Beruflich Dokumente
Kultur Dokumente
PII: S0045-7930(18)30145-2
DOI: 10.1016/j.compfluid.2018.03.041
Reference: CAF 3802
Please cite this article as: D. Deising, D. Bothe, H. Marschall, Direct Numerical Simulation of Mass
Transfer in Bubbly Flows, Computers and Fluids (2018), doi: 10.1016/j.compfluid.2018.03.041
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service
to our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and
all legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT
Highlights
T
fer
IP
• adaptive mesh refinement, load balancing and multi-criterion refine-
CR
ment
1
ACCEPTED MANUSCRIPT
T
a
Mathematical Modeling and Analysis, TU Darmstadt, Darmstadt, Germany
IP
b
Engys Ltd., London, United Kingdom
CR
Abstract
US
Process design, especially in the chemical industry, often requires a detailed
understanding and accurate quantification of interfacial mass transfer. Ex-
perimental investigations, however, are often infeasible as required measure-
AN
ment quantities and locations are inaccessible. Thus, Direct Numerical Simu-
lations (DNS) have become a viable tool for the detailed study of mass trans-
M
fer processes. This paper presents a numerical modelling framework for the
simulation of dilute species transfer based on an algebraic Volume-of-Fluid
ED
Species Transfer (CST) model (Deising et al., 2016) and state-of-the-art high
performance computing techniques are employed.
CE
∗
Corresponding author
Email address: marschall@mma.tu-darmstadt.de (H. Marschall)
1. Introduction
T
aerated bioreactors, which usually do not have any moving parts and aerated
IP
stirred tank reactors (Deckwer, 1985), which are stirred by one or multiple
impellers and play a major role e.g. in waste water treatment, where often
CR
continuous-flow stirred tank reactors are employed. One other two-phase
system of major industrial interest is the dispersed gas-liquid flow in bubble
US
column reactors. This reactor type is widely employed in the chemical and
biochemical industry, whenever mass transfer is needed and the underlying
AN
chemistry is comparably slow. Bubble column reactors are often utilized in
production processes of base chemicals as oxidation, hydrogenation, phos-
genation, alkylation and hydroformylation with a total annual production
M
and interfacial mass transfer processes is of major importance for design and
optimization of such reactors. Experimental studies are expensive and in
PT
2
ACCEPTED MANUSCRIPT
T
2006; Darmana et al., 2007; Radl et al., 2008; Alke and Bothe, 2009; Onea
IP
et al., 2009; Bothe et al., 2011). Typically, in both cases mass transfer effects
CR
on the phase continuity and momentum balance equations are neglected.
Thus, it is emphasised here that the term species transfer implies the numer-
US
ical transport of an inert scalar and should be strictly distinguished from the
expression ’mass transfer ’. Volume effects of mass transfer are accounted for,
e.g., by Bothe and Fleckenstein (2013) and Hayashi and Tomiyama (2011).
AN
The underlying two-phase flow simulation methods to simulate species trans-
fer can be categorised into Front Tracking (Khinast, 2001; Khinast et al.,
M
2003; Koynov et al., 2005; Tukovic and Jasak, 2008; Darmana et al., 2007;
Radl et al., 2008; Aboulhasanzadeh et al., 2012) Level-Set (Yang and Mao,
ED
and Fleckenstein, 2013; Deising et al., 2016; Weiner and Bothe, 2017), and
Arbitrary Lagrangian-Eulerian (ALE) interface tracking (Lehrenfeld, 2015;
CE
3
ACCEPTED MANUSCRIPT
T
This paper aims at the deduction of an improved closure model for the
IP
interfacial transfer of a dilute species from single rising bubbles. This re-
CR
quires the conduction of a comprehensive 3D parameter study which comes
at high computational costs but enables a detailed insight into the underly-
US
ing physical processes of species transfer at single rising bubbles. Unlike in
previous numerical and experimental investigations, in this work the effects
of changes in bubble area and in the local concentration gradients onto the
AN
species transfer are studied separately, allowing of further enhancement of
existing mass transfer correlations.
M
mass transfer length scales due to the thin concentration boundary layer in
close vicinity to the gas-liquid interface. In order to alleviate this problem,
PT
local dynamic adaptive mesh refinement and dynamic load balancing tech-
niques for massively parallel computations using a domain decomposition
CE
4
ACCEPTED MANUSCRIPT
T
of general topology. For the sake of brevity, we only provide an executive
IP
summary of the underlying mathematical model and numerical method.
CR
2.1. Sharp interface continuum model
US
face model, in which material properties are assumed to experience a sharp
discontinuous jump at the location of the phase-separating interface. The
AN
interface Σ (t) is herein assumed to be a surface (i.e. with zero thickness),
which separates two immiscible Newtonian and incompressible fluid phases.
M
read
CE
∇· u = 0 in Ω\Σ(t) , (1)
where Ω = Ω+ (t) ∪ Ω− (t) and Ω+/− (t) denote the respective bulk regions.
For the considered Newtonian fluid flow, the viscous stress tensor reads
T
τ = η ∇u + (∇u) with the dynamic fluid viscosity η. The corresponding
5
ACCEPTED MANUSCRIPT
interface jump conditions are valid at the fluid interface Σ (t) and read
T
where J·K denotes the common jump bracket, nΣ the outer unit vector normal
IP
to the interface and κ = ∇Σ · (−nΣ ) twice the mean interface curvature.
CR
To account for the transfer of a dilute species i, we introduce an additional
set of so-called inert scalar transport quantities for the species concentrations.
US
The corresponding set of transport equations in the bulk phases read as
c+ −
i = Hci at Σ(t) , (7)
ED
where c+
i and ci denote the local concentrations adjacent to the interface in
−
Ω+ (t) and Ω− (t) respectively. The above set of equations for the transport
PT
of species concentration is valid under the assumption that the interface does
not hold mass. The last jump condition represents the so-called Henry law,
CE
6
ACCEPTED MANUSCRIPT
the Finite Volume framework, the method is shortly presented here: to ar-
rive at consistent single field models, application of the CVA technique to
the local instantaneous governing equations of the mathematical model is
indispensable. Application of this technique consists of two steps. The equa-
T
tion is first ’conditioned’ for phase discrimination by multiplication with the
IP
phase-indicator function χ1 (respectively χ2 = 1 − χ1 ), which renders the
CR
equation valid throughout the entire computational domain. In a second
step, volume-averaging is applied to this conditioned equation.
US
The phase indicator, introduced to distinguish between the two phases 1
and 2, separated by a fluid interface SΣ (x, t), is defined as
AN
1 if x ∈ Ω1 (t) at time t
χ1 (x, t) ≡ (8)
0 otherwise.
M
∇χ1 (x, t)
nΣ = . (9)
|∇χ1 (x, t) |
The spatial derivative of the indicator function is thus well-defined and can
PT
|V | V |V |
Z
1 1
Φ := Φ (x + η, t) dη and (11)
|V1 | V1
Z Z
z}|{ 1 1 1
Φ := Φ(x, t) dS = lim Φ(x, t) dS (12)
|SΣ | SΣ aΣ δV →0 δV SΣ
7
ACCEPTED MANUSCRIPT
with aΣ representing the interfacial area density, and applying Gauss and
Leibnitz rules to derivatives (Drew and Passman, 1999; Jakobsen, 2008), the
conditional volume-averaged equation for species transfer with respect to
phase 1 then reads
T
IP
χ1 ∂t c + χ1 ∇· (cu) = χ1 ∇· (D∇c) + χ1 R (13)
1 1 z }| {1
CR
⇔ ∂t(α1 c1 ) + ∇· α1 c1 u1 = ∇· α1 D ∇c1 + α1 R − D∇c · nΣ aΣ .
z }| {1
z}|{
c = H z}|{
1
c .
2
US
z }| {2
D∇c · nΣ aΣ + D∇c · nΣ aΣ = 0 , (14)
(15)
AN
Above equations are the basis for the derivation of a consistent one-field for-
mulation. The volume-averaged jump conditions are herein used for closure
M
modelling of the interfacial terms. More details can be found in Deising et al.
(2016).
ED
∇· u = 0 (16)
CE
∂t α1 + ∇· (α1 u) = 0 (17)
z}|{
(18)
AC
∂t ( ρ u) + ∇· ( ρ u u) = −∇hpi + ρg + ∇· hτ i + fΣ aΣ ,
z}|{
where h· · ·i denotes mean (or mixture) values and fΣ aΣ denotes the volu-
metric force density due to surface tension. Here, we employ the continuous
surface force model (CSF) of Brackbill et al. (1992). Note, that while above
8
ACCEPTED MANUSCRIPT
VOF equations resemble the local instantaneous balance equations, they are
indeed mixture equations derived using the CVA technique.
T
Employing the immersed interface concept which is based on Peskin
IP
(1977), Equation (13) is summed up over both phases 1 and 2 in order to ob-
CR
tain a single field formulation for interfacial species transfer. Exploiting the
surface averaged interfacial jump condition (Eqn. 14), defining the mixture
concentration c ≡ α1 c1 +α2 c2 and ignoring chemical reactions, the following
US
– yet unclosed – single field model is derived
1 2
AN
∂t c + ∇· ( c u) = ∇· α1 D ∇c1 + α2 D ∇c2 . (19)
The basic idea to arrive at a closed single field formulation for interfacial
M
et al. (2010) and Marschall et al. (2012) introduce the follwing approximation
z}|{
CE
1
c+ c c1
H = − = z}|{2 ≈ 2 . (20)
c c c
9
ACCEPTED MANUSCRIPT
T
α1 + H H α1 + 1−α
H
1
IP
Based on above closure relation, it was found in Deising et al. (2016) that
the r.h.s. of (19) can be rewritten in different ways, depending on the
CR
definition of the mean diffusion coefficient which can be chosen arbitrar-
ily. The mean value calculations typically applied are the arithmetic or
cient hDih ≡
1
D D
2
2
α1 D +α2 D
1,
US
harmonic mean. When choosing a harmonic mean mixture diffusion coeffi-
the following single field formulation for interfacial
AN
species transfer is obtained:
1 − H1
∂t c + ∇· ( c u) = ∇· (hDih ∇c) − ∇· hDih c∇α1 . (22)
α1 + 1−α 1
M
This equation was obtained in Haroun et al. (2010) and Deising et al. (2016)
independently, using different mathematical derivations. It is further shown
ED
1 2 1
∂t c + ∇· ( c u) = ∇· (hDia ∇c) + ∇· D − D α1 − 1 ∇c
α1 + 1−α
H
1
" !! #
CE
1 2 2
c 1 D −D 1 D
+ ∇· 1−α1 − D − ∇α1 .
α1 + 1−α
H
1
H α1 + H
H
(23)
AC
10
ACCEPTED MANUSCRIPT
methods yield very similar results. In under-resolved cases, the derived sin-
gle field species transfer models, termed Continuous Species Transfer (CST)
model(s), will lead to an overprediction of the species transfer. For details, it
is referred to Marschall et al. (2012); Deising et al. (2016); Hill et al. (2017).
T
IP
3. Numerical Solution Procedure
CR
3.1. Algebraic VOF
The employed two-phase solver is based on the interFoam solver but con-
US
tains a series of enhancements and additions. To briefly introduce the utilized
interFoam solver, some comments are given here on the employed algorithm
AN
and available literature. Main focus in this introduction is the flux-based
formulation and the algebraic advection algorithm.
The interFoam solver is based on the VOF equations (16 – 18) and suit-
M
in the interFoam solver resembles (16 – 18) – due to the utilized technique
of equation mimicking. However, the solver utilizes a flux-based formulation
PT
for the set of equations, i.e. does not solve for the velocity u but for the face
flux defined as Ff := Sf · u f . The discretization of the surface tension force
CE
11
ACCEPTED MANUSCRIPT
face. Algebraic VOF methods are widely employed also in commercial CFD
codes as Ansys CFX and Fluent but usually exhibit difficulties proprietary
with the transport of sharp fields. The algebraic advection in the solver-
family interFoam utilizes a flux-corrected transport (FCT) method (Zalesak,
T
1979) called MULES (multidimensional universal limiter with explicit solu-
IP
tion) (Weller, 2006). This FCT is unique in the way the limiter is calculated
CR
at each cell face, using an iterative approach. However, unlike the standard
interFoam solver, this work utilizes the Interface Capturing Scheme CICSAM
US
(Ubbink, 1997) in a blended manner (Deising et al., 2016) for advection of
the phase fraction and concentration fields due to its advantage in accuracy
and consistent advection.
AN
common practice to employ methods that allow to follow the bubble during
the simulation. To achieve this, one can either move the computational do-
main with the bubble (Moving Window Technique) as e.g. Bothe and Flecken-
PT
stein (2013); Weiner and Bothe (2017), while observing the flow system from
the same inertial frame, or transform the governing equations into a frame
CE
12
ACCEPTED MANUSCRIPT
Inertial Reference Frame (IRF) technique, where only the phase fraction ad-
vection equation is transformed into the moving reference frame.
A sound theoretical basis for the transformation of equation systems into
reference frames can be found in Muschik and Restuccia (2001); Essén (2002)
T
and an example for a concrete implementation in terms of a non-constant
IP
non-rotating frame is given e.g. in Rusche (2002); Cariglino (2013)
CR
For validation, the two different reference frame formulations have been
compared against each other for a path-instable rising bubble at Eo = 1.2
US
and Mo = 1.267e-10, resulting in Remax ≈ 580. The simulations are con-
ducted on a relatively coarse mesh with a maximum local resolution by cells
of 1/36 bubble diameter size in the region around the interface. Figure 1
AN
shows the results of rise velocity over time and bubble trajectory. Here, the
bubble trajectory was projected onto cylindrical coordinates for easier com-
M
parison. The resulting path of the bubbles shows a zigzag behaviour for both
approaches. However, the plane in which the bubble rises is different. While
ED
the bubble trajectory is very similar for both reference frames (cf. Figure
1b), the rise velocity shows significant differences regarding its transient be-
PT
haviour (cf Fig. 1a). The change in the bubble velocity over time is about
2.5 times larger for the MRF. The cause for this effect is most probably the
CE
additional acceleration term in the momentum equation for the MRF formu-
lation. The bubble shape and velocity streamlines over time for the MRF
case are shown in Figure 1c.
AC
13
ACCEPTED MANUSCRIPT
0.35
MRF
0.3 IRF
300
0.2
200
T
0.15
IP
0.1 100
0.05 MRF
IRF
CR
0 0
0 0.2 0.4 0.6 0.8 1 1.2 −2 0 2 4 6 8 10
time [s] radial position [mm]
US
AN
M
Figure 1: Comparison of MRF and IRF for a path-instable single rising bubble
includes discussion of the bubble path and temporal behaviour of the lateral
PT
velocity components, which are both very similar for the MRF and IRF
with differences being less than 10%. The amplitude of the change in rise
CE
velocity can only be estimated from (Mougin and Magnaudet, 2002, Fig.
6), according to which the bubble rise velocity predicted by the IRF is in
AC
better agreement. Thus, the IRF supposedly yields more accurate results.
Also, it is numerically more stable as it avoids having a controller-dependent
acceleration term in the momentum equation.
14
ACCEPTED MANUSCRIPT
T
reduces parasitic currents in the velocity field. However, this requires an ac-
curate method for the calculation of interface curvature. The implementation
IP
in interFoam computes the curvature from the divergence of the interface nor-
CR
mal, approximated by a simple gradient computation of the volumetric phase
fraction field which is then interpolated to the face centres. This estimation
US
of the curvature is very inaccurate and deviations of the obtained to the exact
curvature are up to several hundred percent (≈ +/− 300 . . . 500 %). Applica-
tion of smoothing algorithms for the phase fraction field prior to calculating
AN
the interface normal as suggested e.g. by Ubbink (1997); Rusche (2002) are
not able to circumvent this problem and offer only a small enhancement of
M
racy by using height functions (Popinet, 2009), which, however, is not trivial
to adapt to unstructured grid arrangements. Therefore, an iso-surface recon-
PT
struction of the phase fraction field based on the work of Batzdorf (2015)
is employed here, which reconstructs the interface as a volume fraction iso-
CE
X
κ=− (nΣ )f,iso · Sf . (24)
f
Actual cell cutting to obtain an interface surface is not performed, the al-
15
ACCEPTED MANUSCRIPT
gorithm is only used for calculating the interface normal vectors and the
interface curvature in cells containing the interface. The reconstruction algo-
rithm of Batzdorf (2015) is incorporated in the solver framework developed
in this work and used for the sole purpose of increasing the accuracy of the
T
interface normal vector and thus the curvature calculation. Further changes
IP
on the surface tension model are not included, as the employed approach
CR
ensures balance of pressure and surface tension forces by using the same
discretization stencils for surface tension and pressure equation.
US
To prove that the underlying algorithm allows for force balancing, sim-
ulations of a 2D liquid disc in zero gravity as presented in Francois et al.
(2006) are performed. The volumetric phase fraction field is initialized nu-
AN
merically exact due to computing the phase fraction values by cutting of a
cylindrical surface with the utilized mesh, employing the library presented in
M
Maric et al. (2013). Tables 1 and 2 show the resulting maximum (spurious)
velocities for a prescribed exact curvature and calculated curvature.
ED
Table 1: Error in maximum velocity |u|max (in m/s) after one time step for the inviscid
static drop in equilibrium when the exact curvature is specified
PT
ρ1 /ρ2 interCSTFoam CSF cell-centred (Francois et al., 2006) SSF face-centred (Francois et al., 2006)
1 9.81e − 19 5.19e − 5 5.43e − 19
103 9.11e − 17 6.15e − 3 4.44e − 18
CE
16
ACCEPTED MANUSCRIPT
Table 2: Effect of the fluid density ratio on the error in maximum velocity |u|max after
one time step for the viscous static drop in equilibrium
T
10 1.99e − 4 1.93e − 5 1.17e − 18 4.40e − 7 2.30e − 6
IP
103 2.47e − 3 5.03e − 4 7.14e − 17 6.22e − 7 2.59e − 6
105 2.01e − 1 5.30e − 2 5.24e − 13 6.25e − 7 2.59e − 6
CR
imental data of Duineveld (1995), who conducted a series of measurements
in an ultra-pure water system (without influence of surfactant). The investi-
US
gated range of bubble diameters is db ∈ [0.3 mm . . . 1 mm]. For the numerical
simulations, the bubble diameters db = {0.35, 0.5, 0.65, 0.85, 0.95 mm} are
AN
chosen. The comparison in Figures 2a and 2b show the results for bub-
ble rise velocity and bubble aspect ratio, respectively. For the simulations,
the blended CICSAM advection scheme and the iso-surface-based curvature
M
model are used. It can be seen that the numerical results deviate only slightly
ED
40
2 Duineveld 1995
35 Moore 1965
standard interFoam
1.8
rise velocity [cm/s]
interCSTFoam
aspect ratio [-]
30
PT
1.6
25
1.4
20 Duineveld 1995
Moore 1965 1.2
CE
15 standard interFoam
interCSTFoam 1
0.4 0.6 0.8 1 0.4 0.6 0.8 1
radius [mm] radius [mm]
AC
from the experimental results. It should be noted that this relatively good
17
ACCEPTED MANUSCRIPT
T
interFoam solver framework as utilized in this work is capable of simulating
IP
rising bubble hydrodynamics with sufficient accuracy.
CR
3.3. Parallelisation and HPC
US
trial processes exhibit high Peclet numbers, which means that the transport
by advection is several orders of magnitude faster than the diffusive transport,
AN
leading to extremely thin concentration boundary layers being typically 2 to
3 orders of magnitude thinner than the hydrodynamic boundary layer. The
M
limiting factor for numerical simulations is thus the Schmidt number, which
directly relates to the decrease in concentration boundary layer thickness rel-
ED
rise (cf. Fig. 6a). The simultaneous application of both techniques massively
reduces the overall computation time. To quantify the speed-up (computa-
tion time reduction factor) compared to utilizing a uniform mesh of the same
resolution at the interface, a comparison for the 3D dam break tutorial case
18
ACCEPTED MANUSCRIPT
with obstacle was conducted (cf. Figure 3). The reference solution was
100
T
30
speed-up [-]
IP
10
CR
3 Uniform mesh
Linear speed-up
AMR & LB
1
1 2 4 8
Figure 3: Scale-up plot for 3D dam break with obstacle using uniform mesh and AMR
AN
with load balancing
cells per iteration, the mapped flux used in the pressure correction equation
and the load balancing interval. These, however, strongly depend on the
CE
local solution and thus can only be reliably assessed by simulating a physical
problem rather than by conventional HPC assessment over a few iterations.
AC
Due to the very high overall simulation time for this study (cf. Table 3),
simulations on less than 8 cores have not been performed.
In the employed OpenFOAM-library both techniques are already present
in principle but required modifications in order to function together. In the
19
ACCEPTED MANUSCRIPT
Table 3: Computational time (in hours) for uniform and dynamic mesh on different num-
bers of cores
mesh type 1 core 8 cores 16 cores 32 cores 64 cores 128 cores 512 cores
uniform mesh − 233 122.4 77.8 52.2 26.5 7.35
T
dynamic mesh 63.4 7.83 5.45 3.15 − − −
IP
recent few years, two different implementations to achieve load balancing
CR
in OpenFOAM have been accomplished, at TTD TU Darmstadt (Batzdorf,
2015), and at Purdue University (T.G. Voskuilen). The former is based on
US
the use of a special decomposition method (clustered decomposition) to main-
tain the refinement history required for AMR, while the latter includes the
load balancing directly into the mesh refinement algorithm. Above proposed
AN
methods have been unified in the scope of this work into one library frame-
work and further improved through major enhancement of the multi-criterion
M
ple refinement criteria and respective target refinement levels for this criterion
by means of a dictionary file, the common input format for OpenFOAM. For
AC
each criterion, minimum and maximum values of the respective quantity and
the number of maximum refinement steps can be specified. Further, domain
regions can be refined by defining geometrical shapes as boxes and spheres.
An exemplary dictionary entry, leading to the refined mesh shown in Figure
20
ACCEPTED MANUSCRIPT
dynamicFvMesh
dynamicRefineBalancedFvMesh ;
refinementControls
{
enableRefinementControl true ;
interface
(
alpha1 (2 5)
);
fields
(
T
alpha1 ( 0 . 0 1 1.1 3)
C1 ( 0 . 0 0 1 0 . 0 5 2 )
);
IP
gradients
(
alpha1 ( 0 . 0 1 2 2)
);
curls
(
CR
U ( 1 0 0 1 e +05 3 )
);
regions
(
cylinderToCell
{
p1 (0.015 0.015 0.015);
p2 (0.015 0.033 0.015);
US
radius 0.006;
}
);
}
5b, is given in Figure 5a for the DNS of species transfer from a single bubble.
PT
Local adaptive mesh refinement leads to new cells and faces, at which
values of the respective transport quantities have to be specified. For fields
CE
and for unrefined cells performs a simple averaging onto the agglomerated
cell. For face-centred fields, however, the employed mapping algorithm in the
OpenFOAM library simply copies the value from local face zero of the parent
cell onto all newly generated faces of the child cells. As the solver relies on
21
ACCEPTED MANUSCRIPT
a divergence free volumetric flux field, the field needs to be corrected after
each AMR step in order to ensure a divergence free flux. For this, a pressure
correction step is introduced after each AMR step, which requires a good
approximation of the flux field on newly generated faces. In this work, the
T
flux field is re-computed from the face-centred velocity field uf , obtained by
IP
projecting the interpolated velocity tangential to the face and using the flux
CR
normal to the face, i.e.
Ff
uf = (I − nf ⊗ nf ) (u)f (CD) + nf , (25)
US |Sf |
which is mapped onto the newly generated faces by averaging of neighbor face
values. This procedure significantly increases the robustness of AMR com-
AN
pared to the standard flux mapping in OpenFOAM, where the face fluxes
are reconstructed from the cell-centred velocity field employing linear inter-
M
polation.
In terms of computational cost per time step, the mesh update takes
ED
about 12% of the time. However, the subsequent flux correction takes about
25−30% of the computation time per time step. The flux mapping algorithm
PT
introduced here significantly reduces that cost due to a better initial solution
and thus faster solver convergence. To maintain a reasonable computational
CE
cost while ensuring that the interface remains well within the refinement
region, AMR is executed every 20 Iteration for all simulations presented here.
Thus, the overall computational cost increase due to dynamic refinement is
AC
22
ACCEPTED MANUSCRIPT
T
threshold is reached. This mainly involves the combination of two steps which
are available within the OpenFOAM framework. At first, a decomposition is
IP
created, which is basically a list of integer values that specify for each cell in
CR
the domain the respective core it belongs to. In a second step, the mesh and
all fields are communicated to the cores based on the aforementioned list.
US
However, several problems with the current load balancing method occur.
One main drawback of the mesh balancing approach introduced by (T.G.
Voskuilen) is that all simulation steps, including the pre-processing, need to
AN
be executed in parallel in order to correctly maintain the refinement history
(cell refinement level). Also, a redistribution onto a different number of
M
23
ACCEPTED MANUSCRIPT
T
IP
CR
(a) Core distribution at time (b) Core distribution at time
t = 0.02 s.
US t = 0.17 s.
fields only when they are flux quantities, thus allowing to recompute the
volumetric flux field from the mapped surface velocity field.
AC
4. Simulation Results
This Section presents the results for species transfer from single rising
bubbles in a quiescent liquid (infinite medium) using the presented VOF/CST
24
ACCEPTED MANUSCRIPT
T
to moderate and high numbers (Sc ∈ {0.03, 0.1, 0.3, 1, 3, 10, 30, 100, 300},
IP
Pe ∈ [10, 104 ]), with realistic diffusion coefficient ratios (up to 104 ) and Henry
CR
coefficients (between 5 and 30).
A detailed model verification and validation study has been shown in
US
Deising et al. (2016) and thus omitted in this paper.
parameter study has been performed to obtain insight into the influence of
the Reynolds and Schmidt numbers onto the species transfer process.
ED
Making use of local adaptive mesh refinement and dynamic load balanc-
ing strategies, in this work the concentration boundary layer is sufficiently
resolved for Peclet numbers up to Pe = 104 (cf. Figure 9b), while still main-
PT
around the interface for all cases in the parameter study is about 224 cells
per bubble diameter. The total mesh cell count is about 4·106 cells, although
AC
it depends for each case – due to the dynamic adaptive mesh refinement –
on the respective bubble shape. Figure 6 shows the concentration profiles
around rising single bubbles for two different material parameter settings.
In addition to the species concentration profiles, concentration iso-surfaces
25
ACCEPTED MANUSCRIPT
T
IP
CR
(a) Eo = 0.3, Mo = 10−11 .
US (b) Eo = 5, Mo = 10−5 .
profile looks similar, while the flow around both bubbles is entirely different.
The velocity and pressure fields for case Eo = 0.3, Mo = 10−11 are given in
ED
Figure 7. Figure 8 shows the cell values (original data, non-smoothed) of the
species concentration at Pe ≈ 6300. In Figure 8b, additionally, the employed
PT
26
ACCEPTED MANUSCRIPT
where AΣ denotes the interfacial area and (cl,Σ − cl,∞ ) is the liquid-sided
difference between the interface and the bulk concentrations and represents
the driving force of the interfacial species transfer. The area AΣ can be
expressed as the area of a spherical volume-equivalent bubble Asphere times
T
the surface ratio Sr := AΣ /Asphere . This approach introduces two unknowns
IP
which need modelling, the species transfer coefficient kl and the interfacial
CR
area AΣ or surface ratio Sr .
In order to calculate kl from experimental studies of mass transfer, the
US
interfacial area needs to be obtained. While for spherical bubbles, AΣ can
be obtained by visual observation and measurement of the bubble diame-
ter, the bubble area for non-spherical bubbles or bubbles undergoing shape
AN
changes (oscillation, wobbling) can only be estimated. From the simulation
results however, all relevant information is readily available. Additionally,
M
Figure 7: Velocity and pressure field for case Eo = 0.3, Mo = 10−11 at t = 0.1379 s
27
ACCEPTED MANUSCRIPT
T
IP
CR
(a) Concentration field.
US (b) Logarithmic concentration
field.
AN
Figure 8: Concentration field for case Eo = 0.3, Mo = 1e-11, Sc = 10 at t = 0.1379 s
commonly only the global mass transfer (kl A) is measured, which lumps the
effects of changes in local concentration gradients and increase in bubble area
ED
and thus contains less information about the actual physics. Due to the lack
of detailed information, experimental results are only partially suitable for
PT
(∇c·nΣ ) db
number (Shloc ), defined as Shloc := cΣ −c∞
from which the global Sherwood
R
number can be defined as Sh := 1
SΣ
Shloc dS ≡ kl db
D
. An important differ-
SΣ
ence between this work and the vast majority of literature is, that we define
the Sherwood number based on the actual bubble area rather than the area
28
ACCEPTED MANUSCRIPT
T
vant correlations is given in (Green and Perry, 2007, pp. 5-71 to 5-73) or in
IP
Colombet et al. (2014). Further relevant correlations can be found in Lochiel
CR
and Calderbank (1964); Oellrich et al. (1973); Clift et al. (1978); Takemura
and Yabe (1998). These correlations are summarized in Table 4.
US
AN
M
ED
PT
CE
AC
29
ACCEPTED MANUSCRIPT
T
IP
CR
Table 4: Sherwood correlations for mass transfer from single rising bubbles
Comments
Correlation
Sh = 1.0(Re Sc)1/3
US
E = Empirical, S = Semiempir-
ical, T = Theoretical
Sh = 1.13(Re Sc)1/2 [T] small (spherical) bubbles Sherwood et al. (1975); Green and Perry (2007)
db
30
Sh = 1.13(Re Sc)1/2 [S] medium to large bubble, Johnson et al. (1969); Sherwood et al.
0.45 cm + 0.2db
M
Sh = 2 + 0.31 Gr1/3 Sc1/3 [S] Calderbank and Moo- Calderbank and Moo-Young (1961);
ED
Sh = 0.42 Gr1/3 Sc1/2 [S] Calderbank and Moo- Calderbank and Moo-Young (1961);
Young correlation, carbon Geankoplis (1993); Green and Perry
dioxide in water/glycerol, (2007); Kirwan (1987); Lee (1992); Shah
CE
T
IP
CR
!0.116
db g 1/3
Sh = 2 + 0.061 Sc0.546 Re0.779
s 2/3
[E] Hughmark correlation, Hughmark (1967); Treybal (1980);
DL
Res with slip velocity, Green and Perry (2007)
air/water-glycerol system
Sh = 2 + 0.651
Sh = 2 +
(Re Sc)1.72
1 + (Re Sc)1.22
0.232(Re Sc)1.72
US
[T]
[T]
Oellrich correlation, small
bubbles, Re → 0, Sc → ∞
Sh = 1.13(Re Sc)1/2 f (χ) [T] aspect ratio χ, Lochiel and Calderbank (1964); Colom-
for oblate spheroids bet et al. (2014)
31
Sh = 1.13(Re Sc)1/2 f (χ) [S] aspect ratio χ ∈ [1, 3], Figueroa-Espinoza and Legendre (2010);
ED
2 1
Sh = 1.13 1 − (2.5 + Pe1/2 ) [E] nearly spherical bubbles, Takemura and Yabe
3 (1 + 0.09 Re2/3 )3/4
0 < Re ≤ 100, Sc 1 (1998); Colombet
et al. (2014)
CE
AC
ACCEPTED MANUSCRIPT
T
IP
CR
US
AN
32
M
ED
PT
CE
AC
ACCEPTED MANUSCRIPT
T
While the number of independent dimensionless numbers describing mass
IP
transfer from rising bubbles is fixed, almost all reported correlations include
CR
only some dependencies and are thus strongly limited in their applicability.
Only very few correlations account for the actual shape of the rising bubble,
US
while the vast majority of mass transfer correlations only account for the ratio
of hydrodynamic to diffusion boundary layer thickness (Schmidt number) and
ratio of inertia/buoyancy to viscous forces (Reynolds or Grashof number).
AN
The most common form of a Sherwood correlation is
∂t c + ∇· (cu) = ∇· (D∇c) ,
33
ACCEPTED MANUSCRIPT
Defining urel trel = xrel then leads to the non-dimensional species transport
equation with one non-dimensional parameter
Drel νrel
∂t∗ c∗ + ∇∗ · (c∗ u∗ ) = ∇∗ · (D∗ ∇∗ c∗ ) (29)
T
νrel xrel urel
1
= ∇∗ · (D∗ ∇∗ c∗ ) .
IP
Re Sc
CR
on the Peclet number (Pe = Re Sc). From this result, the Sherwood corre-
lation (which is a non-dimensionalised form of the concentration gradient at
US
the interface) is assumed to be a function of the Peclet number only.
However, the dimensionless velocity profile and the bubble shape is a func-
AN
tion of the Eötvös and Reynolds numbers. Thus, the Sherwood correlation
is assumed to take the following form:
M
where f (Re, Eo) denotes a correction factor describing the influence of changes
ED
in the local velocity field due to bubble deformation. This form of correlation
can also be found in literature, e.g. in Lochiel and Calderbank (1964), who
PT
introduce an aspect ratio to include the bubble shape influence onto the ve-
locity field. From the above reasoning, it can be deduced that the influence
CE
of the Schmidt number for a given set of material parameters (and thus fixed
velocity field) is approximately constant. From the solution of a potential
AC
34
ACCEPTED MANUSCRIPT
T
evolution of the dynamically changing bubble surface area and its depen-
dency on the global Sherwood number. The main results are visualised
IP
in Figures 9a and 9b. Figure 9a shows that the interfacial area increases
CR
1.7
Eo = 0.3
102
Sc = 0.3
Eo = 1
1.6 Sc = 1
Eo = 2
US
Sc = 3
Eo = 5
Sc = 10
1.5 Eo = 10
Sc = 30
Eo = 20 101.5 Sc = 100
correlation
1.4 a · sqrt(Pe)
Sr
Sh
AN
1.3
101
1.2
1.1
M
1 0 100.5
10 101 102 101 102 103 104
Re Pe
Figure 9: Surface ratio and Sherwood number as function of Reynolds and Peclet number
PT
g(Eo) ≈ 0.726 + 0.317 1+(1.6081 Eo)1.426 ∈ [0.726, 1]. Here, the influence of the
Eötvös number is dominant, which is not surprising. The well-known Grace
AC
diagram (Clift et al., 1978) for instance visualises – although not quantifies –
this effect. Figures 9b and 11a on the other hand show that the global Sher-
wood number decreases with increasing bubble deformation (i.e. increasing
Eötvös number). However, Figure 9b also shows that the decrease in global
35
ACCEPTED MANUSCRIPT
102
Eo = 0.3 , Mo = 3 · 10−7
Eo = 0.3 , Mo = 10−8
Eo = 0.3 , Mo = 10−10
Eo = 0.3 , Mo = 10−11
Eo = 1 , Mo = 10−5
101.5
Eo = 1 , Mo = 10−7
Eo = 1 , Mo = 3 · 10−9
T
Sh
Eo = 2 , Mo = 10−4
Eo = 2 , Mo = 10−6
Eo = 2 , Mo = 5 · 10−9
IP
1
10
Eo = 5 , Mo = 10−3
Eo = 5 , Mo = 10−5
Eo = 10 , Mo = 10−2
Eo = 10 , Mo = 3 · 10−5
CR
Eo = 20 , Mo = 3 · 10−2
100.5
10−2 10−1 100 101 102
Sc
US
Sherwood number due to bubble deformation is strictly limited by a lower
AN
bound, while the upper bound can be approximately by Higbie’s penetration
√
theory (Higbie, 1935), i.e. by Shmax = √2π Pe. Our simulations show that the
upper limit is only surpassed for (nearly) spherical path-instable bubbles (cf.
M
102 0.6
0.55
PT
101.5
Exponent
Sh
0.5
Eo = 0.3 , Mo = 10−10
Eo = 0.3 , Mo = 10−11
CE
101 Eo = 1 , Mo = 3 · 10−9
Eo = 0.3
Eo = 2 , Mo = 5 · 10−9 0.45 Eo = 1
−5
Eo = 5 , Mo = 10 Eo = 2
Eo = 10 , Mo = 3 · 10−5 Eo = 5
penetration model (Higbie) Eo = 10
100.5 0.4
AC
36
ACCEPTED MANUSCRIPT
number on the Schmidt number. It was found from our simulation results
that this functional relation follows in good approximation an exponential
law with a constant exponent and a material system-dependent pre-factor.
This means that the global Sherwood number for any given material system
T
is seen to solely depend on the Schmidt number by
IP
Sh = a · Scb , (30)
CR
where a = g (Eo, Re) and b have been found to be almost constant. This then
results with a set of curves, where all our simulation data points lie on straight
US
lines in the double-logarithmic plot as shown in Figure 10. Fitting b for
different cases yields values of approximately 0.5 (cf. Figure 11b) and thus, as
AN
expected, varies only slightly from the theoretical value of 0.5, obtained, e.g.,
in the analytical solution of potential flow around a sphere. However, Figure
M
11b shows that the exponent does indeed slightly increase with increasing
Reynolds and Eötvös number (i.e. increasing bubble deformation). The
ED
hence depending on the surface ratio and Reynolds number. This form is
also suggested by Lochiel and Calderbank (1964) based on analytical in-
CE
37
ACCEPTED MANUSCRIPT
number, the Schmidt number is found to also affect overall mass transfer,
although this effect is small (cf. Figure 11b).
Eo = 0.3, Sc = 1
102
Eo = 1, Sc = 1
T
Eo = 2, Sc = 1
Eo = 5, Sc = 1
Eo = 10, Sc = 1
IP
Eo = 0.3, Sc = 10
101.5 Eo = 1, Sc = 10
Eo = 2, Sc = 10
Sh ·Sr
Eo = 5, Sc = 10
CR
Eo = 10, Sc = 10
Eo = 0.3, Sc = 100
Eo = 1, Sc = 100
101
Eo = 2, Sc = 100
Eo = 5, Sc = 100
Eo = 10, Sc = 100
100.5
101 102
Pe US
103 104
both upper and lower bound, as well as the dependency of the Sherwood
number on the Schmidt number.
CE
the different curves, the numerical results are plotted separately for each
Eötvös number against the correlation of Takemura and Yabe (1998). As
the correlation is only valid for large Schmidt numbers, deviation from the
numerical results at small Schmidt numbers can be ignored. However, espe-
38
ACCEPTED MANUSCRIPT
T
IP
CR
US
AN
M
ED
Figure 13: Comparison of simulation results with correlation of Oellrich et al. (1973)
PT
cially for very small and very large Reynolds numbers, a deviation from the
proposed correlation can be seen. The deviation at large Reynolds numbers
CE
remains unclear. Overall, it can be seen that the numerical results agree well
with the correlation.
39
ACCEPTED MANUSCRIPT
T
IP
CR
US
AN
M
ED
Figure 14: Comparison of simulation results with correlation of Takemura and Yabe (1998)
PT
40
ACCEPTED MANUSCRIPT
in the inertial frame of a fixed observer. This setup is similar to the moving
window technique employed, e.g., in Bothe and Fleckenstein (2013).
In order to resolve the thin concentration boundary layer, state-of-the-art
high performance computing techniques are employed. The computational
T
mesh is dynamically refined at the interface and in the bubble wake, based
IP
on a multi-criterion refinement strategy. To ensure computational efficiency,
CR
dynamic load balancing is performed based on a clustered hierarchical de-
composition.
US
A parameter study at single rising bubbles has been conducted in order
to investigate the influence of the local flow field and the Schmidt number
onto the species transfer process. It is found that the bubble shape strongly
AN
influences the transfer. Stronger deformation leads to a significant decrease
in the global Sherwood number, while at the same time leading to an increase
M
tively. Further it is shown that a is always close to 0.5 but slightly increasing
with bubble deformation and rise velocity.
CE
41
ACCEPTED MANUSCRIPT
T
higher Eötvös numbers (Eo > 5) and moderate to high Reynolds numbers.
IP
Further, the presented numerical results indicate that the correlations pro-
CR
posed by Oellrich et al. (1973) and Takemura and Yabe (1998) are a good
estimate in a large parameter range. The numerical findings of this work
US
can be utilized to improve the accuracy of detail-reduced simulation meth-
ods. An example is the two-fluid model, which requires an accurate species
transfer correlation over a wide range of bubble shapes and local Reynolds
AN
numbers and also an accurate description of the local interfacial area, which
should also account for changes due to interface deformation.
M
References
ED
42
ACCEPTED MANUSCRIPT
T
transfer processes at fluid particles. Chem. Eng. Sci., 101:283–302, 2013.
IP
D. Bothe, M. Koebe, K. Wielage, and H.J. Warnecke. VOF-simulations of
CR
mass transfer from single bubbles and bubble chains rising in aqueous so-
lutions. In 4th ASME-JSME Joint Fluids Engineering Conference, number
FEDSM2003-45155, July 6-11 2003.
US
D. Bothe, M. Koebe, K. Wielage, J. Prüss, and H.-J. Warnecke. Direct
AN
numerical simulation of mass transfer between rising gas bubbles and wa-
ter. In Martin Sommerfeld, editor, Bubbly Flows: Analysis, Modelling
and Calculation, pages 159–174, Berlin, Heidelberg, 2004. Springer Berlin
M
Heidelberg.
ED
J.U. Brackbill, D.B. Kothe, and C. Zemach. A continuum method for mod-
CE
P.H. Calderbank and M.B. Moo-Young. The continuous phase heat and mass
AC
43
ACCEPTED MANUSCRIPT
R. Clift, J.R. Grace, and M.E. Weber. Bubbles, Drops, and Particles. New
York ; London : Academic Press, 1978.
T
and mass transfer of rising bubbles in a homogenous swarm at large gas
IP
volume fraction. J. Fluid Mech., 763:254–285, 2014.
CR
D. Darmana, N.G. Deen, and J.A.M. Kuipers. Detailed 3D modeling of mass
transfer processes in two-phase flows with dynamic interfaces. Chem. Eng.
Technol., 29(9):1027–1033, 2006.
US
D. Darmana, W. Dijlhuizen, N.G. Deen, M. van Sint Annaland, and J.A.M.
Kuipers. Detailed 3D Modelling of Mass Transfer Processes in Two-Phase
AN
Flows with Dynamic Interfaces. In 6th International Conference on Mul-
tiphase Flow, ICMF 2007, 2007.
M
44
ACCEPTED MANUSCRIPT
T
D.A. Drew. Mathematical modeling of two-phase flow. Annu. Rev. Fluid
IP
Mech., 15:261–291, 1983.
CR
D.A. Drew and S.L. Passman. Theory of Multicomponent Fluids. Springer-
Verlag New York, 1999.
US
M. Dudukovic. Relevance of multiphase reaction engineering to modern tech-
nological challenges. Ind. and Eng. Chem. Res., 46:8574–8686, 2007.
AN
P.C. Duineveld. The rise velocity and shape of bubbles in pure water at high
Reynolds number. J. Fluid Mech., 292:325–332, 6 1995. ISSN 1469-7645.
M
L.C. Evans and R.F. Gariepy. Measure Theory and Fine Properties of Func-
PT
gas bubbles rising through a stationary liquid. Chem. Eng. Sci., 65:6296–
6309, 2010.
45
ACCEPTED MANUSCRIPT
M.M. Francois and N.N. Carlson. The Balanced-Force Volume Tracking Algo-
T
rithm and Global Embedded Interface Formulation for Droplet Dynamics
IP
with Mass Transfer. In ASME 2010 3rd Joint US-European Fluids Engi-
neering Summer Meeting collocated with 8th International Conference on
CR
Nanochannels, Microchannels, and Minichannels, pages 81–88. American
Society of Mechanical Engineers, 2010.
US
M.M. Francois, S.J. Cummins, E.D. Dendy, D.B. Kothe, J.M. Sicilian, and
M.W. Williams. A balanced-force algorithm for continuous and sharp in-
AN
terfacial surface tension models within a volume tracking framework. J.
Comput. Phys., 213:141–173, 2006.
M
C.J. Geankoplis. Transport processes and unit operations. Prentice Hall, 3rd
edition, 1993.
ED
R. Higbie. The rate of absorption of a pure gas into a still liquid during short
periods of exposure. Trans. Am. Inst. Chem. Engrs., 35:365–389, 1935.
46
ACCEPTED MANUSCRIPT
T
G.A. Hughmark. Holdup and mass transfer in bubble columns. Ind. Eng.
IP
Chem. Process Des. Dev., 6:218–220, 1967.
CR
H.A. Jakobsen. Chemical Reactor Modeling: Multiphase Reactive Flows.
Springer-Verlag Berlin Heidelberg, 2008.
US
A.I. Johnson, F. Besik, and A.E. Hamielec. Mass transfer from a single rising
bubble. Can. J. Chem. Eng., 47:559–564, 1969.
AN
J.G. Khinast. Impact of 2-D bubble dynamics on the selectivity of fast gas-
liquid reactions. AIChE J., 47:2304–2319, 2001.
M
J.G. Khinast, A. Koynov, and T.M. Leib. Reactive mass transfer at gas-liquid
interfaces: impact of micro-scale fluid dynamics on yield and selectivity of
ED
2786–2800, 2005.
47
ACCEPTED MANUSCRIPT
T
A.C. Lochiel and P.H. Calderbank. Mass transfer in the continuous phase
IP
around axisymmetric bodies of revolution. Chem. Eng. Sci., 19(7):471 –
484, 1964. ISSN 0009-2509.
CR
T. Maric, H. Marschall, and D. Bothe. voFoam - A geometrical Volume
of Fluid algorithm on arbitrary unstructured meshes with local dynamic
US
adaptive mesh refinement using OpenFOAM. ArXiv e-prints, May 2013.
AN
H. Marschall, K. Hinterberger, C. Schüler, F. Habla, and O. Hinrichsen. Nu-
merical simulation of species transfer across fluid interfaces in free-surface
flows using OpenFOAM. Chem. Eng. Sci., 78(0):111 – 127, 2012.
M
48
ACCEPTED MANUSCRIPT
T
C.S. Peskin. Numerical analysis of blood flow in the heart. J. Comput. Phys.,
IP
25:220–252, 1977.
CR
S. Popinet. An accurate adaptive solver for surface-tension-driven interfacial
flows. J. Comput. Phys., 228:5838–5866, 2009.
US
S. Radl, A. Koynov, G. Tryggvason., and J.G. Khinast. DNS-based pre-
diction of the selectivity of fast multiphase reactions: Hydrogenation of
AN
nitroarenes. Chem. Eng. Sci., 63:3279–3291, 2008.
Y.T. Shah, B.G. Kelkar, S.P. Goobole, and W.-D. Deckwer. Design parame-
PT
ters estimations for bubble column reactors. AIChE J., 28:353–379, 1982.
T.K. Sherwood, R.L. Pigford, and C.R. Wilke. Mass transfer. McGraw-Hill,
CE
1975.
AC
49
ACCEPTED MANUSCRIPT
T
IP
O. Ubbink. Numerical Prediction of Two Fluid Systems with Sharp Inter-
faces. PhD thesis, Department of Mechanical Engineering, Imperial College
CR
of Science, Technology & Medicine, January 1997.
US
transfer based on ALE Interface Tracking. Int. J. Heat Mass Transfer,
104:759–773, 2017.
AN
P.S. Weber. Modeling and Numerical Simulation of Multi-Component Two-
Phase Fluid Systems with Ionic Species. PhD thesis, TU Darmstadt, 2016.
M
PhD thesis.
Limited, 2006.
AC
50