is the density of the top fluid and 𝜌# is the
density of the bottom fluid), with the lighter fluid pushing against the heavier fluid. Any
perturbation or displacement to the interface will cause the equilibrium of a flat interface to
become unstable, and the study of the resultant wave behavior and pattern formation of the
interface and how they evolve over time are of particular interest (Chandrasekhar 1961). A
specific manifestation of the Rayleigh-Taylor instability is the case in which one of the fluid
layers (we consider the top layer) is very thin,
'
comparable to the capillary length 𝑙% = and
()
much smaller than the viscous diffusion length 𝑙+ =
ht
( )"/# . In this regime, the dynamics of the
(
Rayleigh-Taylor (RT) Instability is governed
primarily by viscous (in lieu of inertial) effects,
which occurs when the layers are thick (Ferminger et
al. 1989). Although the RT Instability has been
previously investigated and studied more with thick
fluid layers where the fluid viscosities are often
neglected in equations for simplifications in
calculation, the RT Instability of a thin layer has
numerous natural occurrences in geophysics and
applications in technologies with films (Brunet et al.
2003). In previous experimental investigations of
thin two-dimensional films on a solid plate, with
silicon oil as the top fluid and air as the bottom fluid, Figure 1: (Ferminger et al.) Common pattern formation
numerous two-dimensional patterns and symmetries (and and symmetries formed in the RT instability over time,
transitions between them) are observed in the dynamics which includes six-fold symmetries that become hexagonal
patterns, rolls, and line structures.
of the instability that are heavily dependent on initial
perturbations, such as “rolls” (one-dimensional
undulations parallel to the boundary of the domain), axisymmetric patterns, and hexagonal
patterns, with axial and hexagonal symmetries preferred at large times (Limat et al. 1992,
Ferminger et al. 1989).
In our study, as theoretical and quantitative discussions of experimental studies have
attempted to explain the dominant nature of the hexagonal and axial symmetries, we aim to
numerically study and corroborate this pattern formation of a dripping flat plate to further our
understanding of these phenomena by using the Gerris flow solver through a volume of fluid
(VOF) approach for the simulation of interfacial instability (Popinet 2003). Specially, we use
Gerris to simulate the RT Instability in two and three dimensions, investigating the growth of the
interface instability (wave modes) in the linear regime in 2D and the bidimensional patterns of
the interface as seen on the top of the plate in 3D, both with and without an external supply of
liquid for the top layer, as well as the variation of several other experimental parameters, such as
the thickness and surface tension of the liquid.

II. The Gerris Flow Solver:


Our simulations are executed through the Gerris software, an open source Navier-Stokes
partial differential equations solver in both two and three dimensions that computes fluid
dynamics through quad (2D) and octree (3D) grids that allow an advanced adaptive mesh
refinement (with spatial and time-varying spatial resolution), while also using the Volume-Of-
Fluid (VOF) method for effective tracking of the interface and optimization of numerical
calculations (Popinet 2003). The VOF method uses a discontinuous characteristic function (T = 1
for the top fluid, T = 0 for the bottom fluid, and 0 < T < 1 at the interface) to track the volume
fraction of each fluid in each cell. Nevertheless, one of the weaknesses of the Gerris VOF
method is the accuracy of the interfacial values for useful geometrical properties such as the
curvature of the interface, especially when the resolution of the grid is insufficient. Furthermore,
cases with high density ratios or high viscosity ratios, as well as those that contain the formation
of fast or very small structures such as small droplets in a large domain, can result in very
computationally expensive
simulations (especially since the
default viscosity is 0 for the
solver, so higher numerical values
of the viscosity can result in more
computational time for the
solver).
The Gerris parameter file
specifies a domain for the
simulation through nodes of a
GfsBox, with GfsRefine
establishing the number of
quad/octree levels the code should
use to discretize each GfsBox.
The adaptive refinement of the
mesh for increased resolution at
the interface is controlled using
the gradient of the tracer and the
vorticity. The presence of Figure 2: Excerpt of Gerris parameter file for execution of simulation in 2D with one
Gfsbox domain with periodic boundary conditions (one connection in the box)
viscosity of the liquid fluid is
added with GfsSourceViscosity. The Gerris software is based on non-dimensional units, so all
simulation parameters are based on relative units in lieu of absolute units; thus, the physical size
of the domain is set by the source term gravitational acceleration that is applied to the whole
domain (for instance, setting the acceleration of the system to be -9.81 would result in a domain
size of the box to be 1 meter by 1 meter and other physical parameters, such as density and
viscosity, to be defined in SI units). The initial interface between the two fluids is set by
GfsInitFraction as an implicit function; in our simulations, although experimental trials study a
uniform, flat film interface, we investigate different initial conditions for the initial interface for
the instability, including several runs with superimposed cosine functions to create white noise at
the interface to excite as many possible wave modes as possible. Gerris outputs the current state
of the simulation through the characterization of a variety of cell-centered variables, such as
pressure and interface displacement. Moreover, Gerris uses GfsView as a visualization tool
associated with the solver that enables multiscale visualization of simulation results in the
quad/octree, while also providing an interactive framework for the user.
Parallelization in Gerris computational runs is achieved through domain decomposition;
using the mpi library, we run large simulations in 3D using multiple processors on the MPI
Belenus cluster. In these simulations, Gerris splits the mesh into as many subdomains as the
processors and each processor performs the same instructions as the other subdomains in its own
subdomain, while communication between the processors allows the values at the boundaries to
be synchronized. Nevertheless, numerous users have indicated that the parallel performance of
Gerris still has a lot of challenges and show inconsistencies between single and parallel runs. To
optimize computation and time, although we use parallelization for three-dimensional
calculations, we improve the quantitative synchronization between subdomain boundaries by
using GfsProjectionParams to tune the Poisson solver and introduce coarser grids when going up
the multigrid hierarchy and we limit each simulation to 16 processors.

III. Two-Dimensional Runs (One-Dimensional Experiment)


In our 2D Gerris simulations of the RT
Instability, we primarily investigate the growth of
dominant wave modes (their amplitudes over time)
and the fastest growing wave mode over time in the
linear growth regime of the instability, varying the
boundary conditions, as well as the thickness of the
top fluid layer, its surface tension, and its viscosity.
To set up our simulation, we superimpose two fluids
of stratified density, with the top fluid representing a
liquid with control density 973.0 kg/𝑚0 and dynamic
viscosity 0.02 Pa·s, consistent with the parameters of
experimental trials for silicon oil as the film liquid.
Nevertheless, our control surface tension coefficient
value is 1.0 N/m (a little less than 5 times the
experimental surface tension of silicon oil). The
bottom fluid is given the same physical parameters as Figure 3: Snapshot of the growth of the
air at 15°C (𝜌# = 1.225 𝑘𝑔/𝑚0 and 𝜂# =1.81x109: ). perturbations of the interface in two dimensions
of the RT instability of two stratified fluids
The resolution of the simulation is defined at a separated by a flat interface excited by numerical
maximum of level 8 at the interface for all runs. noise
Because our physical parameters are in SI units, the domain of the size of the square is
1.0mx1.0m; the capillary length is 𝑙% = 32.2 cm. We let the bottom boundary be a constant
outflow.
We perform several runs, including the following:
1. Flat Interface (initial perturbations that disrupt equilibrium caused by numerical noise)
a. Vary the thickness of the top fluid (from 5 cm to 1 mm)
i. With no constant external inflow of the top fluid
1. Periodic Boundary Conditions for left and right sides of the box
domain
2. Non-Slip Conditions set by Dirichlet Boundary Conditions on left
and right sides of the box
ii. With a constant external flow of the top fluid with flow rate Q = 30 cm/s
1. Periodic Boundary Conditions for left and right sides of the box
domain
2. Non-Slip Conditions set by Dirichlet Boundary Conditions on left
and right sides of the box
b. Vary the surface tension coefficient (from 0.2 N/m (experimental value) to 2.0
N/m)
1. Periodic Boundary Conditions for left and right sides of the box
domain
2. Non-Slip Conditions set by Dirichlet Boundary Conditions on left
and right sides of the box
c. Vary the viscosity of the liquid (to confirm numerically simulated droplet
formation)
2. Flat Interface with White Noise (superposition of 1024 cosine functions/frequencies)
a. Vary the thickness of the top fluid (from 5 cm to 1 mm)
i. With no constant external inflow of the top fluid
1. Periodic Boundary Conditions for left and right sides of the box
domain
2. Non-Slip Conditions set by Dirichlet Boundary Conditions on left
and right sides of the box
ii. With a constant external inflow of the top fluid with flow rate Q = 30 cm/s
1. Periodic Boundary Conditions for left and right sides of the box
domain
2. Non-Slip Conditions set by Dirichlet Boundary Conditions on left
and right sides of the box
b. Vary the surface tension coefficient (from 0.2 N/m (experimental value) to 2.0
N/m)
1. Periodic Boundary Conditions for left and right sides of the box
domain
2. Non-Slip Conditions set by Dirichlet Boundary Conditions on left
and right sides of the box
c. Vary the viscosity of the liquid (to confirm numerically simulated droplet
formation)
3. Droplet formation
a. In several runs, we examine the effectiveness of the Gerris software in computing
fluid interface for the development of droplet formation in two dimensions that is
consistent with the Rayleigh-Plateau Instability, but we conclude that this
behavior can only be obtained in three dimensional runs.

For the simulations with variation in surface tension


coefficient and liquid thickness, we examine the
interfacial pattern and the development of the
instability by using Fourier analysis to study the
(exponential) growth rate of the amplitude of the
perturbation in the linear regime (of the most
dominant wave mode), as well as the rate of growth
of each frequency.
We observe large differences in the growth
rate of the most dominant wave mode of the
Figure 4: We plot the interface position using the
VOF Height method on the left and use FFT to instability for simulations without a constant
examine the dominant frequency in the interface at external inflow of liquid that varied surface tension
this particular moment in time in the simulation and film thickness of the liquid, while the addition
of the external liquid inflow (although this flow rate
is very high) results in smaller differences in the growth rate of the amplitude of the instability
for each parameter variation. Over time, as the instabilities of the simulations develop, the
frequency with the greatest amplitude
becomes smaller. Furthermore, analyzing the
change in amplitude over time for each
frequency, we also notice that the increase in
surface tension results in a stabilizing growth
rate for higher wave modes, with higher
wave modes having higher growth rates at
lower wave numbers and a consistent growth
rate at higher wave numbers. Nevertheless,
for lower surface tension coefficient
parameters, there exists a wavenumber with
the greatest growth rate over time.

Figure 5: Compare the effects of a constant supply of


liquid flow from the top of the boundary for various
values of surface tension of the liquid
Figure 6: Compare the effects of a constant supply of liquid flow from the top of the boundary for various values
of the thickness of the top fluid. The left graph shows with no inflow of liquid, while the right shows the
presence of the same trials with an inflow of the top fluid.

Fastest Growing Mode


10

7
Growth Rate

1
0 500 1000 1500 2000 2500
Wavenumber

Figure 7: Compare the effects of surface tension on the growth rate of each wavenumber in the presence of an
external supply of liquid layer. The left graph shows the growth of each mode at surface coefficient value of 0.4,
and the right shows the same graph for a surface coefficient of 1.2. The growth rate is measured as the change
in log(amplitude) over time.

In the simulations simulating droplet formation, we observe that the computational and
physical time required for high resolution droplet formation (in a large, default domain of 1.0 m
x 1.0 m with refinement level maximum of 8) exceeds significantly the time required for that of a
simulation with a significantly smaller resolution (maximum of level 5 or 6) but also smaller
physical boundary parameters (1.0 cm x 1.0 cm) by setting a larger value for the source gravity.
Nevertheless, the precision in the initial interface curvature and consequent development is
comparable for both cases.
IV. Three-Dimensional Runs (Bidimensional Patterns)

In the three-dimensional runs, we primarily analyze and look for the formation of patterns
of dripping formed at the interface, as viewed from the top of the plate over time to investigate
the instability and stability of pattern formation and symmetries over time. We also run several
test trials to evaluate the accuracy and further understand the numerics of the Gerris simulation
in 3D based on changes in resolution and boundary behavior, while also varying the viscosity
ratio and the surface tension coefficient (thus changing the capillary length) and the initial
perturbation interface of the trials.
For instance, we initially observe that in the case of a large domain size, droplet breakup
and formation in 3D demonstrated extremely unsymmetrical patterns that are inconsistent with
the Rayleigh-Plateau Instability, as droplets formed but began to fall to the sides of the box with
non-periodic boundary conditions in lieu of straight down. We concluded that these strange
droplet behaviors reflect potentially the insufficiency of the spatial resolution of the maximum
octree resolution set in the parameter file, and the level of refinement needs to be increased and
mesh adaptiveness needs to be more localized around the interface curvature to provide a clearer
characterization of the interface to study its development and growth more accurately over time.
Because Gerris simulations in 3D require significantly more computation power compared to
their counterparts in 2D, we optimized computational efficiency in our runs by adjusting Gerris
parameters such that the domain size is large enough to simulate the necessary setup as the

Figure 8: Test droplet formation for a


1.0cm cubic Gfsbox domain with
resolution refinement maximum level of
8, which results in an interesting
oscillatory behavior in droplet breakup

experimental trials designed by Limat and Ferminger, while the resolution is at an optimal value
(which we found to be 8 for a 20.0cm cubic domain size) such that it provides accurate fluid
dynamics calculations and interface behavior with its instability growth, while also small enough
to allow reasonable time for us to observe results with shorter time steps. However, in testing
droplet formation and breakup for a 1.0cm cubic box domain, we observe droplet formation, yet
at level 8 resolution refinement of the mesh grid that adapts with vorticity, the formation of
droplets displays an interesting oscillatory behavior which cannot be completely explained by
low resolution errors alone (Figure 8).

To simulate prior
experimental trials conducted by
Limat and Ferminger, we vary our
physical parameters in Gerris to
create a 20cmx20cm initially flat
interface, with 1600 different 2D
wave modes excited by
superimposed cosine functions at
the interface as white noise. We set
the top liquid to have characteristic
values of silicon oil (dynamic
viscosity: 0.096, surface tension:
0.0207, density: 7.784), while the
bottom fluid of the instability to be
air (dynamic viscosity: 3.62e-6,
Figure 9: 3D visualization of dripping from the flat interface, 3D RT
density:0.0098); the units used in instability in Gerris with a maximum mesh refinement level of 8 along the
this parameter file are in relative interface.
with the 20.0cm length of the cubic
box. The capillary length for this set up is 1.47 mm, and we set the thickness initial fluid layer to
have a control value of 0.25 mm, within the range of the experimental trials of 0.2-0.3mm.
Furthermore, in our simulations, we remove small droplets after they break up from the interface
to optimize the efficiency of the 3D simulations. We run our simulations with variation in
external inflow from the top fluid (at 0.00 mm/s, 0.6 mm/s, 0.8 mm/s, and 1.0 mm/s) over the
course of a few days, and we will continue to run these simulations to observe pattern formation
(as experimental patterns were formed within T = 200-300s) over the course of a few weeks to a
few months. We run our simulations in parallel using 16 cores each on the MPIDS Belenus
cluster.
Currently, although hexagonal patterns and symmetries are not clearly observed as
experimental trials, for simulations with higher flow rates, we observe consistent “roll” like
patterns that develop parallel and along the boundary, consistent with several experimental
observations; on the other hand, simulations with lower flow rates demonstrate a well-defined
homogenous distribution of droplet formation/dripping that correspond to consistent dominant
wave modes, which can be further studied using imaging Fourier analysis, as darker spots on the
interface, viewing it from the top, represent thicker areas of the interface. Nevertheless, the
white noise of the flat interface established by the cosine functions has caused consistent droplet
formation/dripping along five consistent “lines” across the interface domain (clearly seen in
Figure 10(g)). It is not clear whether or not this is a numerical/resolution error or whether it
affects the pattern formed in our simulations. We aim to continue monitoring these pattern
developments of these runs to observe any further interesting stabilities/evolutions/transitions of
these patterns.
(A) Figure 19: Interface between two fluids as
viewed from the top with darker areas
representing thicker areas of the interface, or
places where dripping/droplet formation
occurs during the instability. (A) Inflow rate of
1.0 mm/s with evidence of roll-like patterns
along boundary, (B) No inflow, but thickness
of top fluid layer increased to 4.0 mm, (C)
Inflow rate of .8 mm/s, (D) No inflow, but
thickness of top fluid layer increased to
2.0mm, (E) Inflow with rate 0.6 mm/s

(D)
(B)

(C) (E)

V. Future Directions
As we have gained a greater grasp of the Gerris parameters, units, and boundary
conditions, we should rerun the 2D simulations to follow the exact physical parameters as the
experimental trials, as we maintained in 3D by establishing a smaller domain size in lieu of a 1.0
m x 1.0 m box with a smaller flow rate than the very large flow rate used for the constant
external supply of the top fluid. Correcting these physical parameters for the simulation would
potentially provide us with more accurate results for the analysis of the growth of the instability
in different parameter conditions.
As previously mentioned, as experimental trials take between 200-300 seconds to fully
develop the clear hexagonal patterns of dripping, it would be interesting to allow the simulations
run a little longer (perhaps a couple of weeks or months) to see if any specific pattern stabilizes
that is consistent with experimental results. Furthermore, we can also analyze the growth of the
most dominant wavenumber of each of the simulations in the instability through a Fourier
analysis of the images obtained through the top view of the interface. We also aim to conduct
more simulation runs in further variations of the inflow rate, as well as the initial thickness of the
liquid layer, as it has been shown experimentally that the inflow rate regimes can influence
pattern formation (into dripping or sheets). Nevertheless, in examining the velocity vector
gradient of the development of the instability in our simulations in 3D, we observe some velocity
vectors pointing against the direction of gravity (upwards). We will record this inaccuracy in the
consideration of our simulation results. Furthermore, although the shape of the Gerris boundary
conditions should not affect the pattern formation and hexagonal symmetry, as our GfsBox had a
domain much larger than that used in the experiments, simulating the experiment in a circular
domain would provide more consistent results to the experimental trials.

VI. Thanks
Thank you Dr. Stephan Weiss for dedicating countless hours of your time to help guide,
teach, and mentor me in this project throughout the course of the past two months despite his
undoubtedly very busy schedule. I would also like to thank Prasanth Prabhakaran for being very
engaged in the project, offering very useful advice, discussions, and resources. Finally, I would
like to thank the Max Planck Institute for Dynamics and Self-organization, Princeton University,
and the its International Internship Program for providing me with this very wonderful and
rewarding opportunity and all accompanying resources for my summer in Göttingen.

VII. References
Chandrasekhar, S., Hydrodynamic and Hydromagnetic stability, Dover, 1981, first published by
Oxford University Press, 1961.

Ferminger, M., Limat, L., Wesfreid, J., Boudinet, P., & Quilliet, C. (1992). Two-dimensional
patterns in Rayleigh-Taylor instability of a thin layer. Journal of Fluid Mechanics, 236, 349-383.
Doi: 10.1017/S0022112092001447

L. Limat, P. Jenffer, B. Dagens, E. Touron, M. Fermigier, J. Wesfreid, Gravitational instabilities


of thin liquid layers: dynamics of pattern selection, Phys. D: Nonlinear Phenom. 61 (1) (1992)
166–182.

Popinet, S., 2003. Gerris: a tree-based adaptive solver for the incompressible Euler equations in
complex geometries. J. Comput. Phys. 190, 572–600.

Das könnte Ihnen auch gefallen