International Journal of Heat and Fluid Flow 56 (2015) 5159

Contents lists available at ScienceDirect

International Journal of Heat and Fluid Flow

journal homepage: www.elsevier.com/locate/ijhff

EulerEuler large eddy simulations for dispersed turbulent bubbly ows

T. Ma a,, T. Ziegenhein a, D. Lucas a, E. Krepper a, J. Frhlich b
a
b

Helmholtz-Zentrum Dresden-Rossendorf, Institute of Fluid Dynamics, Dresden, Germany

Technische Universitt Dresden, Institut fr Strmungsmechanik, Dresden, Germany

a r t i c l e

i n f o

Article history:
Received 24 February 2015
Received in revised form 8 May 2015
Accepted 17 June 2015
Available online 20 July 2015
Keywords:
Bubble column
Two-uid model
Large eddy simulation
Energy spectra

a b s t r a c t
In this paper we present detailed EulerEuler Large Eddy Simulations (LES) of dispersed bubbly ow in a
rectangular bubble column. The motivation of this study is to investigate the potential of this approach
for the prediction of bubbly ows, in terms of mean quantities. The physical models describing the
momentum exchange between the phases including drag, lift and wall force were chosen according to
previous experiences of the authors. Experimental data, EulerLagrange LES and unsteady EulerEuler
Reynolds-Averaged NavierStokes results are used for comparison. It is found that the present model
combination provides good agreement with experimental data for the mean ow and liquid velocity
uctuations. The energy spectrum obtained from the resolved velocity of the EulerEuler LES is presented
as well.
2015 Elsevier Inc. All rights reserved.

1. Introduction
Many ow regimes in nuclear engineering and chemical engineering are gasliquid ows with a continuous liquid phase and
a dispersed gaseous phase. Computational Fluid Dynamics (CFD)
simulations become more and more important for the design of
the related processes, for process optimization as well as for safety
considerations. Because of the large scales that need to be considered for such purposes, the two-uid or multi-uid approach is
often the most suitable framework. During the last years, clear progress was achieved for modelling dispersed bubbly ows. At
Helmholtz-Zentrum Dresden-Rossendorf, in cooperation with
ANSYS, the inhomogeneous Multiple Size Group (iMUSIG) model
was developed (Krepper et al., 2008). It is based on bubble size
classes for the mass balance as well as for the momentum balance.
This model has been later on extended by adding a continuous gas
phase for a generalized two-phase ow (GENTOP) (Hnsch et al.,
2012). The aim of the GENTOP concept is to treat both unresolved
and resolved multiphase structures. The present study concentrates on the turbulence modelling of the unresolved structures.
Turbulence in the liquid phase is an important issue in bubbly
ows as it has a strong inuence on the local distribution of the
dispersed phase and on the bubble size by bubble fragmentation
and coalescence. Compared to the liquid phase the inuence of
the turbulence in the gas phase is generally negligible because of
the low density of the gas and the small dimensions of bubbles.
Corresponding author.
E-mail address: tian.ma@hzdr.de (T. Ma).
http://dx.doi.org/10.1016/j.ijheatuidow.2015.06.009
0142-727X/ 2015 Elsevier Inc. All rights reserved.

A bubble column provides a good experimental system for the

study of turbulent phenomena in bubbly ows. In bubble columns
a wide range of length and time scales exists on which turbulent
mixing takes place. The largest turbulence scales are comparable
in size to the characteristic length of the mean ow and depend
on reactor geometry and boundary conditions. The smaller scales
depend on the bubble dynamics and hence are proportional to
the bubble diameter. In bubbly ows, the small scales are responsible for the dissipation of the turbulent kinetic energy as in
single-phase ow, but the bubbles can also generate back-scatter,
i.e. energy transfer from smaller to larger scales (Dhotre et al.,
2013). The combination of both effects can yield an overall
enhancement or attenuation of the turbulence intensity.
In the present paper the effect of turbulence modelling is
investigated. In the CFD simulations of bubble columns,
Reynolds-Averaged NavierStokes (RANS) models are used for
modelling turbulence in the traditional way, using isotropic
closures without resolution of turbulent scales. Large Eddy
Simulation (LES) offers the possibility to resolve the large-scale
anisotropic turbulent motion and to model the small scales with
a Subgrid-Scale (SGS) model.
Large eddy simulations for such kind of ows have been performed by different authors employing the EulerEuler approach.
Zhang et al. (2006) used LES with the Smagorinsky model to
simulate a square cross-sectional bubble column, with the gas inlet
placed in the centre of the bottom. They compared the results
obtained with different values of the Smagorinsky constant Cs
and found that too high values lead to an unphysically high effective viscosity which in turn damps the bubble plume dynamics.

52

T. Ma et al. / International Journal of Heat and Fluid Flow 56 (2015) 5159

Such an effect should not be found for bubble columns with a gas
inlet at the bottom that homogeneously distributes the bubbles
over the entire cross section, because the gas volume fraction prole in such bubble columns is at and the impact of the increased
effective viscosity is lower (Ma et al., 2015). Dhotre et al. (2008)
reported LES with a two-uid model for the same experiment of
Zhang et al. (2006). They investigated the inuence of SGS models
in LES using the Smagorinsky model and the dynamic Smagorinsky
model of Germano et al. (1991). It was found that the performance
of both models was similar. In fact, the averaged value of Cs in the
dynamic Smagorinsky model was close to the value of Cs in the
Smagorinsky model. The results obtained compare well with the
experiments. Niceno et al. (2008) employed the one-equation
SGS model of Davidson (1997) with an extra source term in the
transport equation (Peger and Becker, 2001) to represent the
effect of bubble-induced turbulence (BIT) and computed a square
cross-sectional bubble column. It was found that this approach
gives good predictions. However, the unresolved part of the turbulent kinetic energy reaches about 20% with this approach, which
might be too high for a traditional LES (Frhlich and von Terzi,
2008). Furthermore, an extensive discussion of the merits of LES
can be found in the work of Dhotre et al. (2013), providing a systematic evaluation of prior work on the modelling of turbulent
bubbly ows.
The present study employs the EulerEuler LES (EE-LES)
approach. The physical model and the simulation setup are given
in Sections 3 and 4 before the main results are discussed in
Section 5. In particular, the SGS turbulent kinetic energy will be
estimated for zero-equation SGS models to improve the prediction.
2. Experimental data
The simulations are carried out for a rectangular water/air
bubble column at ambient pressure (Akbar et al., 2012) and are
compared with the experimental data for two different gas supercial velocities at the inlet. A schematic sketch of the experimental
setup is shown in Fig. 1. Its width, depth and height are 240, 72 and
800 mm, respectively, and the water level is 700 mm. A distributor
plate containing 35 evenly spaced needles with an inner diameter
of 0.51 mm was positioned at the bottom of the experimental
column. Measurements using a laser Doppler velocimetry system,
an electrical conductivity probe and a high speed camera were

performed for two supercial velocities 3 mm/s (Case 1) and

13 mm/s (Case 2). Liquid velocities, void fraction and bubble velocities were measured along the line 500 mm above distributor plate
in the centre plane (y = 36 mm). More details are provided in the
cited reference (Akbar et al., 2012).
3. Physical modelling
3.1. EulerEuler approach
In this work the EulerEuler two-uid model is used. The conservation equations are discussed in detail in a number of books,
such as (Ishii and Hibiki, 2011), and a broad consensus on this
model has been reached. The governing equations of this approach
are the continuity and momentum equations for each phase
employed here, without mass transfer between the phases.

@
ai qi r  ai qi ui 0;
@t
@ai qi ui
r  ai qi ui ui rai li S i  ai rp ai qi g M i
@t
 rai si :

Here, the lower index i denotes the different phases, with a, q, l and
u the volume fraction, density, molecular viscosity and resolved
velocity, respectively, and S is the strain rate tensor. The index i
can be L to designate the liquid and G to designate the gas. The vector M represents the sum of all interfacial forces acting between the
phases such as drag force, lift force, wall lubrication and turbulent
dispersion force. The unresolved stress tensor s, and all interfacial
forces have to be modelled. The applied modelling is discussed
below.
Eqs. (1) and (2) are usually derived by ensemble averaging.
However, the same form of the equations is obtained if one performs ltering (volume averaging) of the governing equations
(Niceno et al., 2008). This is of practical importance for LES,
because it means that the same numerical tools developed for
ensemble averaged EulerEuler equations, can be used for LES.
The difference then only resides in the model term s.
3.2. Turbulence
3.2.1. Two-phase turbulence
In this study, turbulence is treated differently for the different
phases. Because of the low density of the gas, the turbulence in
the dispersed gas phase is of little relevance and is modelled with
a simple zero equation model here. It was found that this model
has nearly no inuence on the result. For the continuous, liquid
phase, LES was used.
3.2.2. LES for continuous liquid phase
The liquid velocity uL in (1) and (2) represents the resolved
velocity contribution. The corresponding unresolved contributions
are:

u0L uL  uL ;

with uL the true velocity of liquid. The index L is dropped in the

remainder of this paragraph for better readability. The SGS stress
tensor sij ui uj  ui uj , is modelled by the Smagorinsky model
(Smagorinsky, 1963):

1
3

saij sij  skk dij 2msgs Sij with msgs C s D2 jSj;

Fig. 1. Schematic representation of the experiment of Akbar et al. (2012). The
broken line in the gure shows the measurement position. The marker on that line
is the measurement point for the results presented in Figs. 9 and 10.

with saij the anisotropic (traceless) part of the SGS stress tensor sij,
and dij the Kronecker delta. The SGS viscosity, msgs, is a function of

T. Ma et al. / International Journal of Heat and Fluid Flow 56 (2015) 5159

the magnitude of the strain rate tensor, jSj

q
2Sij Sij , and the sub-

grid length scale is l = CsD. Here, the model constant was chosen
to be Cs = 0.15 based on previous experience of Akbar et al.
(2012), while the lter width D was determined by the grid size
p

3
according to D Vol, where Vol designates the volume of the computational cell. Close to the walls the turbulent viscosity is damped
using the formulation of Shur et al. (2008). The trace of the SGS
stress tensor skk in (4) is added to the ltered pressure p, resulting
in a modied pressure P p s3kk . For wall modelling, a blending is
performed between the viscous sublayer and the logarithmic law of
the wall depending on the mesh resolution (ANSYS, 2010).
3.2.3. Bubble induced turbulence
With the EulerEuler approach bubbles are not resolved. The
resolved part of the velocity eld in LES represents only the
shear-induced turbulence, which is assumed to be independent
of the relative motion of bubbles and liquid. The impact of the bubbles traveling through the liquid (possibly exhibiting bubble wake
instability, bubble oscillations, etc.) onto the uid turbulence has
to be modelled. In two equation RANS models, additional source
terms have been developed to describe bubble-induced turbulence. The approximation is provided by the assumption that all
energy lost by the bubbles due to drag is converted to turbulent
kinetic energy in the wake of the bubble. Detailed information
about the BIT models in RANS can be found in the recent review
of Rzehak and Krepper (2013). Such an approach is not suitable
for LES with a zero equation SGS model, because no transport
equation for the turbulent kinetic energy k is available. Here, we
use the common BIT model of Sato et al. (1981). In this model
the bubble inuence on liquid turbulence is included by an
additional extra term contributing to the SGS turbulent viscosity
so that
mol
bub
leff
lsgs
lbub
C B qL aG dB juG  uL j:
L
L lL
L lL ;

The constant CB is a model parameter equal to 0.6, and dB

represents the bubble diameter. In LES, lbub
is added directly in
L
the SGS model, without direct contribution to the total turbulent
kinetic energy, i.e.

a
ij

1
sij  skk dij
3

2msgs mBIT Sij ;

C D maxC D;Sphere ; C D;ellipse ;

53

where

C D;Sphere

24
1 0:1Re0:75 ;
Re

C D;ellipse

2 0:5
Eo :
3

10

Tomiyama et al. (1998) validated this correlation and found

good agreement except for high values of the Etvs number.
3.3.2. Lift force
In a shear ow a bubble experiences a force perpendicular to
the direction of ow. This effect generally is referred to as lift force
and described by the expression of Zun (1980):

F Lift C L qL aG uG  uL  rotuL :

11

For a spherical bubble the shear lift coefcient CL is positive so

that the lift force acts in the direction of decreasing liquid velocity,
i.e. in case of co-current pipe ow in the direction towards the pipe
wall. Experimental investigation by Tomiyama et al. (2002) and
numerical results by Schmidtke (2008) showed that the direction
of the lift force changes its sign if a substantial deformation of
the bubble occurs. From the observation of the trajectories of single
air bubbles rising in simple shear ow of a glycerol water solution
the following correlation for the lift coefcient was derived
(Tomiyama et al., 2002)

8
>
< min0:288 tanh0:121Re; f Eo? ; Eo? < 4
C L f Eo? ;
4 < Eo? < 10
>
:
0:27;
Eo? > 10;

12

with

f Eo? 0:00105Eo3?  0:0159Eo2?  0:0204Eo? 0:474:

13

This coefcient depends on the modied Etvs number Eo\

given by
2

Eo?

gqL  qG d?

14

since skk is not computed anyway.

where d\ is the maximum horizontal extension of the bubble. It is

calculated using the empirical correlation for the aspect ratio by
Wellek et al. (1966)

3.3. Interfacial forces

d? dB

In the Eulerian two-uid model the interaction of the bubbles

and the liquid phase is modelled through exchange terms between
the separate momentum conservation equations of the liquid and
the gas phase. They are still subject to discussion in the community
and vary between researchers. Details of such models including a
complete description of all interfacial transfer were published by
Rzehak and Krepper (2013).

The experimental conditions on which (12) is based, were limited to the range 5:5  log10 Mo  2:8; 1:39  Eo  5:74,
where Mo is the Morten number, and values of the Reynolds number based on bubble diameter and shear rate 0 6 Re 6 10. The
waterair system at normal conditions has Mo = 2.63e11 which
is beyond that range, but good results have nevertheless been
reported for this case, as shown by Lucas and Tomiyama (2011).
With the present parameters, the bubble size where the lift force
changes its direction is 5.8 mm.

3.3.1. Drag force

The drag force is a momentum exchange due to the slip velocity
between the gas phase and the liquid phase. The corresponding gas
phase momentum sink is dened as

F drag 

3
C D qL aG juG  uL juG  uL :
4dB

For the bubble regime investigated in the present study, the

drag coefcient CD mainly depends on the Reynolds number and
the Etvs number Eo. A correlation distinguishing different shape
regimes was suggested by Ishii and Zuber (1979) reading

q
3
1 0:163Eo0:757 :

15

3.3.3. Wall force

A bubble ascending close to a wall in an otherwise quiescent
liquid also experiences a lift force. This wall lift force, often simply
referred to as wall force, has the general form

F wall

2
C W qL ajuG  uL j2 b
y;
dB

16

y is the unit normal vector perpendicular to the wall pointwhere b

ing into the uid. The dimensionless wall force coefcient CW

54

T. Ma et al. / International Journal of Heat and Fluid Flow 56 (2015) 5159

depends on the distance to the wall y and is expected to be positive,

so that the bubble is driven away from the wall.
Based on the observation of single bubble trajectories in simple
shear ow of glycerol water solutions Tomiyama et al. (1995) and
later Hosokawa et al. (2002) developed the functional dependence

C W y f Eo

dB
2y

2
;

17

where in the limit of small Morton number (Hosokawa et al., 2002)

f Eo 0:0217Eo:

18

The experimental conditions on which (18) is based are

2:2  Eo  22 and 2:5  log10 Mo  6:0, which is still
different from the waterair system with Mo = 2.63e11 but a
recent comparison of this formula with other expressions proposed
in the literature (Rzehak and Krepper, 2013) has nonetheless
shown that good predictions can be obtained for vertical upward
pipe ow of air bubbles in water.
3.3.4. Turbulent dispersion force
The turbulent dispersion force is the result of the turbulent
uctuations of liquid velocity. In URANS simulations, this contribution has to be modelled, because only a very small part of the
turbulence is resolved. In LES, the resolved part of the turbulent
dispersion can be calculated explicitly. The unresolved part has
little inuence on bubble dispersion if the bubble size is in the
scale of the lter size (Niceno et al., 2008).
4. Simulation setup
4.1. Polydispersity and iMUSIG
The measured bubble size distributions at the inlet and
measurement plane are given in Fig. 2. Case 1 exhibits a relatively
narrow distribution, while the distribution of Case 2 is much wider.
Obviously, the bubble size distribution near the sparger and
500 mm above the sparger is almost the same for both cases.
Therefore, coalescence and break-up will be neglected. For
modelling the polydispersity, the inhomogeneous multiple size
group (iMUSIG) model as introduced by Krepper et al. (2008)
assigns different velocity groups to the bubble classes used in the
MUSIG model. Each velocity group has therefore its own velocity
eld. This is important, to describe effects like the bubble

size-dependent movement of the gas phase caused by the lift force.

In the present case, the bubble classes are chosen in such a way
that the bubble size distributions are split into two contributions
at the diameter where the lift force changes its sign, which is
dB = 5.8 mm. The resulting bubble classes for Case 2 can be found
in Table 1. In Case 1 bubbles are treated as monodisperse with
dB = 4.37 mm, because almost all bubbles have a positive lift coefcient, so that there is no need for considering different velocity
groups.
4.2. Numerical conditions
4.2.1. Grid Requirement
The rectangular bubble column was discretized with uniform
cubic cells of Dx = Dy = Dz = 4 mm, resulting in about 200,000 cells,
overall. The spatial resolution is in the order of the bubble size
(d = Dx/dB  1). Resolution requirements for EulerEuler methods
are still subject of debate in the literature. The approach was conceived to represent bubble swarms on very coarse grids, far coarser
than individual bubbles. The equations are derived by means of
volume averaging, and the averaging volume then has to be substantial larger than individual bubbles. Once the equations derived,
the problem to be solved is composed of a partial differential equation (PDE) and appropriate boundary conditions. This continuous
problem can be solved numerically on an arbitrarily ne grid, then
yielding a very smooth solution. Indeed, the numerical solution
converges to the exact solution of the EulerEuler PDE, so that
the discretization error is negligible. The modelling error persists,
so that caution is needed only when interpreting the result. The
fact that grid points are very close does not necessarily mean that
the uid motion is resolved with the same detail. The details have
just been removed from the equations and the quantity determined is an averaged quantity. This is the same as solving the
equations of the low-Reynolds ke model on a grid with step size
ner than the Kolmogorov scale which provides the almost exact
solution to the RANS problem, but not a DNS. See also the discussion in Geurts and Frhlich (2002), where the issue is developed in
detail for LES. In the present work, the grid size was chosen according to the experimental energy spectrum. Fig. 3 shows the energy
spectrum of the vertical velocity calculated at the measurement
point in Fig. 1 using a Fast Fourier Transformation (FFT). The
energy spectrum from Akbar et al. (2012a) was transformed from
frequency in time to wavenumber in space using Taylors hypothesis of frozen turbulence. This allows to insert the LES cut-off (lter
size = 4 mm) into the energy spectrum, as done in Fig. 3 with the
dotted line. For Case 2 the grid is ne enough and for Case 1 the
grid scale is located roughly at the beginning of the inertial subrange, which is coarser than what usually would be chosen.
However, as will be reported in Section 5.2 below, the vertical
velocity uctuations are in the order of the vertical mean ow, so
the condition w0 =w  1 for using Taylors hypothesis is actually
not fullled. Fig. 3, nevertheless, provides some hint concerning
the resolution.
4.2.2. Simulation details
The needles introducing the bubbles in the bottom plate of the
container are modelled by small 4 mm  4 mm surfaces at the
respective locations, each representing one needle. Here, the inlet

Table 1
Bubble classes employed in Case 2.

Bubble Class 1
Bubble Class 2
Fig. 2. Measured bubble size distribution at z = 0 and 500 mm (Akbar et al., 2012).

dB (mm)

a (%)

Eo\

CL

5.3
6.3

0.63
0.37

3
7.3

0.288
0.116

T. Ma et al. / International Journal of Heat and Fluid Flow 56 (2015) 5159

Fig. 3. Turbulent energy spectrum of uctuating vertical velocity.

velocity is imposed to equal the gas velocity, the volume fraction

set to 1 for the gas and 0 for the liquid. The gas velocity is obtained
by dividing the total gas ux measured in the experiment by the
number of needles, so that it is assumed to be equal at all needles.
The liquid velocity at the inlet is set to 0. For the initial conditions,
the velocities are set to 0 for both phases and the volume fraction
dened as: aL = 1 and aG = 0. Because the liquid is stagnant and all
cells of the grid are completely lled with liquid at the beginning.
During the entire calculation, a minimal volume fraction 1010 for
each phase is set for numerical robustness, so that all volume fractions are larger than 0 all over the computational domain. At the
walls, a no slip condition is applied for the continuous phase and
a free slip condition for the dispersed phase. At the top of the column a degassing boundary is imposed, which means a slip condition for the continuous phase and an outlet for the dispersed phase.
For the spatial discretization, a central difference scheme is
employed, and a second order backward Euler scheme is used in
time. The simulations were carried out using time steps
Dt = 0.01    0.015 s to satisfy CFL < 1. The results were averaged
over 250 s physical time.
5. Results
5.1. Instantaneous data
To provide a rst impression of the EE-LES simulation result,
Fig. 4 presents selected instantaneous data from Case 2 at an arbitrary instant and compares them to the same data from the
EE-URANS. The right graphs depict the corresponding void fraction,
showing a clear difference between the results for the two turbulence models. The computed streamlines of liquid velocity in
Fig. 4 illustrate that the LES resolves much more of the details of
the ow eld. With URANS, the transient details are not well
resolved. It can be seen that only some large-scale uctuations
are obtained, while smaller ones are damped due to the high turbulent viscosity. Similar differences of the liquid velocity eld with
LES and URANS are also described by Deen et al. (2001) and Dhotre
et al. (2008). This difference in the liquid velocity eld yields a
more homogenous gas volume fraction distribution in the URANS
simulation.
5.2. Time-averaged results
In this section, the results from EE-LES and EE-URANS are compared with previous work of Akbar et al. (2012), who used a
Lagrangian modelling for bubbles i.e. EL-LES. The interfacial force
models used in his reference are comparable to the settings

55

dened in Section 3.3, and the simulation was performed using

two way coupling.
In Figs. 5 and 6, long-time averaged vertical liquid velocity, gas
volume fraction and the uctuations of the vertical liquid velocity
are presented for both cases, together with experimental data of
Akbar et al. (2012) and EL-LES data from the same reference. All
simulations presented were run for 250 s. All proles were taken
along the measurement line from the wall to the centre at a height
of 500 mm, as represented by the broken line in Fig. 1.
Fig. 5 shows the vertical liquid velocity and gas volume fraction
for Case 1. The three predicted liquid velocity proles provide too
high velocities near the wall and somewhat too small velocities in
the centre. Undesired also is the oscillatory behaviour of the EE-LES
near the wall. This can be traced back to an uneven distribution of
the void fraction seen in the second picture of that gure. Indeed,
also the experimental dada exhibit an undulating shape. But the
amplitude is smaller and the region of occurrence somewhat further from the wall. So, apparently, this is a feature of this case
and while not captured quantitatively by the EE-LES, qualitatively
correct. The gas fraction obtained with the other two approaches
also exhibits a local maximum, but in the URANS case no and in
the EL-LES only small undulations of the velocity dada. The wall
peak in the void fraction prole is caused by the lift force modelling and is also described by Krepper et al. (2007), who used a
similar experimental facility. The gas volume fraction of about
1.2% in the centre is obtained in all the simulations and the experimental data. The results from Akbar et al. (2012) using EL-LES
exhibits small oscillations of the averaged gas volume fraction
and vertical liquid velocity in all results reported. The reason could
be a somewhat too small averaging time.
In Fig. 6, the results for high supercial gas velocity at the inlet
(Case 2) are presented. A clear change in the direction of the liquid
velocity can be seen at about 20 mm away from wall. This is a phenomenon caused by the liquid mass balance in the bubble column,
obtained in all three predictions with a quite good quantitative
agreement. In the near wall region the velocity data obtained with
EE-LES match the experimental data better than the other two simulations. The gas volume fraction prole in Fig. 6 is smoother than
in Case 1. The experimental data only exhibit a slight overshoot
between 30 and 60 mm from the wall. The simulation data show
the same tendency of creating a local maximum as observed with
Case 1, albeit much less pronounced. This peak caused by the lift
force modelling making the small bubbles go towards to the wall,
and migrate the big bubbles to the centre, results in a slight second
peak in the position about 40 mm away from the wall. The same
phenomenon appears also in the URANS simulation using the same
lift force modelling. The gas volume fraction results obtained with
EL-LES underpredict the void fraction in the wall region and generally exhibit a very jaggy prole, presumably due to lack of
averaging.
Figs. 7 and 8 show comparisons between experimental and
numerically predicted vertical liquid velocity uctuations for
Cases 1 and 2, respectively. With the EL-LES only the resolved
velocity uctuation is provided in Akbar et al. (2012), the unresolved part is neglected. In the EE-LES the total vertical velocity
uctuation is considered, which can be decomposed into the
resolved part w00 and the unresolved part w0 as mentioned in
Section 3.2. For the resolved part, w00 w00 w  hwi2 is computed
using the difference between the resolved vertical velocity and
the time average of the resolved vertical velocity. For the unresolved part, as can be seen in (4), only the anisotropic part of the
SGS stress tensor is considered in the Smagorinsky model, so that
the information about the isotropic part skk = 2k is lost. Dhotre
et al. (2013) performed a review about application of LES to dispersed bubbly ows, observing that in all papers the collected

56

T. Ma et al. / International Journal of Heat and Fluid Flow 56 (2015) 5159

Fig. 4. Instantaneous data from Case 2 at t = 60 s. The left two graphs show instantaneous streamlines of the liquid coloured with the instantaneous absolute value of the
liquid velocity. The right two graphs show the instantaneous void fraction in the centre plane.

Fig. 5. Comparison of vertical liquid velocity (top) and gas volume fraction
(bottom) for Case 1. Experimental data and EL-LES are from Akbar et al. (2012).
EE-URANS are from Ziegenhein et al. (2015).

SGS kinetic energy ksgs was neglected when zero equation models
were used in LES. Only Niceno et al. (2008) demonstrated the
applicability of a one-equation model for ksgs, so that ksgs could
be explicitly calculated. Here, a method for estimating ksgs will be
introduced based on the SGS dissipation esgs as proposed by
Menter (2013) for single-phase ows:

s
ksgs

msgs esgs
Cl

esgs msgs jSj2 :

19

Fig. 6. Comparison of vertical liquid velocity (top) and gas volume fraction
(bottom) for Case 2. Experimental data and EL-LES are from Akbar et al. (2012).
EE-URANS are from Ziegenhein et al. (2015).

As can be seen in Fig. 7, for Case 1 both LES with EulerEuler

and EulerLagrangian obtain much lower values for the vertical
velocity uctuations than the experimental data. The unresolved
part from EE-LES is quite small and has nearly no inuence on
the evaluation. The underprediction can be related to the limitation of LES for this kind of ow, since large-scale turbulence is
not present with a homogeneous distributed low gas inlet at the
column bottom. With the EE-URANS the experimental data can
be reproduced very well. The total and the unresolved vertical

T. Ma et al. / International Journal of Heat and Fluid Flow 56 (2015) 5159

57

w0 w0 v 0 v 0 u0 u0 23 k. Similar to Case 1, the unresolved part in

Case 2 contributes almost 90% of the total uctuation (Fig. 8). So,
the lateral velocity uctuations in the other two directions in both
cases would have nearly the same proles like the vertical direction. Therefore, it might be a problem at this point that only one
velocity component is evaluated for the EE-URANS approach. The
previous work of Dhotre et al. (2008) also mentions the limit of
URANS models to predict the liquid velocity uctuation in one
direction. The comparison between the total turbulent kinetic
energy and experimental data could give a better agreement.
Unfortunately, the experimental data consists only of one component of the velocity uctuations.
5.3. Energy spectra
Fig. 7. Comparison of vertical liquid velocity uctuation for Case 1. Experimental
data and EL-LES are from Akbar et al. (2012). EE-URANS are from Ziegenhein et al.
(2015).

uctuation proles are nearly the same, so that the resolved uctuations using the EE-URANS method are zero for Case 1. Hence, all
the velocity uctuations come from the used two equation turbulence model with the BIT model. This is in line with the concept of
a RANS model and recovered for URANS as well if instabilities in
the ow are small. Summarizing, this might be a hint that the large
uctuations are in general very low in Case 1 and the
bubble-induced turbulence is dominant. Because the bubbly ow
subgrid models used for the EE-LES method include only dissipation terms, as mentioned in Section 3.2, the total velocity uctuations with this method might be underpredicted.
Fig. 8 shows the same velocity uctuations for Case 2. In this
case, higher large-scale turbulence is expected, because of the
higher supercial velocity. The result of the EE-LES has a better
quantitative agreement with the measured data than the other
two simulations shown in Fig. 8. Especially the measured peak is
only reproduced by the EE-LES and located at the change of sign
of the vertical liquid velocity, as mentioned in Fig. 6. The unresolved velocity uctuations obtained by using the EE-LES method
in Fig. 8 amount to about 10% of the total velocity uctuations.
The result with EL-LES underpredicts the uctuation in this case,
which might be caused by neglecting the unresolved part. The
trend of a peak close to the wall can also be found in the resolved
part of EL-LES.
The results from EE-URANS t the measured prole well in in
Case 1 (Fig. 7), while the agreement is less satisfactory in Case 2
(Fig. 8). URANS with a two equation turbulence model for prediction of velocity uctuation in one direction might be critical,
because the isotropic assumption of turbulence leads to

Fig. 8. Comparison of vertical liquid velocity uctuation for Case 2. Experimental

data and EL-LES are from Akbar et al. (2012). EE-URANS are from Ziegenhein et al.
(2015).

Fig. 9 shows a 200 s time history plot of the resolved vertical

liquid velocity with 20,000 sample points obtained from the
EE-LES of Case 2 at the measurement point shown in Fig. 1.
The energy spectrum obtained with the data extracted from
Fig. 9 is shown in Fig. 10. The velocity signal was transformed using
the Welch method (Welch, 1967) with 10 non-overlapping windows. Each window contains 2000 sample points, and a Hanning
window function was used. As can be seen, the turbulent energy
spectrum with the EE-LES approach exhibits a broad range of frequencies, with a slope steeper than the 8/3 power law in the inertial subrange, which is still in discussion for this type of ow
(Mercado et al., 2010). Here, the steep slope in the simulation is
mainly caused by the BIT model of Sato et al. (1981), which
induced an additional enhancement of the eddy viscosity in (5).
This is a similar effect as increasing the Smagorinsky constant Cs
in the LES, which is discussed by Frhlich (2006). A similar energy
spectrum based on EE-LES results using the Sato model was
obtained by Dhotre et al. (2008). In his work the slope was partly
even over 10/3 in the inertial subrange.
Previous experimental studies have actually attributed the
more dissipative spectrum to the presence of the bubbles suggesting a mechanism of eddy disintegration (Lance and Bataille, 1991).
But in the experimental energy spectrum for Case 2 of Akbar et al.
(2012) shown in Fig. 10, such a dissipative slope does not appear.
The difference between the EE-LES energy spectrum and the experimental energy spectrum could be caused by many different reasons. A determinant reason might be the limitations caused by
the EulerEuler approach. It might not be able to reproduce a similar power law as the experimental one. Since in the EulerEuler
approach, bubbles are not resolved, it is impossible to catch the frequency information in the bubble wake like in the experiment.
Furthermore, the time scale of the bubble wake is also shown in
Fig. 10. The estimation of the angular frequency xB for the bubble
wake is introduced as: xB = 2pfB 2pUB/dB 314 rad/s, with fB
being the frequency of the bubble wake, UB 0.3 m/s the bubble
velocity, and dB 6 mm as the averaged bubble diameter for

Fig. 9. Time history of the liquid velocity obtained with EE-LES at the centre of
measurement line for Case 2.

58

T. Ma et al. / International Journal of Heat and Fluid Flow 56 (2015) 5159

Acknowledgements
This work was carried out in the frame of a current research
project funded by the German Federal Ministry of Economic
Affairs and Energy, project number 150 1411. Partial funding of
the Helmholtz-Alliance LIMTECH is acknowledged. Professor Akio
Tomiyama from Kobe University is greatfully acknowledged for
providing experimental spectra in electronic form.
References

Fig. 10. Temporal energy spectrum of vertical liquid velocity scaled with an
arbitrary reference value. The red dotted line is the estimated frequency of the
bubble wake. (For interpretation of the references to color in this gure legend, the
reader is referred to the web version of this article.)

Case 2. The resolved and reliable angular frequencies in the spectrum of the EE-LES are far away from the requirement to know
bubble wake information (e.g. BIT). Also in the experiment, this
requirement is not reached. Nevertheless, the low frequency contribution can be compared.
6. Conclusions
EE-LES have been carried out for the rectangular bubble column
and compared with the experimental data from Akbar et al. (2012),
previous numerical work with EL-LES (Akbar et al., 2012) and
EE-URANS (Ziegenhein et al., 2015). The bubble-induced turbulence was taken into account in the EE-LES using the Sato model.
The results obtained with the EE-LES approach reproduce the
measured gas volume fraction and liquid velocity proles in the
same way as EL-LES (Akbar et al., 2012) and EE-URANS
(Ziegenhein et al., 2015). Large improvement can be achieved with
the EE-LES method for the turbulence prediction in the case with a
higher gas supercial velocity. A near wall peak in the velocity uctuation can be reproduced. Here, the approach is successful, as the
largest and most energetic scales of motion (comparable in size to
the whole domain) are much stronger than the BIT, which are not
resolved with this approach. For lower gas supercial velocity, LES
may not represent the best option for turbulence prediction in this
case, since large-scale turbulence is not present.
The criterion for a suitable cut-off is discussed. It could be
obtained from the experimental energy spectrum in the wavenumber space, the chosen cut-off for the high supercial velocity is ne
enough. However, for the simulation with a low gas supercial
velocity the cut-off might be too coarse. That might cause the
underprediction of the velocity uctuation for the case with a
low gas supercial velocity in both EE-LES and EL-LES. However,
the uid motion in small scales for this case with the low supercial gas velocity is dominated by the bubble-induced turbulence
and, consequently, an improvement for this case cannot be
achieved with the present EE-LES approach due to the missing of
physical modelling for such small scale uctuations.
A method for estimating the SGS kinetic energy is introduced
and investigated. Considering this unresolved part may improve
the prediction of the velocity uctuation. A similar power law like
experimental energy spectrum might not be reproducible with
EE-LES, since the bubbles are not resolved in EulerEuler approach
and the frequency information related to the bubble wake is lost.
Further tests applying the proposed combinations of interfacial
force models and LES of turbulence to different bubbly ows are
desired and will be conducted in the future.

Akbar, M.H.M., Hayashi, K., Hosokawa, S., Tomiyama, A., 2012. Bubble tracking
simulation of bubble-induced pseudo turbulence. Multiphase Sci. Technol. 24,
197222.
Akbar, M.H.M., Hayashi, K., Hosokawa, S., Tomiyama, A., 2012a. Bubble tracking
simulation of bubble-induced pseudo turbulence. In: 6th Japanese-European
Two-Phase Flow Group Meeting.
ANSYS, Fluent R13 Theory guide, 2010.
Davidson, L., 1997. Large eddy simulations: a note on derivation of the equations for
the subgrid turbulent kinetic energies. Technical Report No. 97/12, 980904,
Chalmers University of Technology, Gothenburg, Sweden.
Deen, N.G., Solberg, T., Hjertager, B.H., 2001. Large eddy simulation of the gas-liquid
ow in a square cross-sectioned bubble column. Chem. Eng. Sci. 56 (2122),
63416349.
Dhotre, M.T., Niceno, B., Smith, B.L., 2008. Large eddy simulation of a bubble column
using dynamic sub-grid scale model. Chem. Eng. J. 136 (23), 337348.
Dhotre, M.T., Deen, N.G., Niceno, B., Khan, Z., Joshi, J.B., 2013. Large eddy simulation
for dispersed bubbly ows: a review. Int. J. Chem. Eng. 2013.
Frhlich, J., 2006. Large eddy simulation turbulenter Strmungen, Teubner-Verlag
(ISBN: 3-8351-0104-8).
Frhlich, J., von Terzi, D., 2008. Hybrid LES/RANS methods for the simulation of
turbulent ows. Prog. Aerosp. Sci. 44, 349377.
Germano, M., Piomelli, U., Moin, P., Cabot, W.H., 1991. A dynamic subgrid-scale
eddy viscosity model. Phys. Fluids A 3, 17601765.
Geurts, B.J., Frhlich, J., 2002. A framework for predicting accuracy limitations in
large-eddy simulation. Phys. Fluids 14, L41.
Hnsch, S., Lucas, D., Krepper, E., Hhne, T., 2012. A multi-eld two-uid concept for
transitions between different scales of interfacial structures. Int. J. Multiph.
Flow 47, 171182.
Hosokawa, S., Tomiyama, A., Misaki, S., Hamada, T., 2002. Lateral migration of single
bubbles due to the presence of wall. In: Proc. ASME Joint U.S. European Fluids
Engineering Division Conference, FEDSM2002, Montreal, Canada, p. 855.
Ishii, M., Hibiki, T., 2011. Thermo-uid Dynamics of Two-Phase Flow, second ed.
Springer.
Ishii, M., Zuber, N., 1979. Drag coefcient and relative velocity in bubbly, droplet or
particulate ows. AlChE J. 25, 843.
Krepper, E., Vanga, B.N.R., Zaruba, A., Prasser, H.-M., Lopez de Bertodano, M.A., 2007.
Experimental and numerical studies of void fraction distribution in rectangular
bubble columns. Nucl. Eng. Des. 237, 399.
Krepper, E., Lucas, D., Frank, T., Prasser, H.M., Zwart, P., 2008. The inhomogeneous
MUSIG model for the simulation of polydispersed ows. Nucl. Eng. Des. 238,
16901702.
Lance, M., Bataille, J., 1991. Turbulence in the liquid phase of a uniform bubbly airwater ow. J. Fluid Mech. 222, 95118.
Lucas, D., Tomiyama, A., 2011. On the role of the lateral lift force in poly-dispersed
bubbly ows. Int. J. Multiph. Flow 37, 1178.
Ma, T., Lucas, D., Ziegenhein, T., Frhlich, J., Deen, N.G., 2015. Scale-adaptive
simulation of a square cross-sectional bubble column. Chem. Eng. Sci. 131,
101108.
Menter, F.R., 2013. Private communication.
Mercado, J.M., Gomez, D.C., Van Gils, D., Sun, C., Lohse, D., 2010. On bubble clustering
and energy spectra in pseudo-turbulence. J. Fluid Mech. 650, 287306.
Niceno, B., Dhotre, M.T., Deen, N.G., 2008. One-equation subgrid scale (SGS)
modelling for EulerEuler large eddy simulation (EELES) of dispersed bubbly
ow. Chem. Eng. Sci. 63 (15), 39233931.
Peger, D., Becker, S., 2001. Modelling and simulation of the dynamic ow
behaviour in a bubble column. Chem. Eng. Sci. 56 (4), 17371747.
Rzehak, R., Krepper, E., 2013. CFD modeling of bubble-induced turbulence. Int. J.
Multiph. Flow 55, 138155.
Sato, Y., Sadatomi, M., Sekoguchi, K., 1981. Momentum and heat transfer in twophase bubble owI. Theory. Int. J. Multiph. Flow 7 (2), 167177.
Schmidtke, M., 2008. Investigation of the Dynamics of Fluid Particles Using the
Volume of Fluid Method. PHD-Thesis, Universitt Paderborn.
Shur, M.L., Spalart, P.R., Strelets, M.K., Travin, A.K., 2008. A hybrid RANS-LES
approach with delayed-DES and wall-modeled LES capabilities. Int. J. Heat Fluid
Flow 29, 16381649.
Smagorinsky, J., 1963. General circulation experiments with the primitive
equations. Mon. Weather Rev. 91 (3), 99165.
Tomiyama, A., Sou, A., Zun, I., Kanami, N., Sakaguchi, T., 1995. Effects of Etvs
number and dimensionless liquid volumetric ux on lateral motion of a bubble
in a laminar duct ow. Adv. Multiphase Flow.
Tomiyama, A., Kataoka, I., Zun, I., Sakaguchi, T., 1998. Drag coefcients of single
bubbles under normal and micro gravity conditions. JSME Int. J. B 41, 472.

T. Ma et al. / International Journal of Heat and Fluid Flow 56 (2015) 5159

Tomiyama, A., Tamai, H., Zun, I., Hosokawa, S., 2002. Transverse migration of single
bubbles in simple shear ows. Chem. Eng. Sci. 57, 1849.
Welch, P.D., 1967. The use of fast Fourier transform for the estimation of power
spectra: a method based on time averaging over short, modied periodograms.
IEEE Trans. Audio Electroacoust. AU-15, 7073.
Wellek, R.M., Agrawal, A.K., Skelland, A.H.P., 1966. Shape of liquid drops moving in
liquid media. AlChE J. 12, 854.

59

Zhang, D., Deen, N.G., Kuipers, J.A.M., 2006. Numerical simulation of the dynamic
ow behavior in a bubble column: a study of closures for turbulence and
interface forces. Chem. Eng. Sci. 61, 75937608.
Ziegenhein, T., Rzehak, R., Lucas, D., 2015. Transient simulation for large scale ow
in bubble columns. Chem. Eng. Sci. 122, 113.
Zun, I., 1980. The transverse migration of bubbles inuenced by walls in vertical
bubbly ow. Int. J. Multiph. Flow 6, 583588.