Numbers in parentheses indicate the pages on which the authors contribution begin.
A. T. ANDREWS IV, Department of Chemical Engineering, Princeton University,
Princeton, NJ 08544, USA (65)
A. G. DIXON, Department of Chemical Engineering, Worcester Polytechnic
Institute, Worcester, MA 01609, USA (307)
L.S. FAN, Department of Chemical and Biomolecular Engineering, The Ohio
State University, 140 West 19th Avenue, Columbus, OH 43210, USA (1)
R. O. Fox, Herbert L. Stiles Professor of Chemical Engineering, Iowa State
University, 3162 Sweeney Hall, Ames, IA 500112230, USA (231)
Currently on sabbatical at:
Swiss Federal Institute of Technology Zurich, ETHZ Institut fur Chemie
und Bioingenieurwissenschaften ETHHonggerberg/HCI H 109 (Gruppe
Morbidelli), CH8093 Zurich, Switzerland
Y. GE, Department of Chemical and Biomolecular Engineering, The Ohio State
University, 140 West 19th Avenue, Columbus, OH 43210, USA (1)
J. A. M. KUIPERS, University of Twente, Faculty of Science & Technology, PO
Box 217, NL  7500 AE Enschede, The Netherlands (65)
M. NIJEMEISLAND, Johnson Matthey Catalysts, Billingham, UK (307)
E. H. STITT, Johnson Matthey Catalysts, Billingham, UK (307)
S. SUNDARESAN, Department of Chemical Engineering, Princeton University,
Princeton, NJ 08544, USA (65)
H. E. A. VAN DEN AKKER, Delft University of Technology, Molenwindsingel 50,
NL 4105 HK Culemborg, The Netherlands (151)
M. A. VAN DER HOEF, Department of Science and Technology, University of
Twente, PO 217, NL  7500 AE Enschede, The Netherlands (65)
M. VAN SINT ANNALAND, Department of Science and Technology, University of
Twente, PO 217, NL  7500 AE Enschede, The Netherlands (65)
M. YE, Department of Science and Technology, University of Twente, PO 217,
NL  7500 AE Enschede, The Netherlands (65)
ix
PREFACE
This issue attempts to give a feeling of the stateoftheart of the application
of computational uid dynamics (CFD) in chemical engineering. It is, however,
not limited to a snapshot but is aimed at providing a perspective: how did we
arrive at the present status and where do we go from here? To do so, contri
butions from ve complementary contributions are brought together. From the
definition of CFD as the ensemble of all computational approaches that solve
for the spatial distribution of the velocity, concentration, and temperature
elds recalled by Fox, it is clear that a selection had to be made as to the topics
covered. In the wake of volume 30 on Multiscale Analysis the present volume
is organized from small to large: from bubbles and droplets in the rst
contribution, to a xed catalyst bed in the last one. The application of direct
numerical simulations (DNS) clearly is still limited to the small scale. Today
subgridscale (SGS) models are required to cover the full spectrum.
The reader will be confronted with some redundancy but this allows each
contribution to stand on its own. Also, a good balance is maintained between
the style of a tutorial and that of a research paper. Those who will read the
complete volume will realize that opinions can vary from looking at CFD as an
alternative for experimentation to emphasizing the need of experimental val
idation. Some contributions are entirely limited to velocity and temperature
elds. Others, on the contrary, emphasize the difculties associated with the
combination of transport and reaction. The latter can introduce stiffness even
for laminar ow. Averaging (e.g. Reynoldsaveraged NavierStokes, RANS) or
ltering (e.g. large eddy simulations, LES), performed to model velocity elds,
does not alleviate this difculty. Clearly, this is still quite a challenge.
The contribution from the Ohio State University by Ge and Fan is dealing
with the simulation of gasliquid bubble columns and gasliquidsolid uidized
beds. A scientist of a major engineering company told me a few years ago that
when he wanted to know how serious an academic group was about CFD, he
would ask whether they could simulate bubble columns. He would only engage
into further conversation if the answer was negative. The group from Columbus
is wise enough to focus on a single air bubble rising in water, and bubble
formation from a single nozzle. In a second part the hydrodynamics and heat
transfer phenomena of a liquid droplet in motion and during the impact process
with a hot at surface, as well as with a particle are studied. The applied
numerical techniques, such as the level set and immersed boundary method, are
outlined and important contributions are highlighted. Next, detailed imple
mentations for particular problems are presented. Finally, numerous simulation
results are shown and compared with experimental data.
xi
The second contribution addresses the different levels of modeling that are
required in order to cover the full spectrum of length scales that are important
for industrial applications. It is a joint paper from Twente and Princeton Uni
versity and claims to put Emphasis on technical details. The latter is a too
modest description of what is really offered to the reader. The recent devel
opments in two leading research groups on the modeling of gasuidized beds
are presented. The holy grail for those interested in the design of industrial units
being the closure of the model equations in general and SGS modeling in par
ticular. The latest developments of both the ltering approach pursued at
Princeton University by Sundaresan and coworkers and the discrete bubble
model developed in Twente by the team of Kuipers are presented. The authors
realize fully that there is still a long way to go, as evidenced by their last
sentence: Finally, the adapted model should be augmented with a thermal
energy balance, and associated closures for the thermophysical properties, to
study heat transport in large scale uidized beds, such as FCCregenerators and
PE and PP gasphase polymerization reactors. This is even more so because
inclusion of reaction kinetics remains beyond the scope of the contribution!
Chemical reactions come into the picture in the context of stirred turbulent
vessels in Chapter 3. Van den Akker from Delft strongly emphasizes the po
tential of LES and DNS for reproducing not only the hydrodynamics of tur
bulent stirred vessels but also for providing a basis for simulating a wide variety
of physical and chemical processes in this equipment. The author advocates the
use of the latticeBoltzmann (LB) technique to this purpose. Van den Akker
certainly belongs to those who believe that one can and should be much more
positive about the merits of CFD so far and about the term at which CFD will
replace and improve existing mixing correlations. To quote him: It may be
easier to measure the local and transient details of the turbulent ows in stirred
vessels and the spatial distributions in e.g. mixing rates and bubble, drop and
crystal sizes computationally than by means of experimental techniques! When
it comes to the design of chemical reactors the authors admit that CFD is
certainly not a panacea. Scaleup of many chemical reactors, in particular the
multiphase types, is still surrounded by a fame of mystery indeed.
The importance of chemicalreaction kinetics and the interaction of the latter
with transport phenomena is the central theme of the contribution of Fox from
Iowa State University. The chapter combines the clarity of a tutorial with the
presentation of very recent results. Starting from simple chemistry and single
phase ow the reader is lead towards complex chemistry and twophase ow.
The issue of SGS modeling discussed already in Chapter 2 is now discussed with
respect to the concentration elds. A detailed presentation of the joint Prob
ability Density Function (PDF) method is given. The latter allows to account
for the interaction between chemistry and physics. Results on impinging jet
reactors are shown. When dealing with particulate systems a particle size dis
tribution (PSD) and corresponding population balance equations are intro
PREFACE xii
duced. The author emphasizes that a balance between the degree of detail or
complexity of the chemistry and that of the physics should be maintained.
The last contribution comes from Dixon (Worcester Polytechnic Institute),
and Nijemeisland and Stitt (Johnson Matthey). The subject is another classic of
reactor engineering: the catalytic xedbed reactor. Heat transfer issues on both
reactor scale and catalyst pellet scale are addressed. Steam reforming is used as a
typical example of a strongly endothermic reaction requiring highheat uxes
through the reactor walls. The presence of the tube wall causes changes in bed
structure, ow patterns, transport rates and the amount of catalyst per unit
volume, and is usually the location of the limiting heattransfer resistance.
Special attention is given to the modeling of the structure of a packed bed.
The importance of wall functions, to be applied not only at the reactor wall but
also at the external pellet surface, is stressed. The authors show ample results of
their own work without neglecting the contributions of others. At the end of this
chapter the reader will be convinced of the importance of the local nonuni
formities in the temperature eld not only within a catalyst pellet but also from
one pellet to the other.
Let me conclude by thanking the authors for their willingness to contribute,
despite health problems for some of them, and for their exibility with respect to
timing.
Guy B. Marin
Ghent, Belgium
April 2006
PREFACE xiii
3D DIRECT NUMERICAL SIMULATION OF GASLIQUID
AND GASLIQUIDSOLID FLOW SYSTEMS USING THE
LEVELSET AND IMMERSEDBOUNDARY METHODS
Yang Ge and LiangShih Fan
and O
+
. The
value of the levelset function is negative in the O
0; x 2 G
40; x 2 O
8
>
<
>
:
(2)
The evolution of f in a ow eld is given by the socalled weakform equation:
@f
@t
V rf 0 (3)
where V is the velocity of uid, and is given by
V
V
; x 2 O
; x 2 G
V
; x 2 O
8
>
<
>
:
(4)
For gasliquid bubble ow, V
and V
+
are the gas and liquid velocities, respec
tively, and the zerolevel set of f marks the bubble interface, which moves with
time. For gasdroplets ows, on the other hand, V
and V
+
represent the
+

+

FIG. 1. The level sets of distance function for a smooth interface over a Cartesian grid.
SIMULATION OF GAS LIQUID AND GAS LIQUID SOLID FLOW SYSTEMS 7
velocity of the liquid and gas phases, respectively, and the zerolevel set of f
denes the droplet surface (Ge and Fan, 2005).
To compute the motion of two immiscible and incompressible uids such as a
gasliquid bubble column and gasdroplets ow, the uidvelocity distributions
outside and inside the interface can be obtained by solving the incompressible
NavierStokes equation using levelset methods as given by Sussman et al.
(1994):
@r
@t
r rV 0 (5)
@rV
@t
r rVV rp r s rg F
s
(6)
where F
s
is the surface tension force which is calculated by (Brackbill et al.,
1992):
F
s
skfdfrf (7)
k(f) is the curvature which can be estimated as r (rf/rf). A smooth d
function is dened as (Sussman et al., 1998; Sussman and Fatemi, 1999):
d
b
f
dH
b
f
df
1
2
1 cospf=b=b; f
ob
0; otherwise
(
(8)
where H
b
(f) follows the Heaviside formulation (Sussman et al., 1998; Sussman
and Fatemi, 1999) given by
H
b
f
1 f4b
0 fob
1
2
1
f
b
1
p
sinpf=b otherwise
8
>
<
>
:
(9)
The surfacetension force F
s
in Eq. (7) is smoothed and distributed into the
thickness of the interface. In order to circumvent numerical instability, the uid
properties such as density and viscosity in the interface region are determined
with a continuous transition:
rf r
H
b
f (10)
mf m
H
b
f (11)
Since the values for r(f), m(f), and the surfacetension force could be distorted
if the variation of rf along the interface is very large, the thickness of the
interface needs to be maintained uniformly, i.e. rf
(12)
dx; 0 fx (13)
until
rd j j 1 OD
2
(14)
where the sin function is dened as
sinf
1; fo0
0; f 0
1; f40
8
>
<
>
:
(15)
In Eq. (12), t is an articial time that has the unit of distance. The solutions
for Eq. (12) are signed distances and only those within a thickness of 35 grid
sizes from the interface are of interest (Sussman et al., 1994, 1998; Sussman and
Fatemi, 1999). Equation (12) needs to be integrated for 35 time steps using a
time step Dt 0.5D.
B. IMMERSED BOUNDARY METHOD
The IBM was originally proposed by Peskin (1977) to model the blood ow
through heart valves. Since then, this method has been extensively modied and
extended to simulate various uid ows in a complex geometrical conguration
using a xed Cartesian mesh (Unverdi and Tryggvason, 1992a,b; Udaykumar,
et al., 1997; Ye et al., 1999; Fadlun et al., 2000; Lai and Peskin, 2000; Kim et al.,
2001). In the IBM, the presence of the solid object in a uid eld is represented
by a virtualbody force eld, which is applied on the computational grid in the
vicinity of the solidow interface through a Dirac delta function (Lai and
SIMULATION OF GAS LIQUID AND GAS LIQUID SOLID FLOW SYSTEMS 9
Peskin, 2000). Various schemes have been proposed to calculate the virtual force
density in the literature. Goldstein et al. (1993) developed a virtual boundary
formulation to simulate the startup ow over a cylinder. In their formation, the
virtual force eld is calculated in a feedback manner in order to satisfy the
boundary condition at the solid surface. MohdYusof (1997) developed an
alternative direct forcing scheme to evaluate the virtual force based on the NS
equation at discrete times. Fadlum et al. (2000) extended the direct forcing
scheme of MohdYusof (1997) to a 3D nitedifference method. Instead of
evaluating and applying the virtual force, the velocity at the rst grid point
outside the solid boundary is estimated through a linear interpolation of the
moving velocity of the boundary and the velocity at the second external grid
point. Conceptually, this velocity interpolation scheme is equivalent to applying
the momentum force inside the ow eld (Kim et al., 2001). This scheme is more
efcient in 3D because it has no adjustable constant and has no extra restric
tion on the scale of the time step, which is required in the feedbackforcing
scheme. Kim et al. (2001) simulated the ow over complex geometry in a nite
volume approach with staggered meshes. The momentum force and mass source
were applied on the immersed boundary to satisfy the noslip boundary con
dition and the ow continuity.
The basic idea of the IBM is that the presence of the solid boundary (xed or
moving) in a uid can be represented by a virtual body force eld
~
F
p
applied on
the computational grid at the vicinity of solidow interface. Thus, the Navier
Stokes equation for this ow system in the Eulerian frame can be given by
@rV
@t
r rVV rp r s rg
~
F
p
(16)
It is noted that the virtual body force
~
F
p
depends not only on the unsteady
uid velocity, but also on the velocity and location of the particle surface, which
is also a function of time. There are several ways to specify this boundary force,
such as the feedback forcing scheme (Goldstein et al., 1993) and direct forcing
scheme (Fadlun et al., 2000). In 3D simulation, the direct forcing scheme
can give higher stability and efciency of calculation. In this scheme, the disc
retized momentum equation for the computational volume on the boundary is
given as
V
t1
V
t
DtRHS
t
F
t
p
(17)
where RHS refers to all the terms in the righthand side of Eq. (16) except the
virtual body force
~
F
p
. The virtual body force F
t
p
is used to maintain the uid
velocity to be equal to the particle velocity at the particle surface (i.e., noslip
boundary condition), which is
V
t1
V
p
t (18)
YANG GE AND LIANGSHIH FAN 10
where V
p
is the particle velocity. Thus, the discrete virtual force can be dened as
F
t
p
V
p
V
t
=Dt RHS
t
(19)
Since the computational grids are generally not coincident with the location of
the particle surface, a velocity interpolation procedure needs to be carried out in
order to calculate the boundary force and apply this force to the control volumes
close to the immersed particle surface (Fadlun et al., 2000).
Other than the virtual momentum force
~
F
p
, a virtual mass source/sink should
also be applied to the particle surface to satisfy the continuity for the control
volume containing the particle surface or the particle (Kim et al., 2001). The
mass source can be calculated by
q
t
1
DV
X
i
a
i
~
V
t
i
~n
i
Ds
i
(20)
where DV is the volume of the computation cell (control volume) and Ds
i
the
surface area of surface i of this cell. For a 3D case, i 1, 2, y, 6. ~n
i
is
the normal vector of each face of the cell.
~
V
t
i
the uid velocity at each face of the
cell. a
i
the ag to indicate whether the virtual body force is applied to face i of
the cell or not. a
i
1 when the force is applied, otherwise it is zero. Therefore,
the continuity equation of the incompressible uid can be written as (Kim et al.,
2001):
r
~
V q (21)
III. System1: Flow Dynamics of GasLiquidSolid Fluidized Beds
The ows in a gasliquidsolid uidized bed or a gasliquid bubble column
are represented by two regimes, the homogeneous and the heterogeneous. In the
homogenous regime, the coalescence of bubbles does not occur and there is little
variation of bubble sizes. However, this is not the case in the heterogeneous
regime. The ow structure in the heterogeneous regime is complex due to sub
stantial coalescence and breakup of bubbles. Both the EE and the EL methods
have proven to be more effective in modeling the homogenous regime than the
heterogeneous regime of gasliquid ow. In the simulation of the heterogeneous
regime of gasliquid ows using either the EE or the EL method, the challenge
lies in the establishment of an accurate closure relationship for the interphase
momentum exchange. The interphase momentum exchange is induced through
the drag force that liquid exerts on the bubble surface, the virtual mass force due
to the bubble and liquid inertial motion, and the lift force caused by the shear
ows around the bubbles. In gasliquid bubble columns and gasliquidsolid
SIMULATION OF GAS LIQUID AND GAS LIQUID SOLID FLOW SYSTEMS 11
uidized systems, the interstitial forces under the bubble coalescence and
breakup conditions are not well established. A computational model based on
the levelset methods given below provides some information on the much
needed closure relationship of the interphase momentum exchange noted above.
A. NUMERICAL PROCEDURE FOR SOLVING THE GAS LIQUID INTERFACE
The levelset technique described in Section II.A is employed to capture the
motion of 3D gasliquid interfaces. The numerical procedures for solving the
gasliquid interface include nding the solution for the timedependent Eqs. (3),
(5), and (6). Given f
n
and V
n
dened at cell centers at one time instant t
n
, f
n+1
,
and V
n+1
can be solved over a time increment at a new time instant
t
n+1
t
n
+Dt following the procedures given below:
Solve Eqs. (5) and (6) to obtain the velocity distribution in the ow eld V
n+1
using the ArbitraryLagrangianEulerian (ALE) scheme (Kashiwa et al., 1994).
Solve Eq. (3) to obtain f
n+1
using the secondorder TVDRungeKutta
method presented as follows:
f
n1
f
n
Dtf
tn
(22)
f
n1
f
n
Dt
2
f
tn1
f
tn
(23)
where f
tn
V
n
Df
n
and the time steps are the same as that used in calculating
V
n+1
, which is determined by restrictions due to the CourantFriedrichsLevy
(CFL) condition, gravity, viscosity, and surface tension.
Solve Eq. (12) to perform the redistancing.
Although, in principle, Eq. (12) would not alter the location of the zerolevel
set of f, in practice, with numerical computation it may not be true. A redis
tance operation is needed to maintain the volume conservation. Therefore, Eq.
(12) is modied to (Sussman et al., 1998):
@d
@t
sinf1 rf
l
ij
f f Lf; d l
ij
f f (24)
where
l
ij
R
O
ij
H
0
fLf; d
R
O
ij
H
0
ff f
(25)
and
f f H
0
f rf
(26)
YANG GE AND LIANGSHIH FAN 12
B. GOVERNING EQUATIONS FOR THE GAS LIQUID SOLID FLOW
The gasliquidsolid ow is characterized by a wide range of physical length
scales, including small to large eddies in the bubble wake, and size in the milli
meter range for solid particles and in the millimeter/centimeter range for gas
bubbles. The accurate description of the gas bubble surface and bubbling ow
requires the use of ne grids, while the tracking of the motion of solid particles
needs the grid size to be much larger than the particle sizes.
For simulation of a gasliquidsolid uidized bed, the locally averaged
NavierStokes equations (Anderson and Jackson, 1967) are used to describe the
liquid phase ow outside the gas bubble, and the gas phase ow inside the gas
bubble. Due to the large grid size used, the liquid phase turbulence needs to be
considered. In this study, a modied coefcient that illustrates the effect of the
bubbleinduced turbulence for a subgrid scale (SGS) stress model is employed.
The levelset method and the numerical procedures described in Sections II.A
and III. are used to simulate the motion and the topological variation of the gas
bubble. The locally averaged governing equations of Eqs. (5) and (6) for liquid
ow outside the bubble and gas ow inside the bubble are given as:
@r
@t
r rV 0 (27)
@rV
@t
r rVV rp r s r s
sg
rg F
D
F
s
(28)
e represents the void fraction of liquid or gas and satises:
p
1 (29)
where e
p
is the void fraction of solid particles. t
sg
the SGS stress term. It is
modeled by the Smagorinsky (1963) model written as
s
sg
ij
n
T
@V
i
@x
j
@V
j
@x
i
(30)
where n
T
is dened as
n
T
C
s
l
2
S j j (31)
for bulk ow, and
n
T
C
s
f yl
2
S j j (32)
for walls with a wall function f(y). C
s
is the Smagorinsky coefcient, l D, and
S j j
2S
ij
S
ij
p
(33)
SIMULATION OF GAS LIQUID AND GAS LIQUID SOLID FLOW SYSTEMS 13
The volumetric uidparticle interaction force F
D
in Eq. (28) is calculated from
the forces acting on the individual particles in a cell:
F
D
P
f
k
d
DO
ij
, (34)
where f
d
is the uidparticle interaction force for a single particle and DO the
cell volume.
C. MODELING THE MOTION AND COLLISION DYNAMICS OF SOLID PARTICLES IN
GAS LIQUID SOLID FLUIDIZATION
The motion of a particle in the ow eld can be described in the Lagrangian
coordinate with the origin placed at the center of the moving particle. There are
two modes of particle motion, translation and rotation. Interparticle collisions
result in both the translational and the rotational movement, while the uid
hydrodynamic forces cause particle translation. Assuming that the force acting
on a particle can be determined exclusively from its interaction with the sur
rounding liquid and gas, the motion of a single particle without collision with
another particle can be described by Newtons second law as
dx
p
dt
V
p
(35)
m
p
dV
p
dt
m
p
g
p
6
d
3
p
rp r s r s
sg
f
d
f
am
f
s
(36)
where x
p
and V
p
are the particle position and particle velocity, respectively, and
d
p
the diameter of the particle. The ve terms on the righthand side of Eq. (36)
represent, respectively, the gravity force, the uid stress gradient force, the total
drag force, the added mass force and the bubblesurfacetensioninduced force.
The Saffman, the Magnus, and the Basset forces are ignored.
Note that the lubrication effect due to particle collisions in liquid is signi
cant. The liquid layer dynamics pertaining to the lubrication effect was exami
ned by Zenit and Hunt (1999). Zhang et al. (1999) used a LatticeBoltzmann
(LB) simulation to account for a closerange particle collision effect and
developed a correction factor for the drag force for closerange collisions, or the
lubrication effect. Such a term has been incorporated in a 2D simulation based
on the VOF method (Li et al., 1999). Equation (36) does not consider the
lubrication effect. Clearly, this is a crude assumption. However, in the three
phase ow simulation, this study is intended to simulate only the dilute solids
suspension condition (e
p
0.423.4%) with the bubble ow time of less than 1 s
starting when bubbles are introduced to the solids suspension at a prescribed e
p
.
YANG GE AND LIANGSHIH FAN 14
The particle collision effect under this simulation condition, therefore, would
be small.
Note that depending on the manner in which the drag force and the buoyancy
force are accounted for in the decomposition of the total uidparticle inter
active force, different forms of the particle motion equation may result (Jackson,
2000). In Eq. (36), the total uidparticle interaction force is considered to be
decomposed into two parts: a drag force (f
d
) and a uid stress gradient force (see
Eq. (2.29) in Jackson, 2000)). The drag force can be related to that expressed by
the WenYu equation, f
WenYu
, by
f
d
f
WenYu
(37)
The Wen and Yu (1966) equation is given by
f
WenYu
1
8
pd
2
p
C
D
2
r V V
p
V V
p
(38)
where the effective drag coefcient C
D
is calculated by
C
D
C
D0
4:7
(39)
In Eq. (39), C
D0
is a function of the particle Reynolds number, Re
p
rd
p
jV V
p
j=m. For rigid spherical particles, the drag coefcient C
D0
can be
estimated by the following equations (Rowe and Henwood, 1961):
C
D0
24
Re
p
1 0:15Re
p
0:687
; Re
p
o1000
0:44; Re
p
! 1000
(
(40)
The added mass force accounts for the resistance of the uid mass that is moving
at the same acceleration as the particle. Neglecting the effect of the particle
concentration on the virtualmass coefcient, for a spherical particle, the volume
of the added mass is equal to onehalf of the particle volume, so that
f
am
1
12
pd
3
p
r
DV
Dt
DV
p
Dt
(41)
When particles approach the gasliquid interface, the surfacetension force acts
on the particles through the liquid lm. The bubblesurfacetension induced
force can be described by
f
s
p
6
d
3
p
sKfdfrf (42)
When the particle inertia overcomes the surfacetensioninduced force, the par
ticle will penetrate the bubbles. Recognizing that particle penetration may not
lead to bubble breakage, details of bubble instability due to particle collision are
given in Chen and Fan (1989a, b).
SIMULATION OF GAS LIQUID AND GAS LIQUID SOLID FLOW SYSTEMS 15
To simulate the particleparticle collision, the hardsphere model, which is
based on the conservation law for linear momentum and angular momentum, is
used. Two empirical parameters, a restitution coefcient of 0.9 and a friction
coefcient of 0.3, are utilized in the simulation. In this study, collisions between
spherical particles are assumed to be binary and quasiinstantaneous. The equa
tions, which follow those of molecular dynamic simulation, are used to locate
the minimum ight time of particles before any collision. Compared with the
softsphere particleparticle collision model, the hardsphere model accounts for
the rotational particle motion in the collision dynamics calculation; thus, only
the translational motion equation is required to describe the uid induced par
ticle motion. In addition, the hardsphere model also permits larger time steps in
the calculation; therefore, the simulation of a sequence of collisions can be more
computationally effective. The details of this approach can be found in the
literature (Hoomans et al., 1996; Crowe et al., 1998).
D. RESULTS AND DISCUSSIONS
The computation performed in this study is based on the model equations
developed in this study as presented in Sections II.A, III.A, III.B, and III.C
These equations are incorporated into a 3D hydrodynamic solver, CFDLIB,
developed by the Los Alamos National Laboratory (Kashiwa et al., 1994). In
what follows, simple cases including a single air bubble rising in water, and
bubble formation from a single nozzle in bubble columns are rst simulated. To
verify the accuracy of the model, experiments are also conducted for these cases
and the experimental results are compared with the simulation results. Simu
lations are performed to account for the bubblerise phenomena in liquidsolid
suspensions with single nozzles. Finally, the interactive behavior between bub
bles and solid particles is examined. The bubble formation and rise from multi
ple nozzles is simulated, and the limitation of the applicability of the models is
discussed.
1. Single Air Bubble Rising in Water
The simulation for a single air bubble rising in water (density: 0.998 kg/cm
3
;
viscosity: 0.01 Pa s; surface tension: 0.0728 N/m) is performed in a 4 4
8 cm
3
3D column. A uniform grid size of 0.05 cm is used for three dimensions
which generates 80 80 160 ( 1.024 10
6
) grid points in the computational
domain. Initially, a spherical air bubble is positioned at rest in this domain with
its center located 0.5 cm above the bottom and the liquid is quiescent. The free
slip boundary conditions are imposed on all six walls. Note that the dimension
of the computational domain is selected based on numerical experiments. It is
found that, under both freeslip and noslip wall boundary conditions, when the
distance of the bubble interface to the wall is more than twice as large as the
YANG GE AND LIANGSHIH FAN 16
bubble diameter, there is practically little effect of the wall boundary conditions
on the simulation results. Thus, with an exception of the cases involving several
bubbles which migrate to the nearwall region through the zigzag motion that
will be discussed in a later section, the simulation results obtained in this study
are not affected by these wall boundary conditions. The time for an air bubble
of 0.8 cm in diameter rising from the initial position to the outlet of the column
is about 0.4 s, which corresponds to a bubble rise velocity of 18.75 cm/s.
The simulation was conducted using a Cray SV1 supercomputer at the Ohio
Supercomputer Center (OSC). The CPU time to compute the entire process of
bubble rising is 4 h.
The simulation results for the positions and the shape changes of the 0.8 cm
air bubble rising in water are shown in Fig. 2. The value for the Eotvos number
(E
0
gDrd
e
2
/s) and the Morton number (M gm
4
Dr/r
2
s
3
) are 8.5 and
2.5 10
7
, respectively. The time increment between two bubble images in
Fig. 2 is 0.05 s. As seen in the gure, the bubble shape undergoes continuous
changes from a sphere initially to an oblate ellipsoidal cap, and uctuates
between an oblate ellipsoidal cap and a spherical cap. The rectilinear motion
of a bubble in water exhibited in the gure, occurs for the rst several frac
tions of a second of a single bubble rising (with symmetric wakes) in a quiescent
liquid even though the bubble Reynolds number (Re
b
) is in the wake shedding
regime (Re
b
4400). The computed results obtained in this study capture such
FIG. 2. Simulated positions and shape variations of a rising bubble in a water column. Initial
bubble diameter 0.8 cm and time increment 0.05 s.
SIMULATION OF GAS LIQUID AND GAS LIQUID SOLID FLOW SYSTEMS 17
phenomena well. The liquideld disturbance would eventually induce an asym
metric wake, yielding wakeshedding phenomena of the rising bubble.
The meshrenement studies are conducted to examine the mesh effect on the
computation results. Simulations of a 0.8 cm air bubble rising in a 4 4 8 cm
3
water column (as shown in Fig. 2) are repeated at three different mesh reso
lutions: from a lower resolution of 40 40 80 grid points with a grid size of
0.1 cm to a higher resolution of 100 100 200 grid points with a grid size of
0.04 cm. The simulated variations of bubble rise velocity and bubble aspect ratio
(height/width) with time are shown in Figs. 3a and b, respectively.
Fig. 3a indicates that the bubblerise velocity measured based on the dis
placement of the top surface of the bubble (U
bt
) quickly increases and ap
proaches the terminal bubble rise velocity in 0.02 s. The small uctuation of U
bt
is caused by numerical instability. The bubblerise velocity measured based on
the displacement of the bottom surface of the bubble (U
bb
) uctuates signi
cantly with time initially and converges to U
bt
after 0.25 s. The overshooting of
U
bb
can reach 4550 cm/s in Fig. 3a. The uctuation of U
bb
reects the unsteady
oscillation of the bubble due to the wake ow and shedding at the base of
the bubble. Although the relative deviation between the simulation results of the
40 40 80 mesh and 100 100 200 mesh is notable, the deviation is insig
nicant between the results of the 80 80 160 mesh and those of the
100 100 200 mesh. The agreement with experiments at all resolutions is
generally reasonable, although the simulated terminal bubble rise velocities
($20 cm/s) are slightly lower than the experimental results (21$25 cm/s). A
lower bubblerise velocity obtained from the simulation is expected due to the
noslip condition imposed at the gasliquid interface, and the nite thickness for
the gasliquid interface employed in the computational scheme.
The aspect ratio shown in Fig. 3b describes the change of the bubble shape
with time during the bubble rising. The simulation and the experimental results
generally agree well, as shown in the gure. It can also be seen that the simu
lation results are only sensitive to the mesh size of 40 40 80 mesh and the
deviation between the results of 80 80 160 mesh and 100 100 200 mesh is
small. Thus, Figs. 3a and b indicate that a reasonable accuracy can be reached
in this bubblerise simulation with a 80 80 160 mesh (grid size of 0.05 cm).
The simulation results on bubble velocities, bubble shapes, and their uctu
ation shown in Fig. 3 are consistent with the existing correlations (Fan and
Tsuchiya, 1990) and experimental results obtained in this study. Bubble rise
experiments were conducted in a 4 cm 4 cm Plexiglas bubble column under
the same operating conditions as those of the simulations. Air and tap water
were used as the gas and liquid phases, respectively. Gas is introduced through a
6 mm nozzle. Note that water contamination would alter the bubblerise prop
erties in the surface tension dominated regime. In ambient conditions, this re
gime covers the equivalent bubble diameters from 0.8 to 4 mm (Fan and
Tsuchiya, 1990). All the airwater experiments and simulations of this study are
carried out under the condition where most equivalent bubble diameters exceed
YANG GE AND LIANGSHIH FAN 18
4 mm.These ow conditions correspond to the bubble inertial regime, and thus,
the extent of water contamination plays a negligible role in the determination of
the bubblerise properties.
The thickness of the gasliquid interface is set as 3D based on the parameters
used in the case of Sussman (1998), with the same densityratio on the interface
and similar Reynolds number. An interface thickness of 5D is also examined in
the simulation and no signicant improvement is observed. The accurate pre
diction of the bubble shape (shown in Figs. 3 and 4) can be attributed, in part,
to the manner in which the surfacetension force is treated as a body force in
the computation scheme. Specically, since the surfacetension force acting on a
solid particle is considered only when a solid particle crosses the gasliquid
interface and the solid particle is considered as a point, the accuracy of the
calculation of this force can be expected if the surface tension is interpreted as a
body force acting on each grid node near the interface.
2. Bubble Formation from an Orice
The airbubble formation from a single orice in water is simulated. The
computational domain is 2 2 4 cm
3
. A uniform grid size of 0.025 cm and
80 80 160 grid points are used to obtain convergent solutions for the bubble
formation process. This meshsize effect is examined by comparing the simu
lation results on the bubbleformation processes with experimental measure
ments. As shown in Fig. 4, decreasing the mesh size or increasing the mesh
resolution from 40 40 80 to 80 80 160 improves the accuracy of the
prediction results on the bubble shape. Further increase in mesh resolution does
not practically change the simulation results.
Simulations are then performed for gas bubbles emerging from a single nozzle
with 0.4 cm I.D. at an average nozzle velocity of 10 cm/s. The experimental
measurements of inlet gas injection velocity in the nozzle using an FMA3306 gas
ow meter reveals an inlet velocity uctuation of 315% of the mean inlet
velocity. A uctuation of 10% is imposed on the gas velocity for the nozzle to
represent the uctuating nature of the inlet gas velocities. The initial velocity of
the liquid is set as zero. An inow condition and an outow condition are
assumed for the bottom wall and the top walls, respectively, with the freeslip
boundary condition for the side walls.
Fig. 5 shows the simulated airbubble formation and rising behavior in water.
For the rst three bubbles, the formation process is characterized by three
distinct stages of expansion, detachment, and deformation. In comparison with
the bubble formation in the airhydrocarbon uid (Paratherm) system, the co
alescence of the rst two bubbles occurs much earlier in the airwater system.
Note that the physical properties of the Paratherm are r
l
870 kg/m
3
,
m
l
0.032 Pa s, and s 0.029 N/m at 25 1C and 0.1 MPa. This is due to the
fact that, compared to that in the airParatherm system, the rst bubble in the
airwater system is much larger in size and hence higher in rise velocity leading
SIMULATION OF GAS LIQUID AND GAS LIQUID SOLID FLOW SYSTEMS 19
to a longer time for its coalescence with the second bubble. Beginning with the
third bubble, the formation and rising behavior of air bubbles in water shows
strongly asymmetric behavior. As is evident from Fig. 5, the bubble rises in a
spiral path or a zigzag path.
FIG. 3. (a) Comparison of the simulation results and experimental results of the bubble rise
velocity. (b) Comparison of the simulation and experimental results of the bubble aspect ratio.
YANG GE AND LIANGSHIH FAN 20
In order to verify the simulation results, experiments on bubble behavior
in bubble columns are carried out under conditions similar to the simulations.
A 3D rectangular bubble column with the dimension of 8 8 20 cm
3
is used
for the experiments. Four nozzles with 0.4 cm I.D. and a displacement of 2.4 cm
are designed in the experiments. For singlenozzle experiments, air is injected
into the liquid bed through one of the orices while the others are shut off. The
outlet air velocity from the nozzle is approximated using the measured bubbling
FIG. 4. Comparison of the experimental measurement and the simulation results with different
resolutions of airbubble formation in water.
SIMULATION OF GAS LIQUID AND GAS LIQUID SOLID FLOW SYSTEMS 21
FIG. 5. Simulation results of airbubble formation from a single nozzle in water. Nozzle size 0.4 cm I.D. and nozzle gas velocity 10 cm/s.
Y
A
N
G
G
E
A
N
D
L
I
A
N
G

S
H
I
H
F
A
N
2
2
frequency and the initial bubble size. A highspeed video camera (240 frames/s)
is used to obtain the images of bubbles emerging from the orice in the liquid.
A common dimensionless number used to characterize the bubble formation
from orices through a gas chamber is the capacitance number dened as:
N
c
4V
c
gr
l
/pD
0
2
P
s
. For the bubbleformation system with inlet gas provided by
nozzle tubes connected to an air compressor, the volume of the gas chamber is
negligible, and thus, the dimensionless capacitance number is close to zero. The
gasow rate through the nozzle would be near constant. For bubble formation
under the constant ow rate condition, an increasing ow rate signicantly
increases the frequency of bubble formation. The initial bubble size also in
creases with an increase in the ow rate. Experimental results are shown in
Fig. 6. Three different nozzleinlet velocities are used in the airwater experi
ments. It is clearly seen that at all velocities used for nozzle air injection, bubbles
rise in a zigzag path and a spiral motion of the bubbles prevails in airwater
experiments. The simulation results on bubble formation and rise behavior
conducted in this study closely resemble the experimental results.
FIG. 6. Experimental results on airbubble formation and bubble rising in water. Nozzle size
0.4 cm I.D. and nozzle gas velocity (a) $6.0 cm/s; (b) $10.0 cm/s; (c) $14 cm/s.
SIMULATION OF GAS LIQUID AND GAS LIQUID SOLID FLOW SYSTEMS 23
3. Gas Liquid Solid Fluidization
As noted earlier, to simulate the bubble motion in a gasliquid bubble column
accurately, ne grid sizes, 0.025 cm for airwater and 0.05 cm for airParatherm
system, should be used in the computation. This negrid computation yields
essentially the results of DNS. These grid sizes are smaller than the size of the
solid particle usually employed for the threephase uidized bed operation. For
the particle size of 0.08 cm used in the present simulation of a threephase
uidized bed, the computational grid size is required to be no less than 0.2 cm
in order to track both the bubble ow and the particle motion. Note that
the system simulated in this study is a dilute liquidsolid bed with a minimum of
particleparticle collisions and uniform particle distribution. Although a grid
size of D410 D
p
as generally used in the Lagrangian simulation of uidparticle
ows is preferable, the grid size used under the current simulation of threephase
ows is acceptable. There were no numerical stability or convergence problems
encountered in the computation. For simulation of the bubble formation in a
gasliquid bubble column, a coarse grid size of 0.2 cm in a 4 4 8 cm
3
domain
with 21 21 41 grid points is used in this study. However, due to this large
grid size used, without any turbulence model, the simulation cannot accurately
track the discrete bubbleformation process. Specically, simulation without
consideration of the turbulent effects, the bubbles with distorted wake structure
are seen to be connected like a jet above a nozzle. An SGS stress model is thus
employed and incorporated into the code for subsequent simulation. The sim
ulation of the gasliquid bubble column system indicates that experimental
results on a bubble formation in an airParatherm medium can be well de
scribed when C
s
values are in a range of 1.0 to 1.2 with a grid size of 0.2 cm.Note
that the values for the Smagorinsky coefcient for single phase ow are 0.10.2.
The results are shown in Fig. 7(a).
Subsequently, simulations are performed for the airParathermsolid uidi
zed bed system with solid particles of 0.08 cm in diameter and 0.896 g/cm
3
in
density. The solid particle density is very close to the liquid density (0.868 g/
cm
3
). The boundary condition for the gas phase is inow and outow for the
bottom and the top walls, respectively. Particles are initially distributed in the
liquid medium in which no ows for the liquid and particles are allowed through
the bottom and top walls. Free slip boundary conditions are imposed on the
four side walls. Specic simulation conditions for the particles are given as
follows: Case (b) 2,000 particles randomly placed in a 4 4 8 cm
3
column;
Case (c) 8,000 particles randomly placed in a 4 4 8 cm
3
column; and Case (d)
8,000 particles randomly placed in the lower half of the 4 4 8 cm
3
column.
The solids volume fractions are 0.42, 1.68, and 3.35%, respectively for Cases
(b), (c), and (d).
The bubbleformation process at different solids concentrations is shown in
Figs. 7(b)(d) and is compared with that without particles as shown in Fig. 7(a).
For the rst 0.3 s, little change is observed in the bubbleformation process
YANG GE AND LIANGSHIH FAN 24
for the three solids concentrations used in this simulation. After 0.4 s, how
ever, signicant changes can be found for the cases with high solids concen
trations. This can be seen from the rst bubble in each case. When the solid
concentration is low or no solids are present, the rst bubble grows on the
FIG. 7. Simulation results of bubble formation and rising in Paratherm NF heattransfer uid
with and without particles. Nozzle size 0.4 cm I.D., liquid velocity 0 cm/s, gas velocity 10 cm/s, and
particle density 0.896 g/cm
3
. (a) No particle; (b) 2000 particles; (c) 8000 particles; (d) 8000 particles.
SIMULATION OF GAS LIQUID AND GAS LIQUID SOLID FLOW SYSTEMS 25
orice and connects to the second bubble. For the high solids concentra
tion cases, the rst bubble is not well connected to the second bubble. This is
particularly true for Case (d) when the bubble rises into the solidsfree region or
freeboard region of the bed. The solid particle entrainment is clearly observed in
Case (d).
FIG. 7 (Continued)
YANG GE AND LIANGSHIH FAN 26
IV. System 2: Deformation Dynamics of Liquid Droplet in
Collision with a Particle with FilmBoiling Evaporation
The phenomena of evaporative liquid droplets impacting onto solid objects at
high temperatures are of relevance to many engineering problems, such as
sprinkler systems in the iron making or metalcasting processes, inkjet spray
painting, impingement of oil droplets on turbine engines, meteorology, and
spray coating of substrates. An evaporative liquid jet in gassolid ow systems
is also of interest to current technology applications in chemical, petroleum, and
materials processing industries, such as FCC, polyethylene synthesis (Kunii and
Levenspiel, 1991; Fan et al., 2001) and microelectronic materials manufactur
ing. In FCC riser reactors, for example, gas oil at a low temperature is injected
into the riser from feed nozzles located at the bottom of the riser and the mist
droplets formed from the spray contact with hightemperature uidized catalyst
particles. The vaporized oil then carries the catalyst particle up through the
riser. In the feed nozzle region, the size of the droplet can be comparable or
signicantly smaller (or larger) than the size of particle. The droplet can always
have a different momentum, thus the collision between the catalytic particles
and oil droplet may have various modes. Fig. 8 shows some of the collision
modes existing in a feednozzle region (Zhu et al., 2000). Smaller droplets may
rebound from the surface of larger particles upon impact, and smaller particles
Large
Droplet
Small
Droplet
FIG. 8. Various modes of dropletparticle collisions.
SIMULATION OF GAS LIQUID AND GAS LIQUID SOLID FLOW SYSTEMS 27
may penetrate through or penetrate but retain inside the larger droplets. Larger
droplets may break into smaller drops during the impact and/or remain at
tached to the particle surface after the collision, which may intensify the particle
aggregation. Clearly, understanding the droplet and particle collision mechanics
are crucial to an accurate account of the momentum and heat transfer between
the droplet and solid object, which is important for prediction of hydrocarbon
product distributions in light of catalytic and the thermalcracking reactions in
the riser. It is also relevant to the design of feed nozzles that provide desired
droplet properties for optimum droplet contact with catalyst particles in the
reactor.
In most of the applications, the solid objects (e.g., the catalyst particles in
FCC reactor) are always under high temperature, and the droplet impact proc
esses are accompanied with intensive evaporation. The nature of the collision
of the droplet with the superheated objects exhibits a great diversity in hydro
dynamic and thermodynamic properties, such as droplet splash and rebound,
wetting or nonwetting contact, nucleate boiling or lm boiling, and Marangoni
effect. Further, the droplet shape, the contact area and the cooling effectiveness
during the impact not only depend on such hydrodynamic forces as the inertia,
pressure, surface tension, and viscous forces but also on the degrees of the
surface superheating and the droplet subcooling (Inada et al., 1985). As
the solid temperature rises to superheated conditions, the characteristics of
liquidsolid contact signicantly change and the evaporation rate affects the
droplet hydrodynamics. Under this condition, the nonwetting contact may
develop during the collision, and the evaporation is under the lmboiling re
gime, or so called Leidenfrost regime (Gottfried et al., 1966). In the Leidenfrost
regime, the vapor pressure generated from the droplet evaporation prevents
the direct contact of the droplet with the solid objects. The heat transfer from
the hot objects to the droplet is also hindered due to the resistance of the vapor
layer existing between the droplet and the solid surface. In this work, a 3D
numerical model is developed and the simulation is conducted to account for
the behavior of the dropletparticle collision in the Leidenfrost regime.
Experimental and numerical studies of droplets impacting onto a at surface
of varied temperatures have been extensively reported in the literature. The
effects of the initial droplet temperature on lmboiling impact are signicant
(Inada et al., 1985; Harvie and Fletcher, 2001b). Depending on the initial
droplet temperature, there are two types of droplet impact: saturated impact
and subcooled impact. The saturated impact involves the initial temperature
at the boiling point of the liquid or saturation temperature of the liquid. The
subcooled impacts, on the other hand, involve the droplet initial impact
temperature below the liquidsaturation temperature. The experimental results
for these two types of impact are briey described below. The modeling and
numerical approaches used for the droplet impingement onto isothermal or
heated at wall are also given.
YANG GE AND LIANGSHIH FAN 28
A. SIMULATION OF SATURATED DROPLET IMPACT ON FLAT SURFACE IN THE
LEIDENFROST REGIME
Wachters and Westerling (1966) rst presented a classication of the dynamic
regimes of the impact based on their experiments in which water drops with a
diameter of 2.3 mm impact on a polished gold surface at temperatures between
200 1C and 400 1C. They studied the saturated impact of water droplets and
found that for the impact with Weo30, where We 2r
l
V
2
R/s, the surface
tension of the droplet dominates the impact process, and the droplet recoils and
rebounds from the surface without disintegration. At 30oWeo80, the droplet
undergoes a similar spreading and recoiling process as that for Weo30. In the
rebounding process, the droplet may disintegrate into several smaller droplets
(secondary droplets) and the shape of the droplet may then become unstable.
For the impact with We480, the impact inertial force (or kinetic energy) is so
large that splashing occurs during the early stage of the impact, while the
droplet breaks up into a number of small droplets. Based on the measured heat
ux on the solid surface, Wachters and Westerling (1966) also estimated the
relative volume decrease of the droplet during the impact. It was found that,
when the solid temperature is higher than 200 1C, the averaged evaporation rate
of the droplet decreases with an increase in the surface temperature. At the
nonwetting condition when the surface temperature reaches 400 1C, the volume
(mass) change of the droplet due to the evaporation during the impact is slight
(0.20.3%) for a wide range of We.
Groendes and Mesler (1982) studied the saturated lm boiling impacts of a
4.7 mm water droplet on a quartz surface of 460 1C. The uctuation of the
surface temperature was detected using a fastresponse thermometer. The maxi
mal temperature drop of the solid surface during a droplet impact was reported
to be about 20 1C. Considering the lower thermal diffusivity of quartz, this
temperature drop implies a low heattransfer rate on the surface. Biance et al.
(2003) studied the steadystate evaporation of the water droplet on a super
heated surface and found that for the nonwetting contact condition, the droplet
size cannot exceed the capillary length.
Ge and Fan (2005) developed a 3D numerical model based on the levelset
method and nitevolume technique to simulate the saturated droplet impact on
a superheated at surface. A 2D vaporow model was coupled with the heat
transfer model to account for the vaporow dynamics caused by the Leidenfrost
evaporation. The droplet is assumed to be spherical before the collision and the
liquid is assumed to be incompressible.
1. Hydrodynamic Model and Numerical Solution
In the levelset method, the free surface of the droplet is taken as the zero in
the levelset function f~x; t as given in Eq. (2). The motion of the interface is
SIMULATION OF GAS LIQUID AND GAS LIQUID SOLID FLOW SYSTEMS 29
traced by solving the HamiltonJacobitype convection equation, given as Eq.
(3), in the computational domain. The mass loss of the droplet due to evapo
ration during the impact process is neglected in surfacetracking based on the
experimental results of Wachters and Westerling (1966). With the levelset
method, the equation of motion of the uid follows the NavierStokes equation
as given by Eqs. (5) and (6). The density and viscosity are dened by Eqs. (10)
and (11).
The computational code used in solving the hydrodynamic equation is de
veloped based on the CFDLIB, a nitevolume hydrocode using a common
data structure and a common numerical method (Kashiwa et al., 1994). An
explicit timemarching, cellcentered Implicit Continuousuid Eulerian (ICE)
numerical technique is employed to solve the governing equations (Amsden
and Harlow, 1968). The computation cycle is split to two distinct phases: a
Lagrangian phase and a remapping phase, in which the Arbitrary Lagrangian
Eulerian (ALE) technique is applied to support the arbitrary mesh motion with
uid ow.
Let f
n
f~x; t
n
and
~
V
n
~
V~x; t
n
be the cellcentered levelset function and
velocity at time t
n
, respectively. The numerical procedures to solve the velocity
eld
~
V
n1
, and the levelset function f
n+1
at t
n+1
t
n
+Dt can be described
below:
(1) Compute the velocity eld
~
V
n1
by solving the governing equation, Eqs.
(56), using the cellcentered ICE technique and ALE technique (Kashiwa
et al., 1994).
(2) Solve the convection equation of f~x; t (Eq. (3)) to obtain the
f
n1
. The
high order (3rd order) essentially nonoscillatory (ENO) upwind scheme
(Sussman et al., 1994) is used to calculate the convective term
~
V
G
rf based
on the updated velocity eld
~
V
n1
. The time advancement is accomplished
using the secondorder total variation diminishing (TVD) RungeKutta
method (Chen and Fan, 2004).
(3) Perform the redistance procedure to obtain the f
n+1
using
f
n1
as the initial
value. The detail of the redistance computation is given by Sussman et al.
(1998).
(4) Calculate the density and the viscosity of the eld using Eqs. (10)(11) with
the updated levelset function f
n+1
.
The time steps (Dt) for calculating the
~
V
n1
and the f
n+1
are the same, which
is determined by the CFL condition and under constraints of the viscous and
surface tension (Sussman et al., 1994).
Considering a surface temperature which is higher than the Leidenfrost tem
perature of the liquid in this study, it is assumed that there exists a microscale
vapor layer which prevents a direct contact of the droplet and the surface.
Similar to Fujimoto and Hatta (1996), the noslip boundary condition is adopted
at the solid surface during the dropletspreading process and the freeslip
YANG GE AND LIANGSHIH FAN 30
condition is applied for the recoiling and rebounding periods. The velocity at
the grid point inside the solid surface is solved together with whole domain but is
reset according to the relative boundary condition (Ge and Fan, 2005).
2. VaporFlow Model
As the thickness of the vapor layer (520 mm) is several orders of magnitude
smaller than the macroscale of the ow eld (i.e., the diameter of the droplet), it
would be impractical to use the same computation mesh for both macroow
and vaporlayer ow (Harvie and Fletcher, 2001a). Thus, a 2D model is de
veloped to simulate the dynamics of the vapor ow between the droplet and the
surface. For the lmboiling impact problem, the vaporlayer model would
allow determination of the evaporationinduced pressure in the vapor layer
without neglecting the inertial force of the vapor ow. In the symmetrical co
ordinates (x,l) shown in Fig. 9, assuming that the gas in the vapor layer is only
saturated vapor and neglecting the temporal term, the continuity and momen
tum equations for incompressible vapor ows with gravitation terms neglected
are given by
@u
x
@x
u
x
x
@u
l
@l
0 (43)
u
x
@u
x
@x
u
l
@u
x
@l
@
@x
P
r
n
@
2
u
x
@l
2
@
2
u
l
@x@l
(44)
u
x
@u
l
@x
u
l
@u
l
@l
@
@l
P
r
n
@
2
u
l
@x
2
@
2
u
x
@x@l
1
x
@u
l
@x
1
x
@u
x
@l
(45)
where u
x
, u
l
are the vaporow velocities in x and l direction, respectively, n is
the kinematical viscosity of the vapor. To determine the relative signicance of
droplet
Vapor flow
O
Solid surface
FIG. 9. Coordinates for the vaporlayer model.
SIMULATION OF GAS LIQUID AND GAS LIQUID SOLID FLOW SYSTEMS 31
each term in these motion equations, an order of magnitude analysis is made by
considering the following dimensionless groups:
x
x
R
; Z
l
d
; u
x
u
x
U
x
; u
l
u
l
U
l
; p
p
rU
2
x
;
t
tU
x
R
; Re
d
dxu
ld
n
(46)
where R is the droplet radius; d the vaporlayer thickness; Re
d
the local evapo
ration Reynolds number; u
ld
(x) the local vapor velocity; and U
x
, U
l
the velocity
scalars in x, l directions, respectively. Two assumptions can be made in ac
counting for the collision process: (a) The vaporlayer thickness is much smaller
than the radius of the droplet; (b) The velocity for the vapor ow is much larger
than the rates of variation of the vaporlayer thickness and breadth. Based on
these assumptions and the order of magnitude analysis, the x momentum equa
tion can be simplied to:
u
x
@u
x
@x
u
l
@u
x
@l
@
@x
P
r
n
@
2
u
x
@l
2
(47)
The boundary conditions are:
l 0; u
x
x; 0 u
l
x; 0 0
l d; u
x
x; d u
l
x; u
l
x; d u
ld
x
x 0;
@
@x
0; x x
b
; p p
b
(48)
where p
b
is the pressure of the ambient gas at the outside edge of the vapor layer.
In the impact process that involves large temperature differences (DT)
between the surface and the droplet, such as the ones considered in this study
(e.g., DT300500 1C), the value for Re
d
is about 0.51.0. Thus, the inertial
force of the vapor ow would be of the same order of magnitude as the viscous
force, and cannot be neglected in Eq. (47) for the vaporow model.
To solve Eq. (47), a variable transformation is considered:
u
l
x; Z Zu
ld
x u
x
x; Z OxFZ (49)
u
ld
(x) can be calculated through the energybalance equation at the
vapordroplet interface. O(x) and F(Z) are singlevariable functions.
With this transformation, the solution for the x momentum can be converted
into that of an ordinary differential equation (ODE) of O(x):
F
00
Z Re
d
ZF
0
Z Re
d
FZ jx (50)
jx
d
2
nOx
@
@x
P=r (51)
YANG GE AND LIANGSHIH FAN 32
The general solution of the Eq. (50) can be obtained in power series form.
Under the condition that Re
d
O1, F(Z) can be approximated by only in
cluding the rst three terms in the power series with good accuracy:
FZ F
d
Z jxZ
1
2
Re
d
24
jx
Z
2
2
Re
d
24
Z
4
(52)
The averaged vaporow velocity is given by
u
x
x O
Z
1
0
F
d
dZ
1
2
u
ld
x
d
2
12g
1
3
20
Re
d
@
@x
p
r
(53)
The vaporcontinuity equation can be expressed by
u
x
x
1
xdx
Z
x
0
x
0
u
ld
x
0
dx
0
(54)
The pressure distribution in the vapor layer can be obtained by solving Eqs.
(53) and (54) using a piecewise integration method (Ge and Fan, 2005). In this
procedure, the thickness of the vapor layer d(x) is obtained from the levelset
function. The u
ld
(x) is calculated by
u
ld
x _ m=r
v
@dx
@t
(55)
where the local evaporation rate m is dened by the heattransfer model. The
vaporpressure force simulated by this model is applied as an interfacial force to
the droplet bottom surface.
3. HeatTransfer Model
Heat transfer occurs not only within the solid surface, droplet and vapor
phases, but also at the liquidsolid and solidvapor interface. Thus, the energy
balance equations for all phases and interfaces are solved to determine the heat
transfer rate and evaporation rate.
Inside the solid surface, the heatconduction equation in 3D coordinates is
@T
s
@t
a
s
@
@x
@T
s
@x
@
@y
@T
s
@y
@
@z
@T
s
@z
(56)
where T
s
(x,y,z) is the solid temperature and a
s
the thermal diffusivity of solid.
The heat transfer within the droplet is described by the following thermal
energy transport equation with neglecting viscous dissipation:
@T
d
@t
u
@T
d
@x
v
@T
d
@y
w
@T
d
@z
a
d
@
2
T
d
@x
2
@
2
T
d
@y
2
@
2
T
d
@z
2
(57)
SIMULATION OF GAS LIQUID AND GAS LIQUID SOLID FLOW SYSTEMS 33
Using the same assumptions that were made in the vaporlayer model, the
energyconservation equation for the incompressible 2D vapor phase can be
simplied to a 1D equation in boundary layer coordinates:
@
2
T
v
@Z
2
0 (58)
The radiative heat transfer across the vapor layer is neglected under the
condition that the solid temperature is lower than 700 1C (Harvie and Fletcher,
2001a,b). On the liquidvapor interface, the energybalance equation is
k
v
T
ss
T
ds
d
_ mL
c
(59)
where k
v
is the thermal conductivity of the vapor; T
ss
and T
ds
are the temper
atures of the solid surface and the droplet surface. The thermal boundary con
dition at the solidvapor interface is
k
v
T
ss
T
ds
d
k
s
@T
s
@Z
(60)
where the heat ux in the Z direction is assumed to be much larger that that in
the x direction.
The numerical method used for solving the heattransfer equation is similar to
that for solving the momentum equation, which is a nitevolume, ALE method
(Kashiwa et al., 1994).
4. Results and Discussion
To validate the model developed in the present study, the simulations are
rst conducted and compared with the experimental results of Wachters and
Westerling (1966). In their experiments, water droplets impact in the normal
direction onto a hot polished gold surface with an initial temperature of 400 1C.
Different impact velocities were applied in the experiment to test the effect of
the We number on the hydrodynamics of the impact. The simulation of this
study is conducted for cases with different Weber numbers, which represent
distinct dynamic regimes.
The simulation shown in Fig. 10 is an impact of a saturated water droplet of
2.3 mm in diameter onto a surface of 4001C with an impact velocity of 65 cm/s,
corresponding to a Weber number of 15. This simulation and all others pre
sented in this study are conducted on uniform meshes (Dx Dy Dz D). The
mesh resolution of the simulation shown in Fig. 10 was 0.08 mm in grid size,
although different resolutions are also tested and the results are compared in
Figs. 11 and 12. The average timestep in this case is around 5 ms. It takes 4000
iterations to simulate a real time of 20 ms of the impact process. The simulation
YANG GE AND LIANGSHIH FAN 34
FIG. 10. Water droplet impacts on a at surface. The initial droplet diameter is 2.3 mm and the
surface temperature is 400 1C. We 15.
SIMULATION OF GAS LIQUID AND GAS LIQUID SOLID FLOW SYSTEMS 35
code is run on the craySV1 supercomputer at the OSC. The computing time of
this case is about 12 h.
Comparing the 3D images simulated and the experimental photographs in
Fig. 10, it can be seen that the droplet shapes are well reproduced by the present
model. During the rst 3.5 ms of the impact (frames 13), a liquid lm with
attened disc shape is formed immediately after the impact. The inertial force
drives the liquid to continue spreading on the solid surface, while the surface
tension and the viscous forces resist the spreading of the liquid lm. As a result,
the droplet spreading speed decreases and the uid mass starts to accumulate at
FIG. 11. Simulated 3D views of the impact for We 15 as a function of the mesh resolution.
YANG GE AND LIANGSHIH FAN 36
the leading edge of the liquid lm (2.53.2 ms). After the droplet spreads to the
maximum extent, the liquid lm starts to shrink back to its center (frames 4
and 5) due to the surfacetension force at the edge of the lm. At 3.55ms (frame
4), the simulated droplet shows a concave structure with a void in the center,
which is also shown in the experimental photograph of 3.85 ms. This structure,
also called a ring structure, has also been widely reported in the literature. In the
crosssectional images at 3.55 ms, the velocity eld shows that the inward ow
rst starts from the outer edge of the liquid lm, which conrms that recoiling
ow is driven by the surface tension. After 4.4 ms, the droplet continues to recoil
and forms an upward ow in the center of the droplet (frames 5 and 6), and this
leads to a bouncing of the droplet up from the surface (frame 7). The peanut
shape droplet (also called dumbbell shape by Harvie and Fletcher (2001b) shown
in the experimental photograph at 14.56 ms is reproduced in the simulation.
The impact process shown in Fig. 10 is also simulated at six different grid
resolutions, i.e., 0.150, 0.120, 0.100, 0.075, 0.060, and 0.050 mm in mesh sizes,
with the corresponding cells per droplet radius (CPR) of 7.6, 9.6, 11.5, 15.3,
19.0, and 23.0, respectively. The comparison of the 3D images among three
resolutions is shown in Fig. 11. The corresponding CPRs of these resolutions
are 9.6, 11.5, and 15.3. It can be found that the simulated droplet shapes are
similar at all three resolutions during the spreading process and even the early
stage of the recoiling process (26 ms). The deviation appears in the late stage of
0 2 4 6 8 101 2 141 61 8
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1.4
1.6
1.8
2.0
0.150mm grid size
0.120mm grid size
0.100mm grid size
0.075mm grid size
0.060mm grid size
0.050mm grid size
Experiment of Wachters and Westerling
6
R
/
R
0
Time after impact(ms)
FIG. 12. Spread factor of the droplet verse time for the impact condition given in Fig. 10 at
different mesh resolutions.
SIMULATION OF GAS LIQUID AND GAS LIQUID SOLID FLOW SYSTEMS 37
the recoiling process (8 ms), while the droplet generated on coarser mesh
(0.12 mm) tends to be more uniform in structure and less elongated in the
vertical direction. The difference in the droplet shape for 0.1 mm mesh and
0.075 mm mesh is relatively small.
Fig. 12 shows the spread factors simulated on meshes with different reso
lutions along with the measurement value of Wachters and Westerling (1966).
The spread factor is dened as the radius of the droplet on the solid surface
divided by the initial radius of the droplet. Although the convergence is not
perfect, the agreement between the experiment and the simulations is relatively
good for all resolutions. Consistent with the results of Fig. 11, the effect of
the mesh resolution on spread factor becomes notable after 8 ms since the
moment of impact, and the coarser resolution tends to yield a slower rebound
ing process.
The simulations were also performed under same conditions as the case of
Fig. 10 but for higher impact velocities. The simulateddroplet dynamics and
heattransfer rate at the solid surface at different impact velocities are given in
Ge and Fan (2005).
B. SIMULATION OF SUBCOOLED DROPLET IMPACT ON FLAT SURFACE IN
LEIDENFROST REGIME
Subcooled impacts, in which the initial temperature of the droplet is below
the liquid saturation temperature, are of primary interest in experiments since
the condition of the spray liquids is often of the ambient temperature in prac
tical applications (Inada et al., 1985; Chandra and Avedisian, 1991; Chen and
Hsu, 1995). Chandra and Avedisian (1991) studied the collision dynamics of a
24 1C nheptane droplet impacting on a metallic surface with a Weber number
equal to 43. The transition from the nucleate boiling to the lm boiling was
identied when the surface temperature rises from the boiling point (170 1C) to
above the Leidenfrost temperature (200 1C) of nheptane. They found that un
der the lmboiling condition, the liquidsolid contact is hindered by the vapor
layer, as evidenced by the disappearance of the bubbles inside the liquid droplet
under the nucleate boiling condition. The contact angle of the liquid to the
surface was also reported to increase with an increase in the surface temper
ature, and reached 1801 in the lm boiling condition. Qiao and Chandra (1996)
measured the temperature drop of a stainless steel surface during the impact of
the subcooled water and the nheptane droplets in low gravity. They found that
when the surface temperature is above the superheat limit, the temperature drop
of the surface is relatively small for the impact of nheptane droplet (less than
20 1C). But for the impact of the water droplet, the temperature drop of the
surface can reach 150 1C, which implies a high heat ux and the intermittent
contact of the liquid and the solid surface. Hatta et al. (1997) found that at low
impact We number, the dynamics of water droplet is almost independent of
YANG GE AND LIANGSHIH FAN 38
the surface materials when the surface temperature is above the Leidenfrost
temperature.
The effects of the subcooling degree of the droplet on the lmboiling impact
are studied by Inada et al. (1985). They found that the heattransfer rate on
the solid surface during an impact of a 4mm water droplet increases signi
cantly with a decrease in the initial droplet temperature. The boiling regimes
were classied to represent different droplet dynamics and heattransfer modes
at various droplet and surface temperatures. Inada et al. (1988) also measured
the thickness of the vapor lm between the impinging droplet and the surface
at various degrees of subcooling. Chen and Hsu (1995) measured the tran
sient local heat ux at the surface of a presuperheated plate, which undergoes
the impingement of subcooled water droplets. A fastresponse microthermo
couple was designed to capture the instantaneous changes of the solidsurface
temperature. Although the droplet dynamics of the impact process was not
presented, they concluded that both the surface temperature and the degree
of the droplet subcooling are crucial to the intermittent contact mode at the
solid surface. At the lmboiling regime with the surface temperature super
heated at 400 1C, a subcooled droplet tends to disintegrate during the impact at
We 55.
In subcooled impact, the initial droplet temperature is lower than the sat
urated temperature of the liquid of the droplet, thus the transient heat transfer
inside the droplet needs to be considered. Since the thickness of the vapor layer
may be comparable with the mean free path of the gas molecules in the sub
cooled impact, the kinetic slip treatment of the boundary condition needs to be
applied at the liquidvapor and vaporsolid interface to modify the continuum
system.
1. Hydrodynamic Model
The ow eld of the impacting droplet and its surrounding gas is simulated
using a nitevolume solution of the governing equations in a 3D Cartesian
coordinate system. The levelset method is employed to simulate the movement
and deformation of the free surface of the droplet during impact. The details of
the hydrodynamic model and the numerical scheme are described in Sections
II.A and 1 V.A.1.
During the subcooled droplet impact, the droplet temperature will undergo
signicant changes due to heat transfer from the hot surface. As the liquid
properties such as density r
l
(T), viscosity m
l
(T), and surface tension s(T) vary
with the local temperature T, the local liquid properties can be quantied once
the local temperature can be accounted for. The droplet temperature is simu
lated by the following heattransfer model and vaporlayer model. Since the
liquid temperature changes from its initial temperature (usually room temper
ature) to the saturated temperature of the liquid during the impact, the linear
SIMULATION OF GAS LIQUID AND GAS LIQUID SOLID FLOW SYSTEMS 39
variation of liquid properties with the temperature is assumed, which is
gT g
0
T T
0
T
sa
T
0
g
sa
g
0
g r
l
; m
l
; s (61)
where T
0
and T
sa
are the initial and saturated temperatures of the liquid re
spectively; g
0
and g
sa
are the liquid property at T
0
and T
sa
, respectively. The
boundary condition adopted at the solid surface is described in Section IV.A.1.
2. Heat Transfer Inside the Droplet and Across the Vapor Layer
For the subcooling impact, especially for the high subcooling degree case in
which the droplet initial temperature is much lower than the saturated temper
ature, the heat transfer within the droplet is signicant and hence affects the
droplet evaporation rate. Neglecting the viscous dissipation, the equation of the
conservation of the thermal energy inside the droplet is given by
@T
@t
~
V rT a
l
r rT (62)
where a
1
is the thermal diffusivity of liquid. At other free surfaces of the droplet,
the adiabatic boundary condition is applied which is given by
~n
G
r
d
T 0 (63)
where ~n
G
is the normal vector of the droplet surface, which can be calculated
based on the levelset function:
~n
G
rf
rf
(64)
r
d
T is the temperature gradient which is evaluated only on the droplet side. The
heatconduction equation inside the solid surface is given in Section IV.A.3.
The heat transfer across the vapor layer and the temperature distribution in
the solid, liquid, and vapor phases are shown in Fig. 13. In the subcooled
impact, especially for a droplet of water, which has a larger latent heat, it has
been reported that the thickness of the vapor layer can be very small and in
some cases, the transient direct contact of the liquid and the solid surface may
occur (Chen and Hsu, 1995). When the length scale of the vapor gap is com
parable with the free path of the gas molecules, the kinetic slip treatment of the
boundary condition needs to be undertaken to modify the continuum system.
Consider the Knudsen number dened as the ratio of the average mean free
path of the vapor to the thickness of the vapor layer:
Kn
l
d
(65)
YANG GE AND LIANGSHIH FAN 40
where l is the mean free path of molecule. Harvie and Fletcher (2001c) analyzed
the kinetic of the molecular behavior at the solid and evaporative surface for
0.01oKno0.1. Based on their simple kinetic theory, the effective temperature
discontinuity at liquidvapor surface and solidvapor surface can be given by
T
s2
T
s1
C
T;s
T
s1
T
d2
(66)
T
d1
T
d2
C
T;l
T
s1
T
d1
(67)
where T
s1
,T
s2
,T
d1
,T
d2
are the interface temperatures of solid surface and droplet
shown in Fig. 13. C
T
is dened by
C
T
Kn
9
4
g
5
4
2 s
t
s
t
(68)
where s
t
is the thermal accommodation coefcient dened by Harvie and
Fletcher (2001c). At the liquidvapor interface, the energy balance equation is
given by
k
v
T
s1
T
d1
d
q
l
_ mL
c
(69)
where k
v
is the thermal conductivity of the vapor; L
c
the latent heat of the
liquid; q
1
the heat ux at the droplet surface at the liquid side, which is given by
q
1
k
1
r
d
T.
Td
Solid surface
Droplet
Td1
Ts2 Ts1
Td2
Ts
qs
ql
Vapor layer
FIG. 13. Temperature distribution and heat ux across the vapor layer.
SIMULATION OF GAS LIQUID AND GAS LIQUID SOLID FLOW SYSTEMS 41
3. VaporLayer Model with Kinetic Treatment at Boundary
The vaporlayer model developed in Section IV.A.2 is based on the contin
uum assumption of the vapor ow. This assumption, however, needs to be
modied by considering the kinetic slip at the boundary when the Knudsen
number of the vapor is larger than 0.01 (Bird, 1976). With the assumption
that the thickness of the vapor layer is much smaller than the radius of the
droplet, the reduced continuity and momentum equations for incompressible
vapor ows in the symmetrical coordinates (x,l) are given as Eqs. (43) and (47).
When the Knudsen number of the vapor ow is between 0.01 and 0.1, the ow
is in the slip regime. In this regime, the ow can still be considered as a
continuum at several mean free paths distance from the boundary, but an
effective slip velocity needs to be used to describe the molecular interac
tion between the gas molecules and the boundary. Based on the simple
kinetic analysis of vapor molecules near the interface (Harvie and Fletcher,
2001c), the boundary conditions of the vapor ow at the solid surface can be
given by
l 0; u
x
x; 0 F
s
@u
x
@l
x; 0; u
l
x; 0 0 (70)
and the boundary conditions at the droplet surface is
l d; u
x
x; d u
x;l
x F
l
@u
x
@l
x; d; u
l
x; d u
l;l
x (71)
where u
x,l
(x), u
l,l
(x) are the velocities of the droplet surface in x and l direction,
respectively; F
s
and F
l
are dened by
F
s
l
s
2 s
v;s
s
v;s
F
l
l
l
2 s
v;l
s
v;l
(72)
where l
s
, l
l
are the mean free path of the gas molecules near solid and droplet
surface.
Equations (43) and (47) can be solved by using the similar procedure as given
by Section IV.A.2 with the boundary condition given by Eqs. (70) and (71).
Thus the vapor pressure can be determined by
px p
0
Z
x
b
x
rjx dx (73)
where x
0
is the radius of the extent of the vapor layer andjris given by
jx
2nC
1
d
2
u
x
u
x
u
ll
d
B
1
B
2
u
2
x
x
B
1
(74)
YANG GE AND LIANGSHIH FAN 42
where
B
1
C
2
1
5
C
1
C
2
2
2C
1
C
3
C
2
2
3
C
1
C
3
C
2
3
B
2
2C
2
1
15
C
1
C
2
3
C
2
2
6
2
3
C
1
C
3
1
2
C
2
C
3
C
1
3
2 2k
l
2k
s
y
l
2k
s
y
l
1 k
l
4k
s
6k
l
k
s
C
2
6 6k
l
2y
l
1 k
l
4k
s
6k
l
k
s
C
3
k
s
C
2
75
where k
l
F
l
=d; k
s
F
s
=d; y
l
u
x;l
= u
x
.
The averaged velocity of the vapor is expressed by Eq. (54). The pressure
distribution in the vapor layer can be obtained by solving Eqs. (54) and (73)
(75) by a piecewise integration method. Details of the solving procedure and how
to use the vapor pressure in ow eld calculation are given in Section IV.A.2.
4. Results and Discussion
Three different subcooled impact conditions under which experiments were
conducted and reported in the literature are simulated in this study. They are:
(1) nheptane droplets (1.5 mm diameter) impacting on the stainless steel surface
with We 43 (Chandra and Avedisian, 1991), (2) 3.8 mm water droplets
impacting on the inconel surface at a velocity of 1 m/s (Chen and Hsu, 1995),
and (3) 4.0 mm water droplets impacting on the copper surface with We 25
(Inada et al., 1985). The simulations are conducted on uniform Cartesian
meshes (Dx Dy Dz D). The mesh size (resolution) is determined by
considering the mesh renement criterion in Section V.A. The mesh sizes in this
study are chosen to provide a resolution of CPR 15.
Fig. 14 shows the comparison of the photographs from Chandra and
Avedisian (1991) with simulated images of this study for a subcooled 1.5 mm
nheptane droplet impact onto a stainlesssteel surface of 200 1C. The impact
velocity is 93 cm/s, which gives a Weber number of 43 and a Reynolds number
of 2300. The initial temperature of the droplet is room temperature (20 1C). In
Fig. 14, it can be seen that the evolution of droplet shapes are well simulated by
the computation. In the rst 2.5 ms of the impact (frames 12), the droplet
spreads out right after the impact, and a disklike shape liquid lm is formed on
the surface. After the droplet reaches the maximum diameter at about 2.1 ms,
the liquid lm starts to retreat back to its center (frame 2 and 3) due to the
surfacetension force induced from the periphery of the droplet. Beyond 6.0 ms,
the droplet continues to recoil and forms an upward ow in the center of the
SIMULATION OF GAS LIQUID AND GAS LIQUID SOLID FLOW SYSTEMS 43
FIG. 14. nHeptane droplet collision with surface at 200 1C. Experimental images (right) are
presented by Chandra and Avedisian (1991). We 45. The size of the last frame is reduced.
YANG GE AND LIANGSHIH FAN 44
droplet (frames 3 and 4), which leads to the bouncing of the droplet up from the
surface (frame 5). The photograph at 8.0 ms shows that the tip of the elongated
droplet separates from the main body of the droplet and the main body of the
droplet then breaks up into smaller secondary drops (frame 5). This
phenomenon was reproduced accurately in the simulation.
Fig. 15 shows the detailed structure of the droplet from a viewing angle of
601. Experimental images show that a hole is formed in the center of the droplet
for a short time period (3.44.8 ms) and the center of the liquid droplet is a dry
circular area. The simulation also shows this hole structure although a minor
variation exists over the experimental images. As the temperature of the surface
is above the Leidenfrost temperature of the liquid, the vapor layer between the
droplet and the surface diminishes the liquidsolid contact and thus yields a low
surfacefriction effect on the outwardly spreading liquid ow. When the droplet
periphery starts to retreat due to the surfacetension effect, the liquid in the
droplet center still ows outward driven by the inertia, which leads to the
formation of the hole structure.
The impact process of a 3.8 mm water droplet under the conditions experi
mentally studied by Chen and Hsu (1995) is simulated and the simulation results
are shown in Figs. 16 and 17. Their experiments involve waterdroplet impact
on a heated Inconel plate with Ni coating. The surface temperature in this
simulation is set as 400 1C with the initial temperature of the droplet given
as 20 1C. The impact velocity is 100 cm/s, which gives a Weber number of 54.
Fig. 16 shows the calculated temperature distributions within the droplet and
within the solid surface. The isotherm corresponding to 21 1C is plotted inside
the droplet to represent the extent of the thermal boundary layer of the droplet
that is affected by the heating of the solid surface. It can be seen that, in the
droplet spreading process (07.0 ms), the bulk of the liquid droplet remains at its
initial temperature and the thermal boundary layer is very thin. As the liquid
lm spreads on the solid surface, the heattransfer rate on the liquid side of the
dropletvapor interface can be evaluated by
q
drop
k
l
T
d2
T
d;bulk
d
T
(76)
where d
T
is the thickness of the thermal boundary layer; T
d2
and T
d,bulk
are the
droplet temperature at the surface and in the bulk, respectively. The thermal
boundary layer thickness can be estimated by (PasandidehFord et al., 2001):
d
T
2d
0
Re
0:5
Pr
0:4
(77)
where Re is the Reynolds number of the impinging liquid ow and is dened
by Re rVd
0
=m; and V the liquid lm velocity. At the early stage of the
spreading, V is close to the initial impact velocity of the droplet, and thus, it
gives a thin thermal boundary layer as shown in frames 1 and 2 of Fig. 16. When
SIMULATION OF GAS LIQUID AND GAS LIQUID SOLID FLOW SYSTEMS 45
FIG. 15. Experimental photos (left) by Chandra and Avedisian (1991) and simulated images
(right) of the spreading droplet on surface at 200 1C. The formation of a hole in the center of the
liquid is captured.
YANG GE AND LIANGSHIH FAN 46
the droplet spreads to the maximum extent and starts to recoil, the liquid
velocity diminishes to zero and the thermal boundary layer is disrupted. Until
then, the temperature rise inside the droplet becomes signicant (frames 3 and 4
of Fig. 16). The simulated temperature distribution inside the solid surface is
also shown in Fig. 16, in which the isotherm of 399 1C is chosen to represent the
area of temperature drop during the impact of the droplet.
The simulation of droplet impact shown in Fig. 16 is conducted under per
fectly symmetrical conditions, which is not easy to achieve in the experiments.
FIG. 16. Simulated temperature eld in the liquid and solid phases.
SIMULATION OF GAS LIQUID AND GAS LIQUID SOLID FLOW SYSTEMS 47
As the droplet is released from the nozzle and moves toward the superheated
surface, some uncontrollable factors such as the angle of dropping, obliquity of
the surface, and perturbation in the ambient conditions render it difcult to
maintain a perfectly normal collision between the droplet and solid surface. The
3D simulation of this study is capable of reproducing the imperfect drop
letsurface impact condition represented by asymmetrical collision. Fig. 17
shows the simulated 3D images of the droplet under the same impact condition
as that shown in Fig. 16 but with a 5 cm/s tangential velocity. The simulated
solid surface temperature is shown in Fig. 18 with comparisons to the experi
mental measurements of Chen and Hsu (1995). It can be seen that with a
small obliquity, there is a signicant effect on the movement behavior of the
droplet. Specically, in this case, the droplet moves away from the impact center
during the recoiling process, which leads to a faster recovery of the solid surface
temperature.
The simulations are further conducted under the experimental conditions of
Inada et al. (1985). In their experiments, 4.0 mm water droplets impact on a
heated platinum surface at a temperature up to 420 1C. The subcooling degree
FIG. 17. Droplet impacts on the at surface with a small tangential velocity. Other conditions are
the same as those in Fig. 16.
YANG GE AND LIANGSHIH FAN 48
of the droplet (dT
sub
) can be varied from 2 1C to 88 1C. The droplet falls down
20 mm before it impacts on the surface, thus the velocity of the impact can be
estimated as 64 cm/s, which gives a Weber number of about 25. In this simu
lation, the initial temperature of surface is xed on 420 1C, which ensure
the nonwetting contact between the droplet and the surface. Fig. 19 shows the
simulated solid temperature compared with the measured value at two loca
tions. T
1
and T
2
were measured at 0.28 and 0.74 mm depth beneath the surface
(Inada et al., 1985). Simulated surface temperatures agreed well with the meas
ured T
1
and T
2
; both decrease until 1314 ms after the impact. The simulated
temperature at the surface of the solid (T
w
) is also shown in Fig. 19. The
maximum temperature drop at the surface is about 60 1C, which occurs at about
13 ms after the impact.
C. SIMULATION OF DROPLET PARTICLE COLLISION IN THE LEIDENFROST REGIME
An efcient numerical model to describe the unsteady, 3D uid ow during
the dropletparticle collision with evaporation is developed (Ge and Fan, 2005).
From the numerical point of view, the droplet and solid surface need to be
0 5 10 15 20 25 30 35 40 45 50
0
100
200
300
400
500
at z= 0.00mm
at z=0.12mm
at z=0.24mm
at z=0.36mm
Experiment of Chen and Hsu(1995)
t
e
m
p
e
r
a
t
u
r
e
o
f
s
o
l
i
d
s
u
r
f
a
c
e
a
t
r
=
0
(
C
)
Time (ms)
FIG. 18. Simulated solid surface temperatures compared with the experiments of Chen and Hsu
(1995). The droplet impacts on the at surface with a small tangential velocity (5 cm/s) as shown in
Fig. 17.
SIMULATION OF GAS LIQUID AND GAS LIQUID SOLID FLOW SYSTEMS 49
captured in the ow eld and the solidow boundary condition at the particle
surface needs to be satised. The numerical method adopted is a combination of
the levelset approach and the IBM. The evaporation effect is accounted for by
the vapor pressure force, which is calculated in a dynamics vaporow model.
Energy balance equations in each phase are solved with interface boundary
condition to give the temperature distribution and evaporation rate.
1. Hydrodynamic Model with LevelSet Method
In this model, two levelset functions (f
d
, f
p
) are dened to represent the
droplet interface (f
d
) and the moving particle surface (f
p
), respectively. The free
surface of the droplet is taken as the zero in the droplet levelset function
f
d
~x; t, and the advection equation (Eq. (3)) of the droplet levelset function
(fd) is solved to capture the motion of the droplet surface. The particle levelset
function (fp) is dened as the signed distance from any given point ~x in the
Eulerian system to the particle surface:
f
p
~x ~x
0
t
R
p
(78)
where ~x
0
t is the position vector of the center of the particle and R
p
the particle
radius. With this denition, f
p
40 when ~x is outside the particle, f
p
o0 when ~x
is inside the particle.
0 5 10 15 20
350
360
370
380
390
400
410
420
430
440
Simulated T
W
Simulated T
1
Measured T
1
Simulated T
2
Measured T
2
S
u
r
f
a
c
e
t
e
m
p
e
r
a
t
u
r
e
(
C
)
time after impact (ms)
FIG. 19. Simulated solid surface temperatures compared with the experiments of Inada et al.
(1985). T
1
and T
2
are the temperatures at locations inside the surface with Z
1
0.28 mm and
Z
2
0.74 mm.
YANG GE AND LIANGSHIH FAN 50
The momentum equation for this 3phase ow system in the Eulerian frame
can be given by
r
@
~
V
@t
r
~
V
~
V
!
rp r~g r 2m
~
D skf
d
df
d
rf
d
~
F
p
f
p
~
F
vapor
79
In this equation, the presence of the solid particle in the uid is represented by
a virtual boundary body force eld,
~
F
p
f
p
, dened by the IBM which will be
discussed in Section IV.C.2.
~
F
vapor
is vapor pressure force exerting on the
dropletparticle contact area due to the effect of the evaporation, which will be
discussed in vaporlayer model of Section IV.C.3.
2. ImmersedBoundary/LevelSet Method for Particle Flow Interaction
In the IBM, the presence of the solid boundary (xed or moving) in the uid
can be represented by a virtual body force eld
~
F
p
f
p
applied on the compu
tational grid at the vicinity of solidow interface. Considering the stability and
efciency in a 3D simulation, the direct forcing scheme is adopted in this model.
Details of this scheme are introduced in Section II.B. In this study, a new velocity
interpolation method is developed based on the particle levelset function (f
p
),
which is shown in Fig. 20. At each time step of the simulation, the uidparticle
boundary condition (noslip or freeslip) is imposed on the computational cells
located in a small band across the particle surface. The thickness of this band can
be chosen to be equal to 3D, where D is the mesh size (assuming a uniform mesh
is used). If a grid point (like p and q in Fig. 20), where the velocity components of
the control volume are dened, falls into this band, that is
Dx4f
p
4Dx (80)
the velocity at these points will be redened using linear interpolation based on
the velocity and particle levelset function (f
p
) at the neighboring grid point. At
grid point p in Fig. 20, which is located in the band and outside the particle
surface, the uid velocity is determined by (for noslip boundary condition):
U
p
V
p
t
f
p
n
X
n
i
U
p
0
;i
V
p
t
f
p
0
;i
(81)
where U
p
0 and f
p
0 are the uid velocity and particle levelset function at grid
point p
0
(in Fig. 20). It should be noted that only the neighboring grid points
located outside the band are chosen to interpolate the velocity at point p. The
velocities at these points (p
0
) are obtained by solving the NS equation like other
points far from the interface. n is the total number of these neighboring points. It
can be seen that when f
p
0, meaning that the grid point p is right on the
particle surface, the velocity at p is equal to the particle velocity: U
p
V
p
(t).
SIMULATION OF GAS LIQUID AND GAS LIQUID SOLID FLOW SYSTEMS 51
The velocity at grid point q (in Fig. 20), which is located inside the small band
and on the particle side, is determined by interpolating the uid velocity at the
neighboring grid points outside the particle surface:
U
q
V
p
t
f
q
m
X
m
i
U
p;i
V
p
t
f
p;i
(82)
where m is the total number of neighboring grid points which are located out
side the particle surface.
3. Vapor Layer and HeatTransfer Model
As the vapor ows in the direction along the spherical surface of the particle,
a boundary layer coordinate (x, l, o) given in Fig. 21 is employed to describe
the vaporlayer equation. In this coordinate, the continuity and momen
tum equations for incompressible vapor ows with gravitation terms neglected
P
p
q
q
P
P
Up
Up
x
x
Particle
interface
Fluid
FIG. 20. Velocity interpolation scheme based on the particle levelset function.
YANG GE AND LIANGSHIH FAN 52
are given by
@u
x
@x
cotx=R
R
u
x
2
R
u
l
@u
l
@l
0 (83)
u
x
@u
x
@x
u
l
@u
x
@l
u
x
u
l
R
@
@x
P
r
n
@
2
u
l
@x@l
2
R
@u
x
@l
@
2
u
x
@l
2
(84)
u
x
@u
l
@x
u
l
@u
l
@l
u
2
x
R
@
@l
P
r
n
R
R
@
2
u
l
@x
2
@u
x
@x
R
@
2
u
x
@x@l
cot
x
R
@u
l
@x
u
x
R
@u
x
@l
!
85
where u
x
, u
l
are the vaporow velocities in x and l direction, respectively.
Based on the assumptions and the order of magnitude analysis described in
Section IV.A.2, and following the similar procedure of solution, the averaged
vaporow velocity is given by
u
x
x O
Z
1
0
F
d
dZ
1
2
u
ld
x
d
2
12g
1
3
20
Re
d
@
@x
p
r
(86)
The vapor continuity equation can be expressed by
u
x
x
1
xdx
Z
x
0
x
0
u
ld
x
0
dx
0
(87)
P
O
Particle surface
p,i
P
i
i
Particle Surface

d,i,
Droplet Surface
i
P
i+1
P
i1
i1
i+1
i1 i
i+3
i+1
i+2
i+2
FIG. 22. Scheme for determining the thickness of the vapor layer.
YANG GE AND LIANGSHIH FAN 54
balance at the interface. Inside the particle, the heatconduction equation is
@T
@t
a
s
r rT (89)
where a
s
is the thermal diffusivity of particle. When the particle surface is in
contact with the droplet through the vapor layer, the boundary condition at the
particle surface is given by
k
v
T
ss
T
ds
d
k
s
~n
p
r
p
T
s
(90)
where ~n
p
is the normal vector of the particle surface, r
p
T the temperature
gradient which is evaluated only on the particle side. This boundary condition,
Eq. (90) is imposed on the computational cells, which are located in a small
band near the particle surface on the solid side.
4. Results and Discussion
Dropletparticle collision conditions under which experiments are conducted
are simulated in this study. In the experiment, the brass particle is heated on
a heating plate with adjustable temperature settings, and a highspeed camera
capable of capturing 500 frames per second will be used to record the drop
letparticle collision process. A droplet is formed by the use of the syringe with
various needle sizes. In the experiments, acetone droplets with 1.62.2 mm di
ameter impact onto the brass particles of sizes 5.5 mm or 3.6 mm.The particle
temperature is 200 1C300 1C, which is much higher than the boiling point of
acetone and ensures that the contact is nonwetting.
The simulation is conducted on uniform Cartesian meshes (Dx Dy
Dz D). The mesh sizes in this study are chosen to give a resolution of
CPR 15. Fig. 23 shows the comparison of the photographs with simulated
images for a 2.1 mm acetone droplet impact onto a 5.5 mm brass particle of
250 1C. The impact velocity is 45 cm/s. The initial temperature of the droplet is
close to the boiling point of acetone (56 1C). A comparison of the images shows
that the impact process is predicted well by this model. Similar to the impact on a
at surface, the droplet spreads on the particle surface at the rst stage
(05.5 ms), then recoils due to the vapor pressure force and surface tension force
(7.513.5 ms), and eventually rebounds away from the particle (17.527.5 ms).
The droplet lm reaches the maximum extent at about 7.0 ms, by which time the
radius of dropletparticle contact area is about 1.5 times that of the original
radius of the droplet. The total contact time is about 17 ms, which is relevant to
the rstorder vibration period of the oscillating drop.
Fig. 24 shows the collision process between a moving particle (1.5 mm in
diameter) and a large water droplet. A water droplet 2.5 mm in diameter has
an initial velocity of 25 cm/s and an initial temperature of 100 1C. The initial
velocity and temperature of the particle are 25 cm/s and 400 1C, respectively.
SIMULATION OF GAS LIQUID AND GAS LIQUID SOLID FLOW SYSTEMS 55
The physical properties of the particle are the same as FCC particles. The
simulation is conducted using a 140 140 200 rectangular mesh covering a
7 mm 7 mm10 mm computational domain. Both the 3D images and the
temperature eld are shown in this gure.
It can be seen in this gure that since the inertia (mass) of the particle is
smaller than that of the droplet, the particle velocity decreases rapidly as soon as
the particle collides with the droplet. After the collision, the surface tension of
0.5ms
1.5ms
13.5ms
5.5ms
17.5ms
21.5ms
27.5ms 7.5ms
0.6
0.5
0.4
0.3
0.2
0.1
0
0 0.1 0.2 0.3 0.4
0.5 0.6 0
0.5
0.6
0.5
0.4
0.3
0.2
0.1
0
0 0.1 0.2 0.3 0.4
0.5 0.6 0
0.5
0.6
0.5
0.4
0.3
0.2
0.1
0
0 0.1 0.2 0.3 0.4
0.5 0.6 0
0.5
0.6
0.5
0.4
0.3
0.2
0.1
0
0 0.1 0.2 0.3 0.4
0.5 0.6 0
0.5
0.6
0.5
0.4
0.3
0.2
0.1
0
0 0.1 0.2 0.3 0.4
0.5 0.6 0
0.5
0.6
0.5
0.4
0.3
0.2
0.1
0
0 0.1 0.2 0.3 0.4
0.5 0.6 0
0.5
0.6
0.5
0.4
0.3
0.2
0.1
0
0 0.1 0.2 0.3 0.4
0.5 0.6 0
0.5
0.6
0.5
0.4
0.3
0.2
0.1
0
0 0.1 0.2 0.3 0.4
0.5 0.6 0
0.5
FIG. 23. Experimental photos (left) and simulated images (right) of the 2.1 mm acetone droplet
impact on 5.5mm particle at 250 1C. Impact velocity V 45 cm/s.
YANG GE AND LIANGSHIH FAN 56
0.8
0.7
t = 0.0ms
t = 10ms
t = 14ms
t = 17ms
t = 24ms
0.6
0.5
0.4
0.3
0.2
0.2
0.4 0.6 0.8
1 1.2
0
0.2
0.4
0.6
0.1
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.2
0.4 0.6 0.8
1 1.2
0
0.2
0.4
0.6
0.1
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.2
0.4 0.6 0.8
1 1.2
0
0.2
0.4
0.6
0.1
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.2
0.4 0.6 0.8
1 1.2
0
0.2
0.4
0.6
0.1
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.2
0.4 0.6 0.8
1 1.2
0
0.2
0.4
0.6
0.1
0.7
400
350
300
250
200
150
100
50
0
0.6
0.5
0.4
0.3
0.2
0.1
0
0 0.2 0.4 0.6 0.8 1
0.7
400
350
300
250
200
150
100
50
0
0.6
0.5
0.4
0.3
0.2
0.1
0
0 0.2 0.4 0.6 0.8 1
0.7
400
350
300
250
200
150
100
50
0
0.6
0.5
0.4
0.3
0.2
0.1
0
0 0.2 0.4 0.6 0.8 1
0.7
400
350
300
250
200
150
100
50
0
0.6
0.5
0.4
0.3
0.2
0.1
0
0 0.2 0.4 0.6 0.8 1
0.7
400
350
300
250
200
150
100
50
0
0.6
0.5
0.4
0.3
0.2
0.1
0
0 0.2 0.4 0.6 0.8 1
FIG. 24. Particle and droplet collision at the same initial velocity. The images on the right show
the temperature eld.
SIMULATION OF GAS LIQUID AND GAS LIQUID SOLID FLOW SYSTEMS 57
the droplet and the vaporlayer pressure induced by the evaporation drive the
particle to the inverse direction. Since the Weber number of the collision is
not very large (We 15), the droplet only undergoes small deformation without
splashing. After collision, the droplet still moves along its original path with
a decreased velocity, while the rebounding velocity of the particle is larger than
that of the droplet. It can also be found that the temperature drop of the particle
is not signicant in this condition.
V. Concluding Remarks
The most signicant advantage of the levelset method and the IBM in solv
ing 3D, threephase ow problems is that the governing equations can con
veniently be discretized and solved on xed, regular structured grids, rather than
resorting to the classic bodytted mesh approach. These methods are efcient
in utilizing computational resources while retaining the accuracy of the com
putational results. The levelset method implicitly captures the motion of the
interface by solving the advection equation of the levelset function, and can be
easily implemented for 3D interface tracking. The levelset methods are par
ticularly effective in handling ow problems that involve changes in the topo
logy of evolving interfaces, and in which the speed of the interface is sensitive to
local surface geometry, such as curvature and normal vector. Although, like
other frontcapture techniques, the levelset method may suffer from some
problems associated with the preservation of mass conservation (Puckett et al.,
1997), the method is able to accurately compute interfacial ows with surface
tension force and other complex physical forces acting on the front. In the
conventional IBM, the moving front is represented by a nite number of dis
crete Lagrangian markers along the interface, which move in the ow. In
general, this arrangement has the property of preservation of the clear interface
position and hence mass conservation; however, with this arrangement, it is
difcult to trace the interface with complicated shape and topological change in
three dimensions. The combination of the levelset method and the IBM as
described in this study can allow accuracy in interface movement, as well as
conservation in moving object mass when simulating the complex threephase
interaction systems.
In system 1, the 3D dynamic bubbling phenomena in a gasliquid bubble
column and a gasliquidsolid uidized bed are simulated using the levelset
method coupled with an SGS model for liquid turbulence. The computational
scheme in this study captures the complex topological changes related to the
bubble deformation, coalescence, and breakup in bubbling ows. In system 2,
the hydrodynamics and heattransfer phenomena of liquid droplets impacting
upon a hot at surface and particle are analyzed based on 3D levelset method
and IBM with consideration of the lmboiling behavior. The heat transfers in
YANG GE AND LIANGSHIH FAN 58
each phase are solved using a microscale vaporow model which is applied to
determine the vapor pressure force during the contact process between the
droplet and the superheated surface. The simulation model developed in this
study is capable of reproducing the droplet impaction process with signicant
heat transfer and phase change (evaporation). Both the saturated impact and
the subcooling impact are considered. The simulation results are found to be in
good agreement with the experimental measurements of the droplet deforma
tion process and the surfacetemperature variation.
Nomenclature
NOTATION
C
D
modied drag coefcient
C
D0
drag coefcient
C
s
Smagorinsky coefcient
~
D stress tensor
d distance function (or diameter)
d
p
diameter of particle
f
am
added mass force
f
d
uidparticle interaction force
f
dr
drag force
f
s
surface tension force
g gravitational acceleration
H
b
heaviside function
k heat conductivity
Kn Knudsen number
l length scale
L
c
latent heat
m
p
mass of particle
_ m mass evaporation rate
n normal vector
p pressure
Pr Prandtl number
q heat ux (or virtual mass source)
R radius of the droplet
Re particle Reynolds number
S shear strain rate
T temperature
t time
U velocity scale of the vapor ow
u velocity of the vapor ow
SIMULATION OF GAS LIQUID AND GAS LIQUID SOLID FLOW SYSTEMS 59
V velocity (or volume)
We Weber number
x position vector
GREEK LETTERS
a thermal diffusivity
D grid size
G gasliquid interface
F function dened in vaporlayer model (4.7)
O cell volume (or function dened in vaporlayer model (4.7))
b half of the thickness of the interface
d d function (or the thickness of the vaporlayer or thermal boundary
layer)
e void fraction
f levelset function
k curvature of the surface
l constant dened in Eq. (24) (or coordinate dened in vaporlayer
model)
m molecular viscosity
n
T
kinematic turbulent viscosity
r density
s surface tension (or thermal accommodation coefcient)
t viscous stress tensor, articial time
x coordinate dened in vaporlayer model
Z dimensionless variable dened in Eq. (46)
n kinematic viscosity of vapor ow
SUBSCRIPTS AND SUPERSCRIPTS
b boundary
d droplet
g gas phase
l liquid phase
o initial
p particle
r radial direction
s solid
v vapor phase
YANG GE AND LIANGSHIH FAN 60
x vaporlayer coordinate
l vaporlayer coordinate
G gasdroplet interface
ij cell index
k particle index
sa saturated
sg subgrid scale
REFERENCES
Agresar, G., Linderman, J. J., Triggvason, G., and Powell, K. G. J. Comp. Phys. 143, 346 (1998).
Amsden, A. A., and Harlow, F. H. J. Comp. Phys. 3, 80 (1968).
Anderson, T. B., and Jackson, R. Ind. Eng. Chem. Fundam. 6, 525539 (1967).
Biance, A. L., Clanet, C., and Quere, D. Phys. Fluids 15(6), 16321637 (2003).
Bird, G. A., Molecular Gas Dynamics. Clarendon Press, Oxford, UK (1976).
Brackbill, J. U., Kothe, D. B., and Zemach, C. J. Comp. Phys. 100, 335 (1992).
Bussmann, M., Mostaghimi, J., and Chandra, S. Phys. Fluids 11(6), 14061417 (1999).
Bussmann, M., Mostaghimi, J., and Chandra, S. Phys. Fluids 12(12), 31213132 (2000).
Chandra, S., and Avedisian, C. T. Proc. R. Soc. Lond. A 432, 1341 (1991).
Chen, C., and Fan, L. S. AIChE J 50, 288301 (2004).
Chen, J. C., and Hsu, K. K. J. Heat Transfer 117, 693697 (1995).
Chen, Y. M., and Fan, L. S. Chem. Eng. Sci. 44, 2762 (1989a).
Chen, Y. M., and Fan, L. S. Chem. Eng. Sci. 44, 117 (1989b).
Crowe, C. T., Sharma, M. P., and Stock, D. E. J. Fluids Eng. 99, 325 (1977).
Crowe, C., Sommerfeld, M., and Tsuji, Y., Multiphase Flows with Droplets and Particles. CRC
Press, NY (1998).
Dandy, D. S., and Leal, L. G. J. Fluid Mech. 208, 161 (1989).
Delnoij, E., Kuipers, J. A. M., and Van Swaaij, W. P. M. Chem. Eng. Sci. 52, 3623 (1997).
Fadlun, E. A., Verzicco, R., Orlandi, P., and Yusof, J. M. Comp. Phys. 161, 3560 (2000).
Fan, L. S., GasLiquidSolids Fluidization Engineering. Butterworths, Stoneham, MA (1989).
Fan, L. S., and Tsuchiya, K., Bubble Wake Dynamics in Liquid and LiquidSolid Suspensions.
ButterworthHeinemann, Stoneham, MA (1990).
Fan, L. S., Lau, R., Zhu, C., Vuong, K., Warsito, W., Wang, X., and Liu, G. Chem. Eng. Sci. 56,
58715891 (2001).
Feng, J., Hu, H. H., and Joseph, D. D. J. Fluid Mech. 261, 95134 (1994a).
Feng, J., Hu, H. H., and Joseph, D. D. J. Fluid Mech. 277, 271301 (1994b).
Feng, Z. G., and Michaelides, E. E. J. Comput. Phys. 202, 2051 (2005).
Fujimoto, H., and Hatta, N. J. Fluids Eng. 118, 142149 (1996).
Fukai, J., Shiiba, Y., Yamamoto, T., and Miyatake, O. Phys. Fluids 7(2), 236 (1995).
Ge, Y., and Fan, L. S. Phys. Fluids 17, 027104 (2005).
Goldstein, D., Handler, R., and Sirovich, L. J. Comp. Phys. 105, 354 (1993).
Gottfried, B. S., Lee, C. J., and Bell, K. J. Int. J. Heat Mass Transfer 9, 11671187 (1966).
Groendes, V., and Mesler, R. Proceedings of the 7th International Heat Transfer Conference,
Munich, Federal Republic of Germany 1982, 1982, pp. 131136.
Harlow, F. H., and Welch, J. E. Phys. Fluid 8, 2182 (1965).
Harlow, F. H., and Shannon, J. P. J. Appl. Phys. 38, 3855 (1967).
Harvie, D. J. E., and Fletcher, D. F. Int. J. Heat Mass Transfer 44, 26332642 (2001a).
Harvie, D. J. E., and Fletcher, D. F. Int. J. Heat Mass Transfer 44, 26432659 (2001b).
SIMULATION OF GAS LIQUID AND GAS LIQUID SOLID FLOW SYSTEMS 61
Harvie, D. J. E., and Fletcher, D. F. J. Heat Transfer 123, 486491 (2001c).
Hatta, N., Fujimoto, H., Kinoshita, K., and Takuda, H. J. Fluids Eng. 119, 692699 (1997).
Hirt, C. W., Amsden, A. A., and Cook, J. L. J. Comput. Phys. 14, 227 (1974).
Hirt, C. W., and Nichols, B. D. J. Comput. Phys. 39, 201 (1981).
Hoomans, B. P. B., Kuipers, J. A. M., Briels, W. J., and van Swaaij, W. P. M. Chem. Eng. Sci. 51,
99118 (1996).
Inada, S., Miyasaka, Y., and Nishida, K. Bull. JSME 28, 26752681 (1985).
Inada, S., Miyasaka, Y., Sakamoto, K., and Hojo, K. J. Chem. Eng. Japan 21, 463 (1988).
Jackson, R., The Dynamics of Fluidized Particles. Cambridge University Press, NY (2000).
Jamet, D., Lebaigue, O., Coutris, N., and Delhaye, J. M. J. Comput. Phys. 169, 624 (2001).
Joseph, D. D., and Lundgren, T. S. Int. J. Multiphase ow 16, 3542 (1990).
Kanai, A., and Mtyata, H. Numerical simulation of bubbles in a boundary layer by Maker
DensityFunction. Proceedings of the 3rd International Conference on Multiphase Flow,
Lion, France (1998).
Karl, A., Anders, K., Rieber, M., and Frohn, A. Part. Part. Syst. Charact. 13, 186191 (1996).
Kashiwa, B. A., Padial, N. T., Rauenzahn, R. M., and Vanderheyden, W. B. Los Alamos National
Laboratory Research Report, LAUR933922 (1994).
Kim, J., Kim, D., and Choi, H. J. Comput. Phys. 171, 132150 (2001).
Kothe, D. B., and Rider, W. J. Los Alamos National Laboratory Research Report, LAUR943384
(1995).
Kunii, D., and Levenspiel, O., Fluidization Engineering. 2nd ed ButterworthHeinemann, Boston
(1991).
Lai, M. C., and Peskin, C. S. J. Comput. Phys. 160, 705719 (2000).
Lapin, A., and Lu bbert, A. Chem. Eng. Sci. 49, 3661 (1994).
Li, Y., Zhang, J., and Fan, L. S. Chem. Eng. Sci. 54, 5101 (1999).
McHyman, J. Physica D 12, 396 (1984).
MehdiNejad, V., Mostaghimi, J., and Chandra, S. Phys. Fluids 15(1), 173183 (2003).
Mittal, R., and Iaccarino, G. Annu. Rev. Fluid Meth. 37, 239261 (2005).
Monaghan, J. J. Comput. Phys. 110, 399 (1994).
Mudde, R. F., and Simonin, O. Chem. Eng. Sci. 54, 5061 (1999).
Osher, S., and Sethian, J. A. J. Comput. Phys. 79, 12 (1988).
PasandidehFord, M., Bhola, R., Chandra, S., and Mostaghimi, J. Int. J. Heat Mass Transfer 41,
29292945 (1998).
PasandidehFord, M., Aziz, S. D., Chandra, S., and Mostaghimi, J. Int. J. Heat Fluid Flow 22, 201
(2001).
Peskin, C. S. J. Computat. Phys. 25, 220 (1977).
Puckett, E. G., Almgren, A. S., Bell, J. B., Marcus, D. L., and Rider, W. J. J. Computat. Phys. 100,
269 (1997).
Qiao, Y. M., and Chandra, S. Int. J. Heat Mass Transfer 39(7), 13791393 (1996).
Rowe, P. N., and Henwood, G. A. Part. 1 Trans. Inst. Chem. Eng 39, 43 (1961).
Sato, T., and Richardson, S. M. Int. J. Numer. Meth. Fluids 19, 555 (1994).
Scardovelli, R., and Zeleski, S. Ann. Rev. Fluid Mechanics 31, 567 (1999).
Sethian, J. A., and Smereka, P. Annu. Rev. Fluid Meth. 35, 341372 (2003).
Smagorinsky, J. Mon. Weather Rev. 91, 99 (1963).
Sokolichin, A., and Eigenberger, G. Chem. Eng. Sci. 49, 5735 (1994).
Sokolichin, A., and Eigenberger, G. Chem. Eng. Sci. 54, 2273 (1999).
Sussman, M., Smereka, P., and Osher, S. J. Comput. Phys. 114, 146 (1994).
Sussman, M., Fatemi, E., Smereka, P., and Osher, S. Computers Fluids 114, 146 (1998).
Sussman, M., and Fatemi, E. SIAM J. Sci. Comput. 20, 1165 (1999).
Tryggvason, T., Bunner, B., Esmaeeli, A., Juric, D., AlRawahi, N., Tauber, W., Han, J., Nas, S.,
and Jan, Y. J. J. Comput. Phys. 169, 708759 (2001).
Tsuji, Y., Kawaguchi, T., and Tanaka, T. Powder Technol 77, 7981 (1993).
YANG GE AND LIANGSHIH FAN 62
Udaykumar, H. S., Kan, H. C., Shyy, W., and TranSonTry, R. J. Comput. Phys. 137, 366405
(1997).
Unverdi, S. O., and Tryggvason, G. J. Computat. Phys. 100, 25 (1992a).
Unverdi, S. O., and Tryggvason, G. Physica D 60, 70 (1992b).
Wachters, L. H. J., and Westerling, N. A. J. Chem. Eng. Sci. 21, 10471056 (1966).
Wen, C. Y., and Yu, Y. H. Chem. Eng. Prog. 62, 100 (1966).
Yabe, T. Interface Capturing and Universal Solution of Solid, Liquid and Gas by CIP Method.
Proceedings of the HighPerformance Computing of MultiPhase Flow, Tokyo, July 1819,
1997.
Ye, T., Mittal, R., Udaykumar, H. S., and Shyy, W. J. Computat. Phys. 156, 209240 (1999).
Yusof, J. M. Combined immersed boundaries/Bsplines methods for simulations of ows in complex
geometries. CTR Annual Research Briefs, NASA Ames/Stanford University (1997).
Zenit, R., and Hunt, M. J. Fluids Eng. 121, 179 (1999).
Zhang, D. Z., and Prosperetti, A. J. Fluid Mech. 267, 185219 (1994).
Zhang, J., Fan, L. S., Zhu, C., Pfeffer, R., and Qi, D. Powder Technol 106, 98 (1999).
Zhang, Z., and Prosperetti, A. J. Appl. Mech.Trans. ASME 70, 6474 (2003).
Zhu, C., Wang, X., and Fan, L. S. Powder Technol 111, 7982 (2000).
SIMULATION OF GAS LIQUID AND GAS LIQUID SOLID FLOW SYSTEMS 63
MULTISCALE MODELING OF GASFLUIDIZED BEDS
M.A. van der Hoef
1
, M. Ye
1
, M. van Sint Annaland
1
, A.T. Andrews IV
2
,
S. Sundaresan
2
and J.A.M. Kuipers
1,
1
Department of Science & Technology, University of Twente, 7500 AE Enschede,
The Netherlands
2
Department of Chemical Engineering, Princeton University, Princeton, NJ 08544, USA
I. Introduction 66
A. GasFluidized Beds 66
B. Numerical Models for Gas and Solid Flows 67
C. The Multi Level Modeling Approach for GasSolid
Flows 70
II. Lattice Boltzmann Model 74
A. From LatticeGas to LatticeBoltzmann 74
B. The Lattice BhatnagarGrossKrook Model 78
C. Modeling Solid Particles 81
D. Results for the GasSolid Drag Force 83
III. Discrete Particle Model 86
A. Introduction 86
B. Particle Dynamics: The SoftSphere Model 89
C. Gas Dynamics 100
D. Interphase Coupling 102
E. Energy Budget 106
F. Results for the Excess Compressibility 107
IV. TwoFluid Model 111
A. Introduction 111
B. GOVERNING Equations 113
C. General Kinetic Theory 115
D. Kinetic Theory of Granular Flow 119
E. Numerical Solution Method 120
F. Application to Geldart A Particles 127
V. Towards IndustrialScale Models 131
A. The Limits of the TwoFluid Model 132
B. StateoftheArt on Dealing with Unresolved Structures 135
C. A Different Approach: The Discrete Bubble Model 141
VI. Outlook 143
References 146
Corresponding author. Tel: +31 53 489 3039; Fax: +31 53 489 2882 Email: j.a.m.kuipers@tnw.
utwente.nl
65
Advances in Chemical Engineering, vol. 31
ISSN 00652377
DOI 10.1016/S00652377(06)310022
Copyright r 2006 by Elsevier Inc.
All rights reserved
Abstract
Numerical models of gasuidized beds have become an important
tool in the design and scale up of gassolid chemical reactors. However, a
single numerical model which includes the solidsolid and soliduid in
teraction in full detail is not feasible for industrialscale equipment, and
for this reason one has to resort to a multiscale approach. The idea is
that gassolid ow is described by a hierarchy of models at different
length scales, where the particleparticle and uidparticle interactions
are taken into account with different levels of detail. The results and
insights obtained from the more fundamental models are used to develop
closure laws to feed continuum models which can be used to compute the
ow structures on a much larger (engineering) scale. Our multiscale
approach involves the lattice Boltzmann model, the discrete particle
model, and the continuum model based on the kinetic theory of granular
ow. In this chapter we give a detailed account of each of these models as
they are employed at the University of Twente, accompanied by some
illustrative computational results. Finally, we discuss two promising ap
proaches for modeling industrialsize gasuidized beds, which are cur
rently being explored independently at the Princeton University and the
University of Twente.
I. Introduction
A. GASFLUIDIZED BEDS
Gasuidized beds consist of ne granular material (usually smaller than
5 mm) that are subject to a gas ow from below, large enough so that the gas
drag on the particles can overcome the gravity and the particles can uidize.
When in the uidized state, the moving particles work effectively as a mixer
resulting in a uniform temperature distribution and a high mass transfer rate,
which are benecial for the efciency of many physical and chemical processes.
For this reason, gasuidized beds are widely applied in the chemical, petro
chemical, metallurgical, environmental, and energy industries in largescale op
erations involving adhesion optimized coating, granulation, drying, and synthesis
of fuels and base chemicals (Kunii and Levenspiel, 1991). Lack of understanding
of the fundamentals of dense gasparticle ows in general has led to severe
difculties in the design and scaleup of these industrially important gassolid
contactors (van Swaaij, 1985). In most cases, the design and scaleup of uidized
bed reactors is a fully empirical process based on preliminary tests on pilotscale
model reactors, which is a very time consuming and thus expensive activity.
Clearly, computer simulations can be a very useful tool to aid this design
and scaleup process. Basically, such simulations can be used for two different
M.A. VAN DER HOEF ET AL. 66
purposes. Firstly, to contribute to our understandingthat is, the simulations
are used to obtain a fundamental insight into the complex dynamic behavior of
dense uidized suspensions, which should lead to an understanding based on
elementary physical principles such as drag, friction, dissipation, etc.; this also
includes the testing of elementary assumptions in theoretical models, such as the
Maxwell velocity distribution of the particles. Secondly, the simulations can be
used as a design tool, where the ultimate goal is to have a numerical model with
predictive capabilities for the dense gasparticle ows encountered in engineer
ingscale equipment. Clearly, it will not be possible to have one single simulation
method that can achieve this, but one rather needs a hierarchy of methods for
modeling the ow phenomena on different length and time scales.
Obviously, these two items are not strictly separated; in contrast, the most
fruitful approach is when they are simultaneously followed, so that they can
mutually benet from each other. In this chapter, we want to focus on the use of
simulation methods as a design tool for gasuidized bed reactors, for which we
consider gassolid ows at four distinctive levels of modeling. However, before
discussing the multilevel scheme, it is useful to rst briey consider the numer
ical modeling of the gas and solid phase separately.
B. NUMERICAL MODELS FOR GAS AND SOLID FLOWS
1. Gas Phase
The description of a gas ow is well established from the micro to macro
scales. The length scale of a gas ow can be characterized by the local Knudsen
number Kn, which is dened as
Kn
l
L
where l is the mean free path of the molecules and L is the characteristic length
scale of the ow. The models to describe a gas ow are schematically shown in
Fig. 1. For largescale systems with Kno0.01, the gas ow can be described by
ordinary uid dynamics where the macroscopic elds (such as density and ve
locity) are formulated by NavierStokes equations in a threedimensional (3D)
coordinate space, together with noslip boundary conditions. A number of well
developed numerical algorithms and meshing techniques in computational uid
dynamics (CFD) can be used to handle very complex geometries (Anderson,
1995). If the system becomes smaller, say 0.01oKno0.1, the NavierStokes
equations still hold, but caution must be exercised for the boundary conditions
because partial slip might exist between the gassolid interfaces. For a rareed
gas where Kn40.1, the continuum assumption breaks down, and the socalled
kinetic theory of (dense) molecular gases should be applied. Kinetic theory
differs from the ordinary uid dynamics as there is just one eld (the density
of molecules) in the phase space. The basic equation in kinetic theory, in the
MULTISCALE MODELING OF GASFLUIDIZED BEDS 67
simplest form, is the Boltzmann equation that describes the evolution of the
density function f in a sixdimensional (6D) space (three coordinates and three
velocity components) (Chapman and Cowling, 1970). At this scale, computa
tional techniques such as molecular dynamics (MD) (Allen and Tildesley, 1990)
and direct simulation Monte Carlo (DSMC) approach (Bird, 1976) can be very
efcient. In these techniques, the motion of molecules is traced on an individual
basis. Gas pressure and other transport coefcients, including gas viscosity and
thermal conductivity, are obtained by methods from statistical mechanics
(Chapman and Cowling, 1970). Molecules can be treated either as hard spheres
or points with certain interaction potentials, depending on their physical prop
erties. In the extreme case where the mean free path is very large compared to the
system sizes (i.e., Kn410), the molecules move freely and just collide with the
walls. This is the limit of free molecular ow, where the system behaves as an
ideal gas.
Clearly, there are two quite different types of models for a gas ow: the
continuum models and the molecular models. Although the molecular models
can, in principle, be used to any length scale, it has been almost exclusively
applied to the microscale because of the limitation of computing capacity at
present. The continuum models present the main stream of engineering appli
cations and are more exible when applying to different macroscale gas ows;
however, they are not suited for microscale ows. The gap between the con
tinuum and molecular models can be bridged by the kinetic theory that is based
on the Boltzmann equation.
2. Solid Phase
The methods used for modeling pure granular ow are essentially borrowed
from that of a molecular gas. Similarly, there are two main types of models: the
continuous (Eulerian) models (Dufty, 2000) and discrete particle (Lagrangian)
models (Herrmann and Luding, 1998; Luding, 1998; Walton, 2004). The con
tinuum models are developed for largescale simulations, where the controlling
equations resemble the NavierStokes equations for an ordinary gas ow. The
discrete particle models (DPMs) are typically used in smallscale simulations or
FIG. 1. The various levels of modeling gas ow.
M.A. VAN DER HOEF ET AL. 68
in the investigation of the detailed physics of granular ow. A kinetic theory of
granular ow (KTGF) has also been proposed to connect the microscale picture
of granular ow to the macroscale description (Jenkins and Savage, 1983; Lun
et al., 1984).
However, a granular ow differs signicantly from a molecular gas ow. The
collisions between molecules are elastic, and the kinetic energy is conserved in
isothermal systems. For the molecular gas, there is a welldened equilibrium
state in the absence of external energy sources, and one can dene a thermal
temperature based on the internal kinetic energy. The interaction between
macroscopic particles, however, is far more complicated. The collision between
two macroscopic particles will come with surface friction and elasticplastic
deformation, which leads to the dissipation of kinetic energy. This inelasticity
forms the primary feature of granular ow that differentiates it from a molec
ular gas (Campbell, 1990). Clearly, without any external energy sources, a
granular system will continuously cool down, and an equilibrium state can
never be reached.
To model granular systems, DPMs using the same techniques as MD meth
ods can be used, where it is assumed that the particle motion can be well
described by the Newtonian equations. However, in order to establish a con
tinuum description, a number of serious difculties are encountered when one
tries to describe the elds in phase space. First, an energy source term and a
dissipative term should be included in the Boltzmann equations, which com
plicates the (approximate) solution. Also, particle sizes may show a certain
distribution even for the same type of materials. It is well known that a differ
ence in particle sizes will result in the segregation of granular materials (e.g., the
Brazil nut effect). Furthermore, in most granular ows the effect of gravity
cannot be ignored, which introduces an anisotropy in the velocity uctuation of
particles. Clearly, the denitions of the particlephase pressure and other trans
port coefcients are not straightforward because normally a homogeneous
equilibrium state does not exist. For these reasons, the construction of a reliable
hydrodynamical model for granular ow offers a great challenge for both sci
entists and engineers (Goldhirsch, 1999).
3. The Interphase Coupling
The prime difculty of modeling twophase gassolid ow is the interphase
coupling, which deals with the effects of gas ow on the motion of solids and
vice versa. Elgobashi (1991) proposed a classication for gassolid suspensions
based on the solid volume fraction e
s
, which is shown in Fig. 2. When the solid
volume fraction is very low, say e
s
o10
6
, the presence of particles has a neg
ligible effect on the gas ow, but their motion is inuenced by the gas ow for
sufciently small inertia. This is called oneway coupling. In this case, the gas
ow is treated as a pure uid and the motion of particle phase is mainly con
trolled by the hydrodynamical forces (e.g., drag force, buoyancy force, and so
MULTISCALE MODELING OF GASFLUIDIZED BEDS 69
on), while the particleparticle interaction is assumed to be irrelevant. With
increasing solid volume fraction up to e
s
o10
3
, the effects of the particle phase
on the gasphase ow pattern will become important. In this region, turbulent
structures encountered in gas ows can be modied by the presence of particles.
It is commonly accepted that particleparticle interactions still do not play a
dominant role in this regime, which we normally refer to as twoway cou
pling. For even higher solid volume fractions (e
s
410
3
), the momentum of
particles will be transported not only by the freeight mechanism but also by
the collisions between particles and particles with the conning walls. This
means that the particleparticle interaction will be very important and four
way coupling should be taken into account. Note that it is precisely this dense
particle regime that is important for the industrial applications of twophase
ows. However, a numerical model that includes the solidsolid and soliduid
interactions in full detail is not feasible for industrialscale equipment, and for
this reason one has to resort to a multilevel approach.
C. THE MULTI LEVEL MODELING APPROACH FOR GAS SOLID FLOWS
As mentioned previously, the construction of reliable models for largescale
gassolid contactors is seriously hindered by the lack of understanding of the
fundamentals of dense gasparticle ows (van Swaaij, 1985). In particular, the
phenomena that can be related to the effective gasparticle interaction (drag
forces), particleparticle interactions (collision forces), and particlewall inter
action are not well understood (Kuipers and van Swaaij, 1998; Kuipers et al.,
1998). The prime difculty here is the large separation of scales: the largest ow
structures can be of the order of meters, and yet, these structures are directly
inuenced by details of the particleparticle and particlegas interactions, which
take place on the scale of millimeters, or even micrometers. As shown above, for
both the gas and particle phase, continuum(Eulerian) and discrete(Lag
rangian) type of models can be applied, depending on the length scales involved.
Thus, in order to model gassolid twophase ows at different scales, one can
FIG. 2. Interphase coupling. Based on Elgobashi, Appl. Sci. Res. (1991).
M.A. VAN DER HOEF ET AL. 70
choose appropriate combinations of the gas and solidphase models, where in
all cases a fourway coupling is used either directly or effectively, depending on
the scale of the simulation. The basic idea is that the smaller scale models, which
take into account the various interactions (uidparticle, particleparticle) in
detail, are used to develop closure laws that can represent the effective coarse
grained interactions in the larger scale models. Note that it is not guaranteed
that some subtle correlations between small and largescale processes exist,
which cannot be captured by effective interactions. However, experience has
shown that in many cases the main characteristics of gassolid ows can be well
described by the use of closure relations. In this chapter, we discuss three levels
of modeling: the lattice Boltzmann model (LBM), the DPM and the twouid
model (TFM) based on the KTGF. In Fig. 3, we show a schematic represen
tation of the three models, including the information that is abstracted from the
simulations, which is incorporated in higher scale models via closure relations,
with the aid of experimental data or theoretical results. We will next give a brief
description of each of these models.
1. TwoFluid Model
At the largest scale, a continuum description is employed for both the solid
phase and the gas phase, and a CFDtype Eulerian code is used to describe the
time evolution of the local mass and momentum density of both phases (see
Refs. Kuipers et al., 1992 and Gidaspow, 1994 amongst others). In a more
sophisticated model, based on the KTGF, also the local granular temperature of
the solid phase is a dynamical variable, and thus included in the update. With
modernday computers, the TFM model can predict the ow behavior of
gassolid ows of systems with a linear dimension of the order of 1 m, denoted
as the engineering scale, corresponding typically to the size of pilot plants,
which is in between the laboratory scale (0.1 m) and the industrial scale (10 m).
The TFM relies heavily on closure relations for the effective solid pressure and
viscosity, and gassolid drag. The basic idea of the multiscale modeling is that
FIG. 3. Multilevel modeling scheme.
MULTISCALE MODELING OF GASFLUIDIZED BEDS 71
these relations are obtained from kinetic theory and from numerical data col
lected in the more detailed scale models.
2. Discrete Particle Model
At one level higher in detail (and thus smaller in scale), we have the DPM.
Here the continuum description for the solid phase is replaced by a description
with discrete particles, which are modeled by spheres (Hoomans et al., 1996,
2000). The ow eld is still continuous and updated by the same methods as in
the TFM, where the scale at which the gas ow eld is described is an order of
magnitude larger than the particles (a CFDgrid cell typically contains
O(10
2
)O(10
3
) particles). The motion of the particles is governed by Newtons
law, where the forces on the particles are integrated using standard schemes for
ordinary differential equations (ODEs). These forces follow from the interac
tion with the uid phase and collisions with the other particles. Therefore, both
a dragforce closure and a collision model have to be specied for this level of
modeling. The advantage of this model is that it can account for the parti
clewall and particleparticle interactions in a realistic manner. This model
allows one to validate (and modify) the viscosity and pressure closures derived
from the KTGF, which are used in the TFM simulations. Still, a closure law for
the effective momentum exchange between the two phases has to be specied for
this model. The system sizes that can be studied are of the order of O(10
5
)
particles, which corresponds (for millimetersized particles) to systems that have
a linear dimension of the order of 0.1 m (i.e., laboratory scale).
3. Lattice Boltzmann Model
At the most detailed level of description, the gas ow eld is modeled at scales
smaller than the size of the solid particles. The interaction of the gas phase with
the solid phase is incorporated by imposing stick boundary conditions at the
surface of the solid particles. This model thus allows one to measure the effec
tive momentum exchange between the two phases, which is a key input in all the
higher scale models. A particularly efcient method to solve the ow eld be
tween the spheres is the LBM (Ladd, 1994; Ladd and Verberg, 2001), although
in principle other direct numerical simulation (DNS) techniques can also be
used. The number of particles in such a simulation is typically around 500,
which is sufciently high to account for swarm effects. The goal of these sim
ulations is to construct drag laws for dense gassolid systems. For low Reynolds
numbers (Re), the functional form of the drag law can be derived from theory
using the CarmanKozeny approximation, where the simulation data is then
used to determine the unknown parameters such as the Kozeny constant. For
higher Reynolds numbers, a theoretical evaluation of the functional form is not
possible and the drag law is simply constructed as the best possible t to the
simulation data, where the functional form is dictated by a compromise between
M.A. VAN DER HOEF ET AL. 72
simplicity and accuracy. For both low and high Reynolds numbers, the drag
laws are validated (and possibly adjusted) on the basis of pressuredrop data.
A graphical representation of the multilevel approach is shown in Fig. 4. All
three models are now commonly accepted and are widely used by a number of
research groups (both academic and industrial) around the world. In a recent
paper, we have given an overview of the three models as they are employed at
the University of Twente, together with some illustrative examples (Van der
Hoef et al., 2004). In this chapter, we will focus on the technical details of each
of the models, much of which has not been published elsewhere. The devel
opment of detailed closure relations from the simulations, as indicated in Fig. 3,
is still ongoing. Some preliminary results for both the dragforce closures and
solid pressure will be presented in the Sections II and III. In this chapter, we will
FIG. 4. Graphical representation of the multilevel modeling scheme. The arrows represent a
change of model. On the left is a uidized bed on a lifesize scale, a section of which is modeled by
the twouid model (TFM) (see enlargement), where the shade of grey of a cell indicates the solid
phase volume fraction. On the right, the same section is modeled using discrete particles. The gas
phase is solved on the same grid as in the TFM. The bottom graph shows the most detailed level,
where the gas phase is solved on a grid much smaller than the size of the particles. Note that in
reality, the separation in scales is much more extreme, and also that the section that can be modeled
by the TFM of the industrialscale uidized bed is much smaller.
MULTISCALE MODELING OF GASFLUIDIZED BEDS 73
only consider monodisperse systems, but nevertheless we will formulate the soft
sphere model in Section III.B for general polydisperse systems where a particle a
has an individual radius R
a
. Apart from Section III.B, the size of the particles
are indicated by a single radius R or a diameter d.
Finally, we note that the TFM can simulate uidized beds only at engineering
scales corresponding typically to the size of pilot plants, and the industrialscale
uidized bed reactors (diameter 15 m, height 320 m) are still far beyond its
capabilities. In Section V, we discuss two promising approaches for modeling
largescale gassolid ow, which are currently being explored independently at
the Princeton University and the University of Twente. We stress, however, that
these approaches are still under development and that they should be recognized
as only preliminary.
II. Lattice Boltzmann Model
As mentioned in Section I, there are two fundamentally different types of
models to describe a gas ow: the continuum models and the molecular models.
In principle, the molecular models can be applied at any length scale; however, in
practice this is limited to microscopic scales only because of the limitation of
computer time. The continuum models present the main stream of engineering
applications and are more exible when applying to different macroscale gas
ows. The gap between the continuum and molecular models can be bridged,
however, by the lattice Boltzmann (LB) simulation model that applies at a me
soscopic scale, which is in between fully microscopic and macroscopic scales.
The LB model that is currently the most widely usedthe lattice Bhatn
agarGrossKrook (BGK) modelis nothing but a nite difference version of
the continuous, macroscopic BGK equation introduced in 1954 (Bhatnagar et
al., 1954). Historically, however, this LB model has evolved from the microscopic
latticegas simulation models for uids, and we will also follow this route here.
A. FROM LATTICEGAS TO LATTICEBOLTZMANN
1. LatticeGas Models
As mentioned earlier, in principle, one can model the dynamics of a simple
classical uid by means of MD simulations. This technique, although straight
forward, is relatively timeconsuming, and therefore not suitable for observa
tion of largescale macroscopic phenomena in the uid. However, one often
does not need such a detailed description of the microdynamics as provided by
MD. In such cases, it would be more efcient to strip the MD model down to its
barest essentials, where the only requirement is that the model behaves like a
uid macroscopically, but is still atomistic in characteri.e., the mechanism
underlying the uid motion is the movement of particles. From the derivation of
M.A. VAN DER HOEF ET AL. 74
the uid dynamics equations it is clear that a key ingredient in such a model
must be local conservation of mass and momentum. The simplest model to
think of would be one with a single species of particles moving on a lattice with
discrete velocities. In 1985, such a model was introduced for twodimensional
(2D) uid ow by Frisch et al. (1986), where particles move on a triangular
lattice. Such a lattice has just enough symmetry to guarantee isotropic, mac
roscopic equations of motion. The rules and basic idea of the FHP model are
illustrated in Fig. 5. Later, the model has been extended to three dimensions as
well. From the update rules it is clear that latticegas cellular automata (LGCA)
make an efcient simulation scheme, in particular on a parallel computer, since
the rules are completely local. Moreover, stability of the algorithm is guaran
teed, since the update involves only bit manipulation, i.e., the update is exact
with no roundoff errors. We will continue with a more formal description of
general LGCA models.
2. Denitions and Equation of Motion
In LGCA models, time and space are discrete; this means that the model
system is dened on a lattice and the state of the automaton is only dened at
regular points in time with separation dt. The distance between nearestneighbor
sites in the lattice is denoted by dl. At discrete times, particles with mass m are
situated at the lattice sites with b possible velocities c
i
, where i A {1, 2, y, b}.
The set c
i
can be chosen in many different ways, although they are restricted by
the constraint that
r
0
r c
i
dt (1)
FIG. 5. Example of the time evolution in a small section of the FHP model. In the gure, each
arrow represents the velocity of a single particle. The particles are situated at the lattice sites. The
update of the lattice consists of two steps. First there are local collisions at all sites, simultaneously,
and such that locally the number of particles and momentum is conserved. Note that only some
cases lead to a new conguration. The next step is a propagation step: all the particles move
simultaneously according to their velocities to neighboring sites. Particles do not interact during this
step. Note that the gure only represents a small section of the lattice; therefore, we can only give the
complete nal conguration of the central site, as the state of the other sites after the propagation
depends on neighboring sites that are not shown. Next there will again be a collision step, etc.
MULTISCALE MODELING OF GASFLUIDIZED BEDS 75
where r and r
0
must be lattice sites. Apart from Eq. (1), it has been proved
essential to have additional symmetry requirements on the velocity set in order
to get isotropic, macroscopic behavior from these models. These requirements
turn out to be that the evenrank tensors that can be constructed from the
velocity set are isotropic up to 4th rank, and the oddrank tensors are zero. The
time evolution of the LGCA consists of two steps:
1. Propagation: All particles move in one time step dt from their initial lattice
position r to a new lattice position r
0
r+c
i
dt.
2. Collision: The particles at all lattice sites undergo a collision that conserves
the total number of particles and the total momentum at each site. The
collision rules may or may not be deterministic.
The state of the automaton at time t can be completely determined by the
boolean variable n
i
(r,t), which is equal to 1(0) if a particle is present (absent)
on site r with velocity c
i
. From this it follows that the local microscopic density ~ r
and ow velocity ~ u at site r are given by
~ pr; t m
i
n
i
r; t; ~ pr; t ~ ur; t m
i
n
i
r; tc
i
(2)
with m as the mass of the particles. The update of n
i
(r,t) (from propagation and
collision) can formally be written by the following equation of motion:
n
i
r c
i
dt; t dt n
i
r; t D
i
nr; t (3)
with
D
i
nr; t
s;s
0
s
0
i
s
i
xs; s
0
P
j
n
sj
j
1 n
j
1sj
With the sum is over all possible states s and s
0
of a single site, and x(s, s
0
) is a
collision function that is equal to one for states s, s
0
where s goes over into s
0
in a
collision, and zero for all other possible pairs of states. Note that expression
in Eq. (3) is the formal expression for the update. In a numerical code, the state
n
i
(r,t) of the systems is represented by a bbit word for every site r at time t. The
collision process can then be done by a very quick table lookup; whereas, for
the propagation step the bits from one word have to be put at the same bit
positions in the words describing the states of the neighboring sites. Despite its
extremely simplied microdynamical behavior, it turns out that the analogy of
these models with the real uid models is very close, such that the theoretical
framework of statistical mechanics of simple uids can be applied, to a great
extent, to these discrete uids. That is, starting from the formal expression in Eq.
(3), it can be proven that the macroscopic equations of motions are, in a well
dened limit, equivalent to the NavierStokes equations (Frisch et al., 1987; Ernst
and Dufty, 1989). This solid theoretical basis makes the LGCA method not just a
toy model of computer scientists but also a numerical scheme that can be seriously
M.A. VAN DER HOEF ET AL. 76
considered for the study of hydrodynamic ow phenomena. However, such
an application is seriously hindered by two big drawbacks of LGCA. Firstly,
their inherent noisiness, which means that massive averaging is required to get
accurate numbers. And secondly, it turns out that the method is not suitable for
modeling uid ow at Reynolds numbers above Re 100, which is related to the
fact that the viscosity cannot be made lower than a certain value, since it is
dictated by the collision function x. It is for these reasons that the current class of
LGCA methods cannot compete with CFD methods for modeling largescale
uid ow.
3. Averaged Equation of Motion
The two drawbacks mentioned above can be overcome, however, by consid
ering the ensembleaveraged version of the microscopic equation of motion,
Eq. (3):
f
i
r c
i
dt; t dt f
i
r; t C
i
f r; t (4)
with f
i
(r,t) /n
i
(r,t)Sthe average occupation number of link i at site r and
time t, which is now a oating number between zero and one, and
C
i
f hD
i
ni
s;s
0
s
0
i
s
i
xs; s
0
j
f
s
j
j
1 f
j
1s
j
(5)
where in the second step the assumption is made that the particle occupation
numbers on a single site are not correlated, so that the average of the product
can be written as the product of the average. The ensemble averaged density r
and ow velocity u follow from Eq. (2):
r h ~ ri m
i
f
i
; ru h ~ r~ ui m
i
f
i
c
i
(6)
where we have omitted the space and time dependence. In its present form, the
collision operator in Eq. (5) is not very useful for simulations, since the update
of f
i
at each site requires the double sum over all possible states, where there are
over 16 million states (2
24
) for the 3D models. This problem can be circum
vented by expanding the collision function about the equilibrium distribution
function f
i
eq
, for which it holds that C(f
i
eq
) 0:
f
i
r c
i
dt; t dt f
i
r; t
b
j1
L
ij
f
j
r; t f
eq
j
r; t (7)
where L is the linearized collision operator:
L
ij
@C
i
@f
j
_ _
f
j
f
eq
j
MULTISCALE MODELING OF GASFLUIDIZED BEDS 77
which can be evaluated directly from Eq. (5). Note that L, which has to be
calculated only once for a given set of collision rules, is now a small b b matrix,
compared to x that is a 2
b
2
b
matrix. Equation (7) can be directly converted
into an algorithm for simulation purpose. The advantage of an LB simulation is
that the system is essentially free of noise. Also, the linearized collision operator
need not necessarily be evaluated from an existing set of microscopic collision
rules x(s, s
0
). One is free to dene any operator L, which has the correct sym
metry and conserves momentum and number of particles. As an example, for
the 2D hexagonal lattice, one can derive from the requirements of symmetry and
conservation that L should have the following form:
L
a b c d c b
b a b c d c
c b a b c d
d c b a b c
c d c b a b
b c d c b a
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
(8)
with
a
4
21
1
3
l; b
1
7
1
6
l; c
4
21
1
6
l; d
1
7
1
3
l
where l can take any value between 0 and 2 and is related to the kinematic
viscosity via
m
1
4l
1
8
_ _
r
dl
2
dt
This gives the possibility to make the viscosity arbitrarily small, so that sim
ulations can be performed also at Reynolds numbers higher than 100.
B. THE LATTICE BHATNAGARGROSSKROOK MODEL
In the linearized LB equation Eq. (7), the ensemble averaged effect of the
particleparticle collision is now represented by a relaxation of the distribution
function f
i
to the equilibrium function f
i
eq
, where the matrix L
ij
does not nec
essarily have to correspond to an existing set of collision rules. The question
now arises if L can be simplied even further to the form L
ij
ad
ij
, so that the
LB equation takes the form (with t dt/a)
f
i
r c
i
dt; t dt f
i
r; t
dt
t
f
i
r; t f
eq
i
r; t (9)
M.A. VAN DER HOEF ET AL. 78
At rst sight this does not seem possible because the specic form of L, for
instance Eq. (8), which followed from the requirement of symmetry. However, it
turns out that b, c, and d can be put to zero, if an equilibrium function f
i
eq
with
different weights is used for the different directions, so that the lack of symmetry
can be remedied (Qian et al., 1992; Ladd, 1994; Succi, 2001). Specically, the f
i
eq
should take the form
f
eq
i
a
ci
r 1
c
i
u
c
2
s
c
i
u
2
2c
4
s
u
2
2c
2
s
_ _
(10)
where the weight a
c
i
only depends on the magnitude c
i
of the velocity c
i
con
nected to the link direction i, and c
s
is the speed of sound. For the popular
D3Q19 model (3D, 19 velocities), there are 6 velocities with c
i
1dl/dt, 12
velocities with c
i
2
p
dl=dt; and one zero velocity, c
i
0 (see Fig. 6). The
parameters that yield the proper equilibrium distribution in Eq. (10) are a
0
1/3, a
1
1/36, and a
2
p
1=18, and the speed of sound is usually set to c
s
1/3(dl/dt) (Ladd and Verberg, 2001; Succi, 2001).
To rst order, expression in Eq. (9) represents the nite difference form of the
wellknown BGK equation:
@
@t
r c
i
_ _
f
i
1
t
f
i
f
eq
i
_
(11)
FIG. 6. The D3Q19 model.
MULTISCALE MODELING OF GASFLUIDIZED BEDS 79
It can be shown that this equation will yield the familiar conservation of mass
and momentum equations
@
@t
r r ru 0;
@
@t
ru r ruu r p (12)
where the pressure tensor is equal to
p p
I m ru ru
T
_
l
2
3
m
_ _
r u
s
I
_
with
I the unit tensor, (ru)
ab
r
a
u
b
, ru
T
ab
r
b
u
a
, and the pressure p, shear
viscosity m, and bulk viscosity l given by
p c
2
s
r; m tc
2
s
r; l 0 (13)
As said, the lattice BGK in Eq. (9) is the nite difference version of Eq. (11) to
rst order in dt. To second order in dt, however, Eq. (9) represents the nite
difference version of a slightly different expression:
@
@t
r c
i
_ _
f
i
1
t
f
i
f
eq
i
_
dt
2t
@
@t
r c
i
_ _
f
i
f
eq
i
_
(14)
In the route to NavierStokes, it turns out that f
i
eq
in the second term on the
righthand side (RHS) does not play a role, so that we can rewrite Eq. (14) as a
normal BKG equation with a different prefactor on the RHS:
@
@t
r c
i
_ _
f
i
dt
t dt=2
f
i
f
eq
i
_
We thus nd that the lattice BGK model describes, to second order in dt, the
uid according to the NavierStokes equation with a viscosity
m t
1
2
dt
_ _
c
2
s
r
where the extra term (1/2)dt is due to the nite difference scheme. As a dem
onstration of how well the simple lattice BGK scheme in Eq. (9) can describe
uid ow, we show in Fig. 7 the velocity prole from a lattice BGK simulation
for forced Poiseuille ow and shear ow; it can be seen that excellent agreement
is found with the theoretical results that follow from the NavierStokes equa
tion. Note that for the simulations shown in Fig. 7 we used stickboundary
M.A. VAN DER HOEF ET AL. 80
conditions at the walls of the channel. The implementation of such conditions is
similar to the boundary conditions that are required to model large, solid ob
jects in the LB model, which is described in Section II.C.
C. MODELING SOLID PARTICLES
In order to simulate large, moving particles in the LB model, we should dene
additional rules that describe the interaction of the LB gas with the surface of
the solid particles. One essential ingredient of the movingboundary rules is that
these rules result, on average, in a dissipative force on the suspended particle.
An obvious choice of rules is those according to which the gas next to the solid
particle moves with the local velocity of the surface of the solid particle. In this
way, one models the hydrodynamic stick boundary condition; for a spherical
particle suspended in an innite 3D system, moving with velocity v, this will give
rise to a frictional force on the particle F 3pmdv, at least in the limit of low
Reynolds numbers Re rdv/m, where d is the hydrodynamic diameter of the
particle and m is the shear viscosity. A particular efcient and simple way to
enforce stickboundary rules was introduced by Ladd (1994). First the bound
ary nodes are identied, which are dened as the points halfway two lattice sites
that are inside and outside the particle and closest to the actual surface (see
Fig. 8, left graph, solid squares). For a static particle, the boundary rule is
simply that a distribution moving such that it would cross the boundary is
bounced back at the boundary node. Since this node is halfway the link, the
bounceback rule has the effect so (see Fig. 8):
f
i
r; t dt f
i
0 r; t (15)
Forced Poiseuille flow
Lattice Boltzmann
Theory
Fluid velocity
Shear flow
Lattice Boltzmann
Theory
Fluid velocity
v
0
FIG. 7. Velocity prole for Poiseuille ow and shear ow. The points are the LBM data, and the
solid lines are the theoretical proles. For the simulations, we used the D3Q18 model with 25 lattice
sites across the channel.
MULTISCALE MODELING OF GASFLUIDIZED BEDS 81
where i and i
0
are opposite links. This rule ensures that the uid velocity at the
boundary node indeed vanishes; the momentum at the boundary node at time
t+(1/2)dt is given by
r
b
u
b
f
i
0 r; t c
i
0
f
i
r; t dt c
i
(16)
Inserting Eq. (15) and using c
i
c
i
0
gives r
b
u
b
0. For nonstatic particles, the
local uid velocity must be set equal to the local boundary velocity v
b
. This can
be achieved by a simple modication of the bounceback rule:
f
i
r; t dt f
i
0 r; t a c
i
(17)
where a is chosen such that u
b
v
b
. Note that only the component of v
b
in the
direction of the link can be set in this way. For more details we refer to the
papers by Ladd (1994) and Ladd and Verberg (2001). In Fig. 9 (right graph), we
show the LBM simulation results for the velocity of a single freefalling sphere
in an (effectively) unbounded uid. As can be seen from Fig. 9, the boundary
rule results in a terminal velocity according to the StokesEinstein friction force.
Note that the actual plateau value of the velocity is slightly smaller than the
theoretical prediction. This can be attributed to the fact that the radius of the
particle is not well dened because of the irregular shape of boundarynode
surface of the sphere. In fact, the freefalling sphere experiment (or a similar
experiment with periodic boundary conditions) is used for calibration purposes,
i
i
i
i
site
boundary
r
site
s
node
FIG. 8. Left graph: example of a boundary node map for a disc in a 2D hexagonal lattice. Right
graph: illustration of the bounceback rule on an enlarged section of the boundary. The distribution
at site r that moves at time t into direction i
0
, instead of arriving at the (virtual) site s, is bounced at
the boundary node, and thus arrives back at site r at time t+dt, but is now headed in the opposite
direction i.
M.A. VAN DER HOEF ET AL. 82
i.e., the effective hydrodynamic radius of the spheres is obtained from the ter
minal velocity, where it is assumed that the StokesEinstein relation holds.
D. RESULTS FOR THE GAS SOLID DRAG FORCE
The drag force from the gas phase on an assembly of spheres can, in principle,
be obtained from the terminal velocity in a sedimentation experiment. However,
the drag force can also be directly measured in the simulation from the change
in gas momentum due to the boundary rules. The change in gas momentum per
unit time, required to maintain stickboundary conditions for particle a, is equal
to minus the total force F
gs,a
that the gas phase exerts on particle a. This total
force has two contributions (see also Section III.D), namely the drag force F
d,a
due the uidsolid friction at the surface of the spheres and a force
F
p,a
V
a
rp due to the static pressure gradient rp, which drives the gas ow
past the spheres (V
a
pd
3
/6 is the volume of the sphere). From a balance of
forces it follows that Vrp
a
F
gs,a
, with V the total volume of the system;
eliminating rp from the expressions gives F
d,a
eF
gs,a
with e the volume
fraction of the gas phase. The procedure to obtain the drag force for mono
disperse systems in an LB simulation is then as follows. N particles with di
ameter d are distributed randomly in a box of n
x
n
y
n
z
lattice sites, so that
the gas fraction equals e 1Npd
3
/(6n
x
n
y
n
z
dl
3
). Some typical values are
N 54 and d 25dl, where periodic boundary conditions are used. All spheres
are forced to move with the same constant velocity v
sim
in some arbitrary di
rection, so that the array of spheres moves as a static conguration through the
system. A uniform force is applied to the gas phase, to balance the total force
0 1000 2000 3000 4000 5000
time
0.000
0.002
0.004
0.006
P
a
r
t
i
c
l
e
v
e
l
o
c
i
t
y
Sedimentation of a single sphere
Simulation
v
term
= mg/3d
FIG. 9. Velocity of a single sphere in a 3D LB gas. The black line is the data from LBM, which has
the proper functional form v(t) v
N
(1 exp(gt/v
N
)). The grey line is theoretical terminal velocity,
which is slightly higher than v
N
.
MULTISCALE MODELING OF GASFLUIDIZED BEDS 83
N
a1
F
g!s;a
from the moving particles on the gas phase. From this it follows
that in a frame of reference where the particles are static, the supercial ow
velocity is equal to v
sim
, so that Re pdv
sim
/m, where m is the viscosity. Once
an equilibrium state is obtained, the average value
F
g!s
h
N
a1
F
g!s;a
=Ni
t
is
determined, with / S
t
a time average. The momentum exchange coefcient b,
as dened in Eq. (44), is then determined via
b
F
g!s
1
2
v
sim
V
a
since the relative velocity uv
a
in Eq. (44) corresponds to v
sim
/e in the LB
simulations and F
d,a
eF
gs,a
. Note that the dimensionless quantity bd
2
/m will
only depend on the Reynolds number and the gas fraction e. Ergun (1952)
showed that when the experimental data for b for different e and Re is plotted in
a single {log(x), log(y)} graph with
x
Re
1
y
bd
2
m1
1
Re
all data fall onto a single curve y 150/x+1.75, which corresponds to
bd
2
m1
1501 1:75Re (18)
which is the famous Ergun equation. In Fig. 10, we show the data from ex
tensive LB simulations (Van der Hoef et al., 2004; Beetstra et al., 2006) for a
range of gas fractions and Reynolds numbers, on the same {log(x), log(y)}
graph. We nd that our LBM data deviates substantially from the Ergun
equation: for low Re numbers the Ergun equation underestimates the drag
force, whereas for high Re numbers the Ergun equation overestimates the drag
force. A simple remedy would be to use different coefcients in the Ergun
equation, for instance 180 instead of 150 and 1.0 instead of 1.75. However, it
can be seen from Fig. 10 that not all data obey the functional form y A/x+B.
Note also that the Ergun equation was derived for packed beds, and is not
expected to be valid for high gas fractions; for that range, normally the Wen and
Yu Eq. (46) is used. However, we nd that this equation signicantly under
predicts the drag force at higher Reynolds numbers (see Fig. 10). Based on our
data from the LB simulations, we suggest the following new dragforce cor
relation that we write in the form of an Erguntype equation (Beetstra et al.,
2006; Van der Hoef et al., 2005):
bd
2
m 1
A1 BRe (19)
M.A. VAN DER HOEF ET AL. 84
only with coefcients that depend on both e and Re:
A 180
18
4
1
1 1:5
1
p
_ _
and
B
0:31
1
31 8:4Re
0:343
1 10
3 1
Re
22:5
Expression in Eq. (19) is within 8% of all simulation data up to Re 1000.
Since this relation has been derived very recently (Beetstra et al., 2006), it has
not been applied yet in the higher scale models discussed in Sections III and IV.
However, the expression by Hill et al. in Eq. (47) derived from similar type of
LBM simulations is consistent with our data, in particular when compared to
the large deviations with the Ergun and Wen and Yu equations. So, we expect
that the simulation results presented in Section IV.F using the Hill et al.
correlation will not be very different from the results that would be obtained
with expression in Eq. (19). A more detailed account of the derivation of
expression in Eq. (19) and a comparison with other dragforce relations can be
found in Ref. Beetstra et al. (2006).
10
1
10
1
10
3
10
5
Re / (1 )
10
0
10
1
10
2
10
3
(
d
2
/
/
(
(
1
)
R
e
)
Ergun
WenYu (=0.8)
Hill et al (=0.5)
Equation (19) (=0.5)
LBM (=0.40.8)
FIG. 10. Normalized drag force at arbitrary Reynolds numbers and gas fractions. The symbols
represent the simulation data, the solid line the Ergun correlation Eq. (18), the dashed line the
WenYu correlation Eq. (46) for e 0.8, and the grey line the correlation by Hill et al. (2001a,b) Eq.
(47) and the longdashed line Eq. (19), both for e 0.5.
MULTISCALE MODELING OF GASFLUIDIZED BEDS 85
III. Discrete Particle Model
A. INTRODUCTION
DPMs offer a viable tool to study the macroscopic behavior of assemblies of
particles and originate from MD methods. Initiated in the 1950s by Alder and
Wainwright (1957), MD is by now a welldeveloped method with thousands of
papers published in the open literature on just the technical and numerical
aspects. A thorough discussion of MD techniques can be found in the book by
Allen and Tildesley (1990), where the details of both numerical algorithms and
computational tricks are presented. Also, Frenkel and Smit (1996) provide a
comprehensive introduction to the recipes of classical MD with emphasis on
the physics underlying these methods. Nearly all techniques developed for MD
can be directly applied to discrete particles models, except the formulation of
particleparticle interactions. Based on the mechanism of particleparticle in
teraction, a granular system may be modeled either as hardspheres or as
softspheres.
1. HardSphere Model
In a hardsphere system, the trajectories of particles are determined by mo
mentum conserving binary collisions. The interactions between particles are
assumed to be pairwise additive and instantaneous. In the simulation, the col
lisions are processed one by one according to the order in which the events
occur. For not too dense systems, the hardsphere models are considerably
faster than the softsphere models. Note that the occurrence of multiple col
lisions at the same instant cannot be taken into account.
Campbell and Brennen (1985) reported the rst hardsphere discrete particle
simulation used to study granular systems. Since then, the hardsphere models
have been applied to study a wide range of complex granular systems. Hoo
mans et al. (1996) used the hardsphere model, in combination with a CFD
approach for the gasphase conservation equations, to study gassolid two
phase ows in gasuidized beds. By using this model, they studied the effect of
particleparticle interaction on bubble formation (Hoomans et al., 1996) and
the segregation induced by particlesize differences and density differences
(Hoomans et al., 2000). This model has been further used in connection with
the kinetic theory of granular dynamics by Goldschmidt et al. (2001), high
pressure uidization by Li and Kuipers (2002), and circulating uidized beds by
Hoomans (2000).
Similar simulations have been carried out by other research groups. Ouyang
and Li (1999) developed a slightly different version of this model. Helland et al.
(1999) recently developed a DPM in which hardsphere collisions are assumed,
but where a timedriven scheme (typically found in the softsphere model) is
used to locate the collisional particle pair. Effect of the gas turbulence has also
M.A. VAN DER HOEF ET AL. 86
been taken into account in some hardsphere models by Helland et al. (2000),
Lun (2000), and Zhou et al. (2004).
At highparticle number densities or low coefcients of normal restitution e,
the collisions will lead to a dramatical decrease in kinetic energy. This is the so
called inelastic collapse (McNamara and Young, 1992), in which regime the
collision frequencies diverge as relative velocities vanish. Clearly in that case, the
hardsphere method becomes useless.
2. SoftSphere Model
In more complex situations, the particles may interact via short or long
range forces, and the trajectories are determined by integrating Newtonian
equations of motion. The softsphere method originally developed by Cundall
and Strack (1979) was the rst granular dynamics simulation technique
published in the open literature. In softsphere models, the particles are
allowed to overlap slightly and the contact forces are subsequently calculated
from the deformation history of the contact using a contactforce scheme. The
softsphere models allow for multiple particle overlap, although the net contact
force is obtained from the addition of all pairwise interactions. The softsphere
models are essentially timedriven, where the time step should be carefully cho
sen in calculating the contact force. The softsphere models that can be found in
literature mainly differ from each other with respect to the contactforce scheme
that is used. A review of various popular schemes for repulsive interparticle
forces is presented by Scha fer et al. (1996). Walton and Braun (1986) developed
a model that uses two different spring constants to model the energy dissipation
in the normal and tangential directions. In the force scheme proposed by
Langston et al. (1994), a continuous potential of an exponential form is used,
which contains two unknown parameters: the stiffness of the interaction and an
interaction constant.
A 2D softsphere approach was rst applied to gasuidized beds by Tsuji
et al. (1993), where the linear springdashpot modelsimilar to the one
presented by Cundall and Strack (1979)was employed. Xu and Yu (1997) in
dependently developed a 2D model of a gasuidized bed. However in their sim
ulations, a collision detection algorithm that is normally found in hardsphere
simulations was used to determine the rst instant of contact precisely. Based on
the model developed by Tsuji et al. (1993), Iwadate and Horio (1998) incorporated
van der Waals forces to simulate uidization of cohesive particles. Kafui et al.
(2002) developed a DPM based on the theory of contact mechanics, thereby
enabling the collision of the particles to be directly specied in terms of material
properties such as friction, elasticity, elastoplasticity, and autoadhesion.
It is also interesting to note that softsphere models have also been applied to
other applications such as gasparticle heat transfer by Li and Mason (2000)
and coal combustion by Zhou et al. (2003). Clearly, these methods open a new
way to study difcult problems in uidized bed reactors.
MULTISCALE MODELING OF GASFLUIDIZED BEDS 87
3. Comparison between Hard and SoftSphere Models
Although both hard and softsphere models have been used in the simulation
of granular ow, each has its own characteristics that make them very efcient
in some cases, while inefcient in others. The two types of models are compared
in Table I. Hardsphere models use an eventdriven scheme because the inter
action times are (assumed to be) small compared to the freeight time of
particles, where the progression in physical time depends on the number of
collisions that occur. In contrast, in the softsphere models a time step that is
signicantly shorter than the contact time should be used. This directly implies
that the computational efciency of the softsphere model (compared to the
hardsphere model) decreases when the ratio of the freeight time to the con
tact time increases, which is the case when the system becomes less dense. In the
softsphere models, a slight deformation of particles is allowed, so that multiple
contacts between several pairs of particles are possible, which should never
happen in the eventdriven models. As mentioned above, a lower coefcient of
normal restitution may lead to the inelastic collapse in hardsphere simulations.
Incorporation of cohesive forces, especially the pairwise forces, is quite
straightforward in softsphere models. This is because the collisional process in
the softsphere model is described via the Newtonian equations of motion of
individual particles, that is, in terms of forces. In the hardsphere system, the
update is not via forces (since they are, loosely speaking, either zero or innite),
but via a momentum exchange at contact. This means that for shortrange
forces, such as the cohesive force, a kind of hybrid method for the interaction at
close encounters has to be devised, which is not straightforward. In contrast, for
systems with different size particles, it is the softsphere model that poses some
difculties. In a softsphere system using a linear springdashpot scheme, the
spring stiffness is dependent on the particle size. This means that in principle a
different spring stiffness should be used for calculating the contact forces be
tween particles with different sizes, otherwise the computing efciency will drop
substantially.
TABLE I
COMPARISON BETWEEN HARD AND SOFTSPHERE MODELS. THE SYMBOLS INDICATE GOOD (++),
NORMAL (+), AND NOT SUITABLE ()
Hardsphere Softsphere
Computing efciency ++ +
Multiple contacts ++
Dense systems ++
Incorporation of cohesive force + ++
Energy conservation during collisions ++ +
Multiple particle sizes ++ +
M.A. VAN DER HOEF ET AL. 88
In the following, we focus on the softsphere method since this really is the
workhorse of the DPMs. The reason is that it can in principle handle any
situation (dense regimes, multiple contacts), and also additional interaction
forcessuch as van der Waals or electrostatic forcesare easily incorporated.
The main drawback is that it can be less efcient than the hardsphere model.
B. PARTICLE DYNAMICS: THE SOFTSPHERE MODEL
1. The Equations of Motion
The linear motion of a single spherical particle a with mass m
a
and coordinate
r
a
can be described by Newtons equation:
m
a
d
2
r
a
dt
2
F
contact;a
F
pp;a
F
ext;a
(20)
where the RHS is the total force on the particle, which has three basic con
tributions:
(i) The total contact force F
contact,a
is the sum of the individual contact forces
exerted by all other particles in contact with the particle a, which are nat
urally divided into a normal and a tangential component:
F
contact;a
b2contactlist
F
ab;n
F
ab;t
b2contactlist
R
a
n
ab
F
ab;t
with R
a
as the radius of particle a.
We will next give a more detailed description of the contact force, the co
hesive force, and the integration of the equations of motionEqs. (20) and (21).
The description of the forces resulting from interaction with the gas phase is
given in Section III.D, whereas the dynamics of the gas phase itself is described
in Section III.C.
2. Contact Force
The calculation of the contact force between two particles is actually quite
involved. A detailed model for accurately computing contact forces involves
complicated contact mechanics (Johnson, 1985), the implementation of which is
extremely cumbersome. Many simplied models have therefore been proposed,
which use an approximate formulation of the interparticle contact force. The
simplest one was originally proposed by Cundall and Strack (1979), where a
linearspring and dashpot model is employed to calculate the contact forces (see
Fig. 11 and 12). In this model, the normal component of the contact force
between two particles a and b can be calculated by
F
ab;n
k
n
d
n
n
ab
Z
n
v
ab;n
(22)
where k
n
is the normal spring stiffness, n
ab
the normal unit vector, Z
n
the normal
damping coefcient, and v
ab,n
the normal relative velocity. The overlap d
n
is
given by
d
n
R
b
R
a
r
b
r
a
j j
FIG. 11. Graphical representation of the linear springdashpot softsphere model. From Hoo
mans, Ph.D. thesis, University of Twente (2000).
M.A. VAN DER HOEF ET AL. 90
with R
a
and R
b
denoting the radii of the particles. The normal unit vector is
dened as
n
ab
r
b
r
a
r
b
r
a
j j
(23)
The relative velocity of particles a and b is
v
ab
v
a
v
b
R
a
o
a
R
b
o
b
n
ab
(24)
where v
a
and v
b
are the particle velocities, and o
a
and o
b
the angular velocities.
The normal component of the relative velocity between particle a and b is
v
ab;n
v
ab
n
ab
n
ab
(25)
For the tangential component of the contact force, a Coulombtype friction law
is used:
F
ab;t
k
t
d
t
Z
t
v
ab;t
for F
ab;t
m
f
F
ab;n
m
f
F
ab;n
t
ab
for F
ab;t
4m
f
F
ab;n
_
(26)
where k
t
, d
t
, Z
t
, and m
f
are the tangential spring stiffness, tangential displace
ment, tangential damping coefcient, and friction coefcient, respectively. In
FIG. 12. The coordinate system used in the softsphere model.
MULTISCALE MODELING OF GASFLUIDIZED BEDS 91
Eq. (26), the tangential relative velocity v
ab,t
, and tangential unit vector t
ab
are
dened as
v
ab;t
v
ab
v
ab;n
t
ab
v
ab;t
v
ab;t
; c cos j; s sin j; q 1 c
and
j arcsin n
ab
n
0;ab
_ _
Are Eqs. (27) and (28) sufcient to describe the tangential displacement dur
ing particleparticle contact? In the absence of friction, the answer is yes. When
we consider friction during particleparticle contactas pointed out by Brendel
FIG. 14. The transformation of tangential displacement vector.
MULTISCALE MODELING OF GASFLUIDIZED BEDS 93
and Dippel (1998)the use of Eqs. (27) and (28) may give rise to unphysical
behavior for dense systems due to the allowance of an arbitrarily large tangen
tial displacement (Eq. 28). In a dilute system, this will not be a problem since the
multipleparticle contacts do not happen frequently. In this case if the contacts
ends, the tangential displacements will be set to zero. In contrast for dense
systems, multipleparticle contacts are very common and the contact history for
a specic particle could be very long. The long contact history causes a relatively
large tangential displacement, which means that an extra friction force should
be taken into account. This problem can be overcome, however, by using the
method proposed by Brendel and Dippel (1998), where the tangential displace
ment during the friction is calculated by d
t
m
f
F
ab,n
t
ab
/k
t
, so that
d
t
d
t
0
H
_
t
t
0
v
ab;t
dt for F
ab;t
m
f
F
ab;n
m
f
F
ab;n
t
ab
=k
t
for F
ab;t
4m
f
F
ab;n
_
_
_
(29)
4. Collision Parameters
To solve the Eqs. (20) and (21), we have to specify ve parameters: normal
and tangential spring stiffness k
n
and k
t
, normal and tangential damping
coefcient Z
n
and Z
t
, and the friction coefcient m
f
. In order to get a better
insight into how these parameters are related, it is useful to consider the equa
tion of motion for the overlap in the normal direction dn:
m
eff
d
::
n
k
n
d
n
Z
n
d
:
n
(30)
which follows from Eq. (20) when only the normal contact force is taken into
account. In Eq. (30), m
eff
is the reduced mass of the two interacting particles a
and b:
1
m
eff
1
m
a
1
m
b
Equation (30) is the wellknown differential equation of the damped harmonic
oscillator, the solution of which is
d
n
t u
0
=O expCt sinOt (31)
d
:
n
t u
0
=O expCtCsinOt OcosOt (32)
with u
0
_
d
n
0 as the initial relative velocity, and
O
O
2
0
C
2
_
O
0
k
n
=m
eff
_
C Z
n
=2m
eff
d
:
n
0
exppC=O (33)
Thus, we can calculate the normal damping coefcient via
Z
n
2 ln e
m
eff
k
n
p
p
2
ln
2
e
_ ea0
Note that for e 0 we get O 0 according to Eq. (33), so that in that case
Z
n
2
k
n
m
eff
p
.
We can follow a similar procedure for the tangential springdashpot system.
So, the tangential damping coefcient is determined by
Z
t
2 ln e
t
m
0
eff
k
t
_
p
2
ln
2
e
t
_ e
t
a0
where m
0
eff
2 m
eff
/7 is the reduced mass of the twoparticle system interaction
via a tangential linear spring. Note that m
0
eff
is different from m
eff
, since in a
tangential direction both the rotational and translational momentum must be
considered. In the case of particlewall contact, we shall simply treat particle b
as a big particle with an innite radius, so that we have
m
eff
m
a
m
0
eff
2
7
m
a
The contact force between two particles is now determined by only ve pa
rameters: normal and tangential spring stiffness k
n
and k
t
, the coefcient of
normal and tangential restitution e and e
t
, and the friction coefcient m
f
. In
principle, k
n
and k
t
are related to the Young modulus and Poisson ratio of the
solid material; however, in practice their value must be chosen much smaller,
otherwise the time step of the integration needs to become impractically small.
The values for k
n
and k
t
are thus mainly determined by computational efciency
and not by the material properties. More on this point is given in the Section
III.B.7 on efciency issues. So, nally we are left with three collision parameters
e, e
t
, and m
f
, which are typical for the type of particle to be modeled.
MULTISCALE MODELING OF GASFLUIDIZED BEDS 95
5. Cohesive Force
Cohesion between particles can arise from a variety of sources including van
der Waals forces, liquid bridging (i.e., capillary forces), sintering, and so on. Of
these forces, which become increasingly important as the particle size decreases,
the van der Waals force is generally accepted as the dominating cohesive force in
gasuidized beds of ne particles (Geldart A and C particles), and will be
considered next. The van der Waals force is present between any two molecules
(polar or nonpolar) and follows from the interaction of the uctuating dipole
moments on the molecules. According to the London theory, the potential
energy of two molecules i and j at distance r
ij
, due to the van der Waals in
teraction, is equal to jr C
6
r
6
ij
. The total energy between two macroscopic
bodies a and b, made of the same material, then equals:
V C
6
i on a f g
j on b f g
r
6
ij
C
6
r
2
_
v
a
_
v
b
dr
a
dr
b
r
a
r
b
j j
6
(34)
where in the last step we replaced S
i on a
by
_
V
a
dr
a
r(and similarly for b), where
r is the density of the material, which is justied since the number of molecules
present in the particles is very large. For two spheres with radii R
a
and R
b
,
where the centers are at position r
a
and r
b
, respectively, expression Eq. (34) can
be evaluated analytically (Hunter, 1986; Israelachvili, 1991):
V r
ab
A
6
2
r
2
ab
4
2
r
2
ab
ln
r
2
ab
4
r
2
ab
_ _ _ _
with
r
2
ab
r
2
ab
R
a
R
b
2
R
a
R
b
; r
ab
r
b
r
a
j j; A p
2
r
2
C
6
(35)
The parameter A is known as the Hamaker constant. The force on sphere a then
follows via
F
coh;a
@Vr
ab
@r
ab
n
ab
32 A
3 R
a
R
b
r
ab
n
ab
r
4
ab
r
2
ab
4
2
(36)
with n
ab
dened by Eq. (23). When the spheres are nearly touching (r
ab

R
a
+R
b
), and for R
a
R
b
R, the force in Eq. (36) can be simplied to
F
coh;a
AR
12s
2
ab
n
ab
s
ab
r
ab
2R
Note that Eq. (36) exhibits an apparent numerical singularity in that the van der
Waals interaction diverges if the surface distance between two particles
approaches zero. In reality, such a situation will never occur because of the
M.A. VAN DER HOEF ET AL. 96
shortrange repulsion between particles. In the present model, we have not
included such a repulsion; however, we can avoid the numerical singularity by
dening a cutoff (maximal) value of the van der Waals force between two
spheres. In practice, it is more convenient to use the equivalent cutoff value for
the intersurface distance, s
ab
0
, instead of for the interparticle force.
The Hamaker constant A can, in principle, be determined from the C
6
co
efcient characterizing the strength of the van der Waals interaction between
two molecules in vacuum. In practice, however, the value for A is also inu
enced by the dielectric properties of the interstitial medium, as well as the
roughness of the surface of the spheres. Reliable estimates from theory are
therefore difcult to make, and unfortunately it also proves difcult to directly
determine A from experiment. So, establishing a value for A remains the main
difculty in the numerical studies of the effect of cohesive forces, where the
value for glass particles is assumed to be somewhere in the range of 10
21
joule.
6. Integrating the Equations of Motion
In the following section, we only consider the integration of the equation of
linear motion Eq. (20); the procedure for the equation of rotational motion, Eq.
(21), will be completely analogous. Mathematically, Eq. (20) represents an in
itialvalue ordinary differential equation. The evolution of particle positions
and velocities can be traced by using any kind of method for ordinary differ
ential equations. The simplest method is the rstorder integrating scheme,
which calculates the values at a time t+dt from the initial values at time t (which
are indicated by the superscript 0) via:
v
a
v
0
a
a
0
a
dt; r
a
r
0
a
v
a
dt (37)
where a
a
is the acceleration:
a
a
F
contact;a
F
pp;a
F
ext;a
m
a
(38)
The rstorder integration scheme, however, will introduce a drift in the energy;
from Eq. (37), we have
v
a
a
0
a
dt
2
v
a
2
a
0
a
dt
2
2v
a
a
0
a
dt v
0
a
2
so
1
2
v
a
2
v
a
a
0
a
dt
1
2
a
2
a
dt
2
1
2
v
0
a
2
(39)
The rst term on the left of Eq. (39) is the reduced kinetic energy of the particle
at time t+dt, the second term is the work due to all kinds of external forces, and
the rst term on the right is the reduced kinetic energy at time t. The remaining
MULTISCALE MODELING OF GASFLUIDIZED BEDS 97
term a
2
a
dt
2
=2 is always positive, and this energy is introduced into the system
solely due to the numerical method, for each time step.
In the past decades, a large number of methods have been proposed to
achieve better energy conservation: for example, the Gear family of algorithms
and the family of Verlet algorithms (Frenkel and Smit, 1996). In our 3D code,
we have incorporated yet another type of method developed by Beeman, which
has a somewhat better energy conservation than the Verlet algorithm (Frenkel
and Smit, 1996). In the Beeman method, the position and velocity of particle a
are calculated via
r
a
r
0
a
v
0
a
dt
2
3
a
0
a
1
6
a
1
a
_ _
dt
2
v
a
v
0
a
1
3
a
a
dt
5
6
a
0
a
1
6
a
1
a
_ _
dt
where the superscript (1) denotes the values at time tdt. Note that the Bee
manVerlet algorithm is not self starting, so it requires the storage of the old
value of the acceleration a
(1)
.
7. Efciency Issues: Spring Constants and Neighbor Lists
To perform simulations of relatively large systems for relatively long times, it
is essential to optimize the computational strategy of discrete particle simula
tions. Obviously, the larger the time step dt, the more efcient the simulation
method. For the softsphere model, the maximum value for dt is dictated by the
duration of a contact. Since there are two different springdashpot systems in
our current model, it is essential to assume that t
contact,n
t
contact,t
, so that
p
2
ln e
2
k
n
=m
eff
p
2
ln e
t
2
k
t
=m
0
eff
m
eff
m
0
eff
2
7
Based on the discussion in previous sections, we can calculate the time step by
dt
1
K
N
t
contact;n
1
K
N
p
2
ln e
2
k
n
=m
eff
_
which must typically be less than 1% of the particle diameter.
A second way of speeding up the simulation is the use of neighbor lists and
cell list, which was originally developed for MD simulations (Allen and
Tildesley, 1990). The neighbor list contains a list of all particles within the
cutoff sphere of a particular particle, so that the separations do not need to
be calculated at each step, which is shown in Fig. 15. The neighbor list cutoff
s
cutoff
should be dened with care. A too small cutoff value may result in
some neighboring particles to be excluded from the list. In contrast, however,
a big cutoff value will greatly reduce the computational efciency. To speed
up the searching for neighbors, the particles in each uid cell in this research
are put into a corresponding list. All neighbors of a particle will then be
found either in the cell containing the particle or in an adjacent cell.
FIG. 15. The scheme of neighbor list and cell lists. The particle of interest is black; the grey
particles are within the neighbor list cutoff.
MULTISCALE MODELING OF GASFLUIDIZED BEDS 99
C. GAS DYNAMICS
In the DPM the gas phase is treated as a continuum phase, the dynamics of
which can be described by a set of volumeaveraged NavierStokes equations
(Kuipers et al., 1992). From mass conservation, we have
@r
@t
r ru 0 (40)
where r is the gas density, e the local porosity, and u the gas velocity. Mo
mentum conservation gives that
@ru
@t
r ruu rp S
p
r t rg (41)
where p is the gas phase pressure, t the viscous stress tensor, g the gravitational
acceleration, and S
p
a source term that describes the momentum exchange with
the solid particles present in the control volume:
S
p
1
V
_
N
part
a1
bV
a
1
u v
a
dr r
a
dV (42)
Here V represents the local volume of a computational cell and V
a
the volume of
particle a. The dfunction ensures that the drag force acts as a point force at the
(central) position of this particle. In Eq. (42), b is the momentum transfer
coefcient, which will be discussed in more detail in Section III.D. The gas
phase density r is calculated from the ideal gas law:
r
pM
RT
where R is the universal gas constant (8.314 J/(mol K)), T the temperature,
and M the molecular mass of the gas. The equation of state of the ideal gas
can be applied for most gases at ambient temperature and pressure. The
viscous stress tensor t is assumed to depend only on the gas motion. For gas
uidized beds, the general form for a Newtonian uid (Bird et al., 1960) can
be used:
t l
2
3
m
_ _
r u
I mru ru
T
(43)
with l the gas phase bulk viscosity, m the gas phase shear viscosity, and
I the
unit tensor. Normally, the bulk viscosity of the gas phase can be set equal to
zero (Bird et al., 1960). Note that no turbulence modeling is taken into
M.A. VAN DER HOEF ET AL. 100
account. For dense gassolid uidization, this can be justied since the tur
bulence is completely suppressed in the particle bed due to the high solids
volume fraction.
The numerical method for solving the set of Eqs. (40) and (41) is similar to the
method that is used in the TFM, which is discussed in detail in Section IV.E.
The time step by which the gasphase is updated is typically one order of
magnitude larger than the time step dt that is used for updating the softsphere
system. The boundary conditions are taken into account by utilizing ctitious
cells at the boundaries and a agmatrix concept, which allows different bound
ary conditions to be applied for each single cell. A variety of boundary con
ditions can be applied by specication of the value of the cell ag (i, j, k), which
denes the relevant boundary condition for the corresponding cell (i, j, k). A
typical set of boundary conditions used in a 2D simulation is shown in Fig. 16.
In Table II, we explain the meaning of each type of boundary condition.
Normally, the bottom distributor is dened as inux cells formulated by (i, j,
k) 4, where the void fraction is set to a constant value of 0.4.
FIG. 16. The typical set of boundary conditions used in 2D simulations.
MULTISCALE MODELING OF GASFLUIDIZED BEDS 101
D. INTERPHASE COUPLING
For dense gassolid twophase ows, a fourway coupling is required; how
ever, the coupling between particles is managed in a natural way in DPMs. The
task is, therefore, only to nd a twoway coupling between the gas and the solid
phases, which satises Newtons third law. Basically, the gas phase exerts two
forces on particle a: a drag force F
d,a
due the uidsolid friction at the surface of
the spheres, and a force F
p,a
V
a
rp due to the pressure gradient rp in the gas
phase. We will next describe these forces in more detail, along with the pro
cedure to calculate void fraction, which is an essential quantity in the equations
for the gassolid interaction.
1. Drag Force
The drag force that the gas phase exerts on a particle a, consistent with the
source term S
p
in expression Eq. (41), reads
F
d;a
V
a
b
1
u v
a
(44)
where b is the momentum exchange coefcient. The commonly used drag cor
relations for b in the simulation of gasuidized beds are the Ergun (1952)
equation for denser beds (eo0.8):
bd
2
m1
1501 1:75Re
a
(45)
and the Wen and Yu (1966) equation for dilute systems (e40.8):
bd
2
m1
3
4
C
d
Re
a
1:65
(46)
TABLE II
VALUES FOR THE CELL FLAG, WHICH DEFINE THE BOUNDARY CONDITIONS
(i, j, k) The type of cell
1 Interior cell, no boundary conditions have to be specied
2 Impermeable wall, freeslip boundaries
3 Impermeable wall, noslip boundaries
4 Inux cell, velocities have to be specied
5 Prescribed pressure cell, freeslip boundaries
6 Continuous outow cell, freeslip boundaries
7 Corner cell, no boundary conditions have to be specied
M.A. VAN DER HOEF ET AL. 102
with Re
a
red
a
uv
a
/m the Reynolds number of particle a and C
d
the drag
coefcient, for which the expression by Schiller and Nauman (1935) is used:
C
d
241 0:15Re
0:687
a
=Re
a
Re
a
o10
3
0:44 Re
a
410
3
_
Note that the validity of both the Ergun and Wen and Yu equations has re
cently been questioned on the basis of LB data, and alternative dragforce
correlations have been proposed. From LB simulations, Hill et al. (2001a, b)
suggest the following relation for Stokes ow (lim Re
a
0):
bd
2
m1
A
o
1
with
A
o
180 o0:6
18
3
1
3
2
p 1
1=2
135
64
116:141
_ _
10:6811
2
8:481
3
8:161
4
40:6
_
_
whereas for Re
a
440, they found that the drag force increases linearly with Re
a
:
bd
2
m1
A
2
1 0:6057
3
1:908
3
1 0:209
2
_
Re
a
(47)
In the paper by Hill et al. (2001b), values for A
2
are only given for a nite
number of gas fractions
1
; however, A
2
is nearly the same as A
o
(Koch and Hill,
2001). Note that in Section II.D we suggest a different expression for b, also on
the basis of lattice Boltzmann simulations.
2. Force from the Pressure Gradient
The force on particle a due to the pressure gradient rp in the gas phase is
equal to
F
p;a
V
a
rp
Note that the reaction of this force (thus the twoway coupling) is incorporated
in the momentum conservation equation of the gas phase in the rst term on the
RHS of Eq. (41). The local value for rp at r
a
is obtained from a volume
weighted averaging technique using the values of the pressure gradients at the
eight surrounding grid nodes. The volumeweighted averaging technique used to
1
In Ref. Hill et al. (2001b), values are listed for F
2
, which relates to A
2
via F
2
A
2
(1e)/(18e
3
).
MULTISCALE MODELING OF GASFLUIDIZED BEDS 103
obtain the localaveraged value
Q of a quantity Q
ijk
from the eight surrounding
computational nodes is shown in Fig. 17. The localaveraged value is calculated
as follows:
Q
Q
1
V
8
Q
2
V
7
Q
3
V
6
Q
4
V
5
Q
5
V
4
Q
6
V
3
Q
7
V
2
Q
8
V
1
DX DY DZ
where
V
1
d
x
d
y
d
z
V
2
~
d
x
d
y
d
z
V
3
d
x
~
d
y
d
z
V
4
~
d
x
~
d
y
d
z
V
5
d
x
d
y
~
d
z
V
6
~
d
x
d
y
~
d
z
V
7
d
x
~
d
y
~
d
z
V
8
~
d
x
~
d
y
~
d
z
with
~
d
x
DX dx;
~
d
y
DY dy;
~
d
z
DZ dz, and the distances d
x
, d
y
,
and d
z
necessary in this averaging techniqueare calculated from the position
of the particle in the staggered grid (see also Fig. 24). Note that the same
technique of volume weighting is also used to obtain local gas velocities and
local void fractions at the position of the center of the particle.
3. Void Fraction Calculation
From the position of each particle, we can calculate its contribution to the
local solid volume fraction e
s
in any specied uid cell. This local void fraction,
e 1 e
s
, is one of the key parameters that controls the momentum exchange
between the phases and should be determined with care.
FIG. 17. The scheme of volumeweighted averaging.
M.A. VAN DER HOEF ET AL. 104
For a 2D situation, the void fraction e(i, j) can be calculated on the basis
of the area occupied by the particles in the cell of interest. A particle can be
present in multiple cells, however, as shown in Fig. 18. Hoomans et al. (1996)
developed a method to account for the multiple cell overlap. The area of A
ii,jj
is
given by
A
ii;jj
R
2
a
d
1
d
2
1
2
d
1
1 d
2
1
_
d
2
1 d
2
2
_
arccos d
1
arcsin d
2
_ _
(48)
and area A
i,jj
by
A
i;jj
R
2
a
1
2
p d
1
d
2
1
2
d
1
1 d
2
1
_
d
2
1 d
2
2
_
arccosd
1
arccosd
2
_ _
(49)
with d
1
d
1
/R
a
and d
2
d
2
/R
a
(see Fig. 18). The area A
ii,j
can be calculated by
an equation similar to Eq. (49). However, the void fraction calculated in this
way is based on a 2D distribution of disks, whereas the empirical dragforce
correlations are derived for 3D systems. To correct for this inconsistency, the
void fraction calculated on the basis of area (e
2D
) is transformed into a 3D void
fraction (e
3D
) using the following equation:
3D
1
2
3
p
_
1
2D
3=2
FIG. 18. The multiple cell overlap of a single particle. From Hoomans, Ph.D. thesis, University of
Twente (2000).
MULTISCALE MODELING OF GASFLUIDIZED BEDS 105
In a true 3D situation, we can calculate the void fraction on the basis of actual
volume of the particles. However, no analytical expression is available for
volume V
ii,jj
. Hoomans et al. (1996) suggested the approximation
V
ii;jj
%
V
ii
V
a
V
jj
V
a
V
a
(50)
with V
ii
V
ii,jj
+V
ii,j
and V
jj
V
ii,jj
+V
i,jj
. The volume of the sphere caps V
ii
and V
jj
can be calculated exactly by
V
ii
V
a
1
4
1 d
1
2
2 d
1
V
jj
V
a
1
4
1 d
2
2
2 d
2
with d
i
d
i
/R
a
being the distance from the center of the particle to the cell
boundary relative to the radius of the particle and V
a
4pR
a
3
/3 the volume of
the particle. The error in the calculation of the porosity that is introduced by the
approximation in Eq. (50) is negligibly small when the particle radii are an order
of magnitude smaller than the size of the CFDgrid cell, which is required in any
case in order to have a gridindependent value of the porosity. In this context, it
is noteworthy that recently a new method has been developed that can generate
a gridindependent estimate of e, even when the size of the particles is of the
order of the size of the grid cells (Link et al., 2005).
E. ENERGY BUDGET
To relate the discrete particle simulations to the KTGFs, it is very useful to
analyze the detailed information of the energy evolution in the system. The total
energy balance of the system is obtained by calculating all relevant forms of
energy as well as the work performed due to the action of external forces.
N
part
a1
m
a
v
a
v
a
E
rot
kin
1
2
N
part
a1
I
a
o
a
o
a
(51)
N
part
a1
m
a
g r
a
N
part
a1
b
k
n
d
2
ab;n
k
t
d
2
ab;t
The work done by the external forces and the cohesive force in one time step
dt
W
ext
dt
N
part
a1
F
d;a
F
p;a
F
coh;a
v
a
Also, the energy dissipated during the particleparticle contact process has to
be considered and is determined by the following:
Energy dissipated by the normal and tangential spring in one time step dt
E
ds
dt
N
part
a1
b
Z
n
v
ab;n
v
ab;n
Z
t
v
ab;t
v
ab;t
N
part
a1
b
m
f
F
ab;n
t
ab;n
v
ab;t
b
F
ab
r
ab
(52)
where the sums are over all particles, with the restriction that a6b. In Eq. (52),
F
ab
is the interaction force between particles a and b. For the softsphere model
MULTISCALE MODELING OF GASFLUIDIZED BEDS 107
as presented in the previous sections, in the absence of cohesive forces, F
ab
is
equal to the sum of F
ab,n
and F
ab,t,
as given by Eqs. (26) and (21), and then only
when particles a and b are in contact. Furthermore, in Eq. (52),
y is the average
granular temperature, which can be dened as the average over the total volume
of the local granular temperature dened by Eq. (59). In the absence of any drift
velocity, m
y 2E
trans
kin
=3N, with E
trans
kin
as the total translational kinetic energy
given by Eq. (51). Note that for the hardsphere model there are no forces, and a
different procedure is required. In that case, the solids pressure (and thus the
excess compressibility) can be obtained from the average number of collisions
per unit time (Allen and Tildesley, 1990).
We have performed simulations for 500 particles with periodic boundary
conditions and no gas phase present. Owing to the inelastic collisions, the par
ticles will continuously dissipate energy, which would eventually cause the par
ticles to come to a quiescent state. In this work, we therefore drive the system by
two different techniques: (1) rescaling the particle velocities every time step,
according to the desired granular temperature; (2) accelerating the particles
randomly. Method (2) is most robust but less efcient. The rescaling procedure,
however, does not attain an equilibrium state for high solid fractions. For this
reason, the random acceleration procedure is used to simulate the denser system
with a solid fraction higher than 0.45, while the rescaling procedure is used for
lower solid fractions. For more details on the procedures, we refer to a recent
paper (Ye et al., 2005). All the parameters are normalized by the particle radius,
particle density, and granular temperature.
First, we should check whether the softsphere model gives results compa
rable to those from the hardsphere model, since the approximate theories of
granular ow are based on the latter model. To this end, we carried out several
sets of simulations with particles starting from either random positions or face
centered cubic (FCC) positions. The thermodynamic properties of the hard
sphere system for these two congurations have been well documented by many
researchers (Alder and Wainwright, 1957; Carnahan and Starling, 1969; Hoover
and Ree, 1969; Erpenbeck and Wood, 1984). In Fig. 19, we show our simulation
results for smooth, elastic, and cohesiveless spheres in periodic boundary do
mains, where at the start of the simulation the particles are placed in an FCC
grid. For such systems, Hoover and Ree (1969) observed a phase transition
from the uid state to the solid state at y 7.27. As can be seen, both the hard
sphere and softsphere simulations clearly display this transition point. For the
uid state, our simulation data from both models is in very good agreement
with the CarnahanStarling equation of state (Carnahan and Starling, 1969).
y
ES
4
s
2
2
s
1
s
3
(53)
The conclusion is that the softsphere model can be used as an alternative for
the hardsphere model, as far as the calculation of the excess compressibility is
concerned.
M.A. VAN DER HOEF ET AL. 108
Next, we consider a system of inelastic spheres (ISs). As can be seen from Eq.
(81), the KTGF predicts that the excess compressibility y
IS
of ISs is a linear
function of the coefcient of normal restitution e,
y
IS
1 e
2
y
ES
(54)
where y
ES
is the excess compressibility of elastic spheres (ESs). In Fig. 20, we
show our simulation results for the excess compressibility of ISs, both for the
softsphere and the hardsphere model. The solid fraction in the initial cong
uration is xed at 0.05. It is shown that for this dilute system, the simulation
results of both models are in very good agreement with the prediction in
Eq. (54) from the KTGF (solid line). Note that the Eq. (54) is derived under the
assumption that the particles are only slightly inelastic, i.e., e$1.0.
In Fig. 21, the excess compressibility is shown as a function of the solid
fraction for different coefcients of normal restitution e. These results are
compared with the Eq. (54), where the excess compressibility y
ES
is taken from
either the MaAhmadi correlation (Ma and Ahmadi, 1986) or the Car
nahanStarling correlation. As can be seen, the excess compressibility agrees
well with both correlations for a solid fraction e
s
up to 0.55. For extremely dense
systems, i.e., e
s
40.55, the MaAhmadi correlation presents a much better
estimate of the excess compressibility, which is also the case for purely elastic
particles (see Fig. 23).
FIG. 19. Simulation results for both the softsphere model (squares) and the hardsphere model
(the crosses), compared with the CarnahanStarling equation (solidline). At the start of the sim
ulation, the particles are arranged in a FCC conguration. Spring stiffness is K 70,000, granular
temperature is y 1.0, and coefcient of normal restitution is e 1.0. The system is driven by
rescaling.
MULTISCALE MODELING OF GASFLUIDIZED BEDS 109
Up to this point, we have neglected the cohesive van der Waals forces between
the particles, which is only justied if particles are larger than say 1 mm. Pres
ently, the van der Waals forces have not been included in the KTGF; a rst step
would be to consider the effect of such forces on the excess compressibility by
also including the interparticle force of Eq. (36) in F
ab
of Eq. (52). In Fig. 22, the
results for the excess compressibility for different Hamaker constants A are
shown. For simplicity, a coefcient of normal restitution e 1.0 is used. We
consider two different Hamaker constants: A 3.0 10
12
and A 3.0 10
10
(in units where r
s
1, R 1, and y 1). From Fig. 22, we see that for these
two Hamaker constants, the simulation results differ only slightly from the
prediction in Eq. (54), where y
ES
is calculated from the MaAhmadi correlation,
which suggests that cohesion has only a weak inuence on the excess com
pressibilityat least for the values of Hamaker constant that we studied. In this
context, it should be noted that the quantication of the cohesive force is not
straightforward since there is no reference force (such as gravitational force) in
these systems. We consider these systems as slightly cohesive since the ratio of
the cohesive potential and the average kinetic energy per particle is small, i.e.,
j 6.25 10
8
$6.25 10
6
. At the same time, the ratio between the cohesive
force and contact force ranges from 1.11 10
5
to 1.11 10
3
. If a strong
cohesive force is present, particles in the system may form complicated struc
tures, whereas a homogeneous state is one of the basic assumptions underlying
0.0 0.2 0.4 0.6 0.8 1.0
e
0.5
0.6
0.7
0.8
0.9
1.0
y
I
S
/
y
E
S
y
IS
= y
ES
(1+e)/2
Hardsphere
Softsphere
FIG. 20. Excess compressibility y
IS
for a system of inelastic hard spheres, as function of the
coefcient of normal restitution, for one solid fraction (e
s
0.05). The excess compressibility has
been normalized by the excess compressibility y
ES
of the elastic hard spheres system. Other sim
ulation parameters are as in Fig. 19.
M.A. VAN DER HOEF ET AL. 110
the KTGF. A more detailed analysis of the effect of the cohesive force on the
excess compressibility can be found in Ref. Ye et al. (2005).
IV. TwoFluid Model
A. INTRODUCTION
In the EulerEuler models, i.e., the TFMs, it is assumed that both the gas and
the solid phase are interpenetrating continua. This continuous approach is es
pecially useful and computationally costeffective when the volume fractions of
the phases are comparable, or when the interaction within and between the
phases plays a signicant role in determining the hydrodynamics of the system.
As discussed before, it is relatively straightforward to model the gas phase, for
instance by the use of wellestablished CFD techniques. The challenge is to
establish an accurate hydrodynamic description of the particulate phase.
0.0 0.2 0.4 0.6
Solid fraction
s
0
10
20
30
40
y
I
S
0.0 0.2 0.4 0.6
Solid fraction
s
0
10
20
30
40
y
I
S
0.0 0.2 0.4 0.6
Solid fraction
s
0
10
20
30
40
y
I
S
0.0 0.2 0.4 0.6
Solid fraction
s
0
10
20
30
40
y
I
S
e = 1.00
e = 0.95
e = 0.90
e = 0.80
FIG. 21. The excess compressibility from softsphere simulations, with random initial particle
positions, for different coefcients of normal restitution e: (a) e 1.0 (topright); (b) e 0.95 (top
left); (c) e 0.90 (bottomright); (d) e 0.80 (bottomleft). The simulation results (symbols) are
compared with Eq. (54) based on the MaAhmadi correlation (solid line) or the CarnahanStarling
correlation (dashed line). The spring stiffness is set to k
n
70,000.
MULTISCALE MODELING OF GASFLUIDIZED BEDS 111
Anderson and Jackson (1967, 1968, 1969) and Ishii (1975) have separately de
rived the governing equations for TFMs from rst principles. Although the
details of constructing the averaged equations are different, the nal equations
are essentially the same. The TFMs differ signicantly from each other as
different closures for the solid stress tensor are used.
There are basically three types of approaches to dene the solid stress tensor,
or more specically the solid viscosity. In the early hydrodynamic models
developed by Jackson and his coworkers (Anderson and Jackson, 1967; And
erson et al., 1995), Kuipers et al., (1992), and Tsuo and Gidaspow (1990)the
viscosity is dened as an empirical constant, and also the dependence of the
solid phase pressure on the solid volume fraction is determined from experi
ments. The advantage of this model is its simplicity, the drawback is that it does
not take into account the underlying characteristics of the solid phase rheology.
In another class of models, pioneered by Elghobashi and AbouArab (1983)
and Chen (1985), a particle turbulent viscosity, derived by extending the concept
of turbulence from the gas phase to the solid phase, has been used. This is the
socalled k model, where the k corresponds to the granular temperature and
is a dissipation parameter for which another conservation law is required. By
coupling with the gas phase k turbulence model, Zhou and Huang (1990)
developed a k model for turbulent gasparticle ows. The k models do not
FIG. 22. The effect of the cohesive force on the excess compressibility. The coefcient of normal
restitution is e 1.0, and granular temperature is T 1.0. The Hamaker constant is A 3.0 10
12
(circles) and 3.0 10
10
(crosses).
M.A. VAN DER HOEF ET AL. 112
include the effect of particleparticle collisions, and so these models are re
stricted to dilute gasparticle ows.
Signicant contributions to the modeling of gassolid ows have been made
by Gidaspow and coworkers (1994), who combined the kinetic theory for the
granular phase with continuum representations for the particle phase. There are
a number of other studies using this approach. Sinclair and Jackson (1989)
predicted the coreannular regime for steady developed ow in a riser. Ding et
al. (1990) simulated a bubbling uidized bed. Transient simulations and com
parisons to data were done by Samuelsberg and Hjertager (1996). Nieuwland et
al. (1996) investigated a circulating uidized bed using the KTGF. Detamore et
al. (2001) have performed an analysis of scaleup of circulating uidized beds
using kinetic theory.
One of the strengths of the KTGF, although still under development, is that it
can offer a very clear physical picture with respect to the key parameters (e.g.,
particle pressure, particle viscosity, and other transport coefcients) that are
used in the TFMs. The TFMs based on KTGF requires less ad hoc adjustments
compared to the other two types of models. Therefore, it is the most promising
framework for modeling engineeringscale uidized beds.
B. GOVERNING EQUATIONS
In the TFM, both the gas phase and the solid phase are described as fully
interpenetrating continua using a generalized form of the NavierStokes equa
tions for interacting uids. The continuity and momentum equations for the gas
phase are given by expressions identical to Eqs. (40) and (41), except for the
gassolid interaction term:
@r
@t
r ru 0 (55)
@ru
@t
r ruu rp bu u
s
r s rg (56)
with t as the viscous stress tensor of the gas phase given by Eq. (43). The
continuity and momentum equations for the particle phase are given by a sim
ilar set of equations:
@
s
r
s
@t
r
s
r
s
u
s
0 (57)
@
s
r
s
u
s
@t
r
s
r
s
u
s
u
s
s
rp rp
s
bu u
s
r s
s
s
r
s
g (58)
where e
s
1e and u
s
is the velocity of the solid phase. Note that r
s
is the
material density of the solid phase, so that the local mass per unit volume is
equal to r
s
e
s
. Obviously, the numerical scheme for updating the solid phase is
now analogous to (and synchronous with) that of the gas phase, the details of
MULTISCALE MODELING OF GASFLUIDIZED BEDS 113
which are given in Section IV.E. Since the concept of particles has disappeared
completely in such a modeling, the effect of particleparticle interactions can
only be included indirectly, i.e., via the effective solid pressure p
s
and the effec
tive solid stress tensor s
s
. A description that allows for a more detailed de
scription of particleparticle interactions follows from the KTGF, which
expresses the pressure and the solid stress tensor as a function of the local
granular temperature y, which is dened from the uctuation in the velocity of
the individual solid particles. More precisely, the granular temperature at r is
dened as
y
1
3
1
N
r
N
r
a1
v
a
u
s
2
_ _
(59)
where /.S is an ensemble average, and the sum is over all N
r
particles in a small
control volume dV around r. Note that also the solid density and velocity as
they appear in Eqs. (57) and (58) can be dened from the positions and mom
enta of the individual particles by similar type of averages
2
:
s
r
s
1
dV
N
r
a1
m
a
_ _
s
r
s
u
s
1
dV
N
r
a1
m
a
v
a
_ _
(60)
For particles of equal mass, we thus have e
s
r
s
mn with n the local number
density of particles. From the KTGF, the time evolution of the granular tem
perature is given by
3
2
@
@t
s
r
s
y r
s
r
s
yu
s
_ _
p
s
I t
s
: ru
s
r q
s
3by g (61)
with q
s
the kinetic energy ux and g the dissipation of kinetic energy due to
inelastic particle collisions. In Eqs. (58) and (61), there are three unknown
quantities (pressure, stress tensor, and energy ux), which must be expressed in
terms of the three basic hydrodynamic variables (density, velocity, and tem
perature), in order to get a closed set of equations. This is the subject of the
KTGF, and the resulting closures will be presented in Section IV.D. However,
before doing so, we will rst give a brief description of the general principles of
kinetic theory.
2
Note that for dV0 the local density and momentum density can be written as
e
s
p
s
a
m
a
d(r r
i
) and e
s
p
s
u
s
a
m
a
v
a
d (r r
i
), which are the expressions that are usually found
in literature.
M.A. VAN DER HOEF ET AL. 114
C. GENERAL KINETIC THEORY
In this section, we will only discuss the basic principles of kinetic theory,
where for detailed derivations we refer to the classic textbook by Chapman and
Cowling (1970), and a more recent book by Liboff (1998). Of central impor
tance in the kinetic theory is the single particle distribution function f
s
(r, v),
which can be dened as the number density of the solid particles in the 6D
coordinate and velocity space. That is, f
s
(r, v, t) dv dr is the average number of
particles to be found in a 6D volume dv dr around r, v. This means that the
local density and velocity of the solid phase in the continuous description are
given by
r
s
r; t
_
1
1
m f
s
r; v; tdv (62)
and
r
s
r; tu
s
r; t
_
1
1
mv f
s
r; v; tdv (63)
where the local density is dened as r
s
r
s
s
with r
s
as the material density of
the solid particles. The granular temperature, dened by Eq. (59), follows from
r
s
r; tyr; t
1
3
_
1
1
mv u
s
2
f
s
r; v; tdv (64)
The evolution of the oneparticle distribution function f
s
can be described by the
Boltzmann equation
@
@t
f
s
r; v; t v rf
s
r; v; t C (65)
which is basically a continuity equation, where the second term on the lefthand
side (LHS) represents the change of f
s
in time due to streaming and the collision
function C on RHS represents the change of f
s
due to particleparticle inter
actions. Conservation of mass, momentum, and energy in a collision gives that
C should satisfy
_
1
1
Cdv 0;
_
1
1
Cvdv 0;
_
1
1
Cu
2
dv 0
Taking the same integrals (
_
ydv,
_
yvdv, and
_
yu
2
dv) of the Boltzmann
equation Eq. (65), making use of Eqs. (62) and (63), yields
@
@t
r
s
r r
s
u
s
0 (66)
MULTISCALE MODELING OF GASFLUIDIZED BEDS 115
@
@t
r
s
u
s
r r
s
u
s
u
s
r p (67)
3
2
@
@t
r
s
y r r
s
yu
s
_ _
p : ru
s
r q
s
(68)
with
p
_
1
1
mVVf
s
r; v; tdv; q
_
1
1
mV
2
2
Vf
s
r; v; tdv; V v u
s
In principle, one should solve the Boltzmann equation Eq. (65) in order to arrive
at explicit expressions for the pressure tensor p and heat ux q, which proves not
possible, not even for the simple BGK equation Eq. (11). However, one can
arrive at an approximate expression via the ChapmanEnskog expansion, in
which the distribution function is expanded about the equilibrium distribution
function f
s
eq
, where the expansion parameter is a measure of the variation of the
hydrodynamic elds in time and space. To second order, one arrives at the
familiar expression for p and q
p p
s
I s
s
; q
s
k
s
ry (69)
with
I is the unit tensor, and
s
s
m
s
ru
s
ru
s
T
_
l
s
2
3
m
s
_ _
r u
s
I
_
(70)
where ru
ab
r
a
u
b
; ru
T
ab
r
b
u
a
. Inserting the above expression for p and q
into Eqs. (67) and (68) will give the NavierStokes equations, where the pa
rameters k
s
, l
s
, m
s
, and ps can be calculated (at least in princeiple) when the
collision function C is known. For the simple BGK equation Eq. (11), this will
result in the relations of Eq. (13). For an accurate description of the solid phase,
however, one requires a much more detailed expression for C, which contains
the details of the particleparticle interactions. Although this is a laborious
route, it opens a possibility for making a link between the microscopic details
of particle collisions and the macroscopic transport coefcients. Apart from
the details of the particleparticle interactions, C does also depend on the joint
probability function f
2
s
(r
1
, v
1
, r
2
, v
2
, t), provided that the interactions between
the particles are pairwise additive (generally for nbody interactions, C will
depend on f
s
(n)
). In order to get a closed equation, f
2
s
should be described in
term of f
s
. If the velocities v
1
and v
2
are not correlated, one can write
f
2
s
r
1
; v
1
; r
2
; v
2
; t gr
12
;
s
f
s
r
1
; v
1
; t f
s
r
2
; v
2
; t
M.A. VAN DER HOEF ET AL. 116
where g(r
12
, e
s
) is the pair distribution function, which depends only on the
distance r
12
jr
2
r
1
j and the solid fraction.
For sufciently low density, g 1, the collision function takes the form
C
1
m
2
_
dOsO
_
dv
0
v v
0
f
s
~ vf
s
~ v
0
f
s
vf
s
v
0
_
(71)
where we have omitted the r, t argument of f
s
. In Eq. (71), ~ v, ~ v
0
, are the velocities
of the two particles involved after the collision, which can be constructed from
the initial velocities v, v
0
from conservation of energy and momentum:
~ v v aa v v
0
~ v
0
v
0
aa v
0
v
with a as the unity vector along the line connecting the two centers of the
particle before the collision. Furthermore, in Eq. (71), s(O) represents the
crosssection and O is the solid angle in which the particle is scattered. More
details on these concepts can be found in the standard literature (Chapman and
Cowling, 1970; Liboff, 1998). Using this form of the collision function, it can be
shown that pressure p
s
, shear viscosity m
s
, and thermal conductivity k
s
in Eqs.
(69) and (70) are given by
m
id
s
5
96
pr
s
d
y
p
_
; k
id
s
75
384
pr
s
d
y
p
_
; p
id
s
s
r
s
y (72)
where d is the diameter of the particles, and the superscript id (ideal) indicates
that the expressions are for the limit of a dilute gas, for which the pressure is
given by the ideal gas law.
For high densities, g cannot be set equal to one, and the collision function
becomes much more complex and so is not given here. It turns out, however,
that instead of using the full radial distribution function, it is sufcient to use
the value at contact r R, so that we dene a new function:
w
s
gR;
s
In the standard Enskog theory (SET), the shear viscosity and thermal con
ductivity of ESs are found to be equal to
3
m
ES
s
m
id
s
1
wbr
s
4
5
0:7614wbr
s
_ _
br
s
(73)
3
See Chapman and Cowling (1970). Note that the true expression for m
ES
s
reads
m
ES
s
c
1
m
id
s
1
xbr
s
4
5
4
25
1
12
pc
2
wbr
s
br
s
, with c
1
c
2
1.016. In most expressions in literature, c
1
is set equal to 1; in expression Eq. (82) of Gidaspow, both c
1
c
2
1, which is the cause of the
slightly different coefcient 0.771, compared to 0.7614 in Eq. (75). For practical purposes, the
difference is negligible. Similar remarks can be made about k.
MULTISCALE MODELING OF GASFLUIDIZED BEDS 117
k
ES
s
k
id
s
1
wbr
s
6
5
0:7574wbr
s
_ _
br
s
(74)
with b 2e
s
pd
3
/3 m, so that bp
s
4e
s
. Note that the pressure of a dense system
is directly related to the radial distribution function at contact (Chapman and
Cowling, 1970; Hansen and McDonald, 1986):
p
ES
s
p
id
s
1 y
ES
y
ES
wbr
s
4w
s
with y
ES
the excess compressibility of the elastic hardsphere system. Thus in the
Enskog theory, the transport coefcients are completely determined by the y
ES
:
m
ES
s
4m
id
s
s
1
y
ES
4
5
0:7614y
ES
_ _
(75)
k
ES
s
4k
id
s
s
1
y
ES
6
5
0:7574y
ES
_ _
(76)
Various expressions for y
ES
have been proposed in literature based on the virial
coefcients and simulation data. Most of these have the following general form:
y
ES
n0
c
n
4
s
n1
1
s
=
cp
b
(77)
with e
cp
the closepacked solid fraction, at which the pressure diverges. In
Table III, we summarize the parameters found by different authors. A com
parison of expression in Eq. (77) with the MD simulation data from Alder and
Wainwright (1960) and Woodcock (1981) is shown in Fig. 23. In our current
TABLE III
VALUES FOR THE PARAMETERS IN EQ. (77)
CS MA SSM TC
e
cp
1 0.64356 0.6435 0.6875
a 1 3 1 1
b 3 0.67802 0.76 1
c
o
1 1 1 1
c
1
1/8 0.625 0.3298 0.2613
c
2
0 0.2869 0.08867 0.05968
c
3
0 0.070554 0.01472 0.005905
c
4
0 0 0.0005396 0.001191
c
5
0 0 0.0003574 0.0004455
c
6
0 0 0.0005705 0.0004818
c
7
0 0 0.0001212 0.00003636
c
8
0 0 0.0001151 0.00008182
Note: CS: Carnahan and Starling (J. Chem. Phys. 51, 635 (1969)); MA: Ma and Ahmadi (J. Chem.
Phys. 84, 3449 (1986)); SSM: Song, Stratt and Mason (J. Chem. Phys. 88, 1126 (1988)); TC: To
bochnik and Chapin (J. Chem. Phys. 88, 5824 (1988)).
M.A. VAN DER HOEF ET AL. 118
version of the TFM, we use the expression by Ma and Ahmadi (1986) (see also
Fig. 21). Alder et al. (1970) have also measured the shear viscosity in MD
simulations of dense hardsphere systems. It was found that the Enskog ap
proximation in Eq. (75) is very accurate up to e
s
0.3; however, for higher solid
fractions the theory signicantly underestimates the shear viscosity up to a
factor of two for e
s
E0.5.
D. KINETIC THEORY OF GRANULAR FLOW
In the KTGF, the dissipation of energy in collisions is included in the Enskog
theory. Currently, only the effect of the coefcient of normal restitution has
been considered, although it is anticipated that friction also plays an important
role. The derivation of the constitutive equations for ISs can be found in the
book by Gidaspow (1994) and the papers by Jenkins and Savage (1983), Lun
et al. (1984), Ding and Gidaspow (1990), and Nieuwland et al. (1996). Here,
we will present the expressions for p
s
, m
s
, and K
s
from the book of Gidaspow
(1994) (Eqs. (T.9.1), (9.183), (9.250), (9.262), (9.268), and (9.272)):
p
IS
s
1 21 e
s
g
s
r
s
y (78)
m
IS
s
5
96
pr
s
d
y
p
_
2
1 eg
1
4
5
1 e
s
g
_ _
2
4
5
2
s
r
s
dg1 e
y
p
_
(79)
0.4 0.5 0.6 0.7
solid fraction
s
0
20
40
60
80
y
E
S
CarnahanStarling
MaAhmadi
SongStrattMason
MD (Woodcock)
MD (Alder & Wainwright)
FIG. 23. Comparison of the expressions from Eq. (77) and Table III with data from MD sim
ulations.
MULTISCALE MODELING OF GASFLUIDIZED BEDS 119
k
IS
s
75
384
pr
s
d
y
p
_
2
1 eg
1
6
5
1 e
s
g
_ _
2
2
2
s
r
s
dg1 e
y
p
_
(80)
where g is the value of the radial distribution function of a hardsphere uid at
contact and e is the coefcient of normal restitution. Note that from Eq. (78) it
follows that the excess compressibility of the IS system is equal to
y
IS
21 e
s
g
1 e
2
y
ES
(81)
that is, the dissipation in the collisions reduces the excess compressibility by a
factor of (1+e)/2. Replacing 2(1+e)e
s
g in Eqs. (79) and (80) by y
IS
and using
expression in Eq. (72) for m
s
id
and k
s
id
gives
m m
id
s
4
s
1
y
IS
1
2
5
y
IS
_ _
2
48
25p
y
IS
_ _
m
id
s
4
s
1
y
IS
4
5
0:771y
IS
_ _
(82)
k k
id
s
4
s
1
y
IS
1
3
5
y
IS
_ _
2
32
25p
y
IS
_ _
k
id
s
4
s
1
y
IS
6
5
0:767y
IS
_ _
(83)
which are of the same form as the Enskog expressions in Eqs. (75) and (76), with
y
ES
replaced by y
IS
.
3
It thus turns out, like for the elastic hard spheres, that the
constitutive equations are completely determined by the excess compressibility,
and that the general form of the Enskog equations is not affected by the dis
sipation of energy in the collisions.
Note that in the granular temperature equation Eq. (61), there is one extra
term that is absent in the SET, namely the dissipation of uctuating kinetic
energy g. From the KTGF follows that
g
3
2
1 ey
IS
r
s
s
y
4
d
y
p
_
r u
s
_ _
E. NUMERICAL SOLUTION METHOD
Owing to the tendency of inelastic particles to contract in highdensity clus
ters, and the strong nonlinearity of the particle pressure near the maximum
packing density, special attention has to be paid to the numerical implemen
tation of the model equations. Most classic constant property TFMs are
solved using computational methods based on the implicit continuous Eulerian
(ICE) method pioneered by Harlow and Amsden (1975). The implementation is
based on a nite difference technique and the algorithms closely resemble the
SIMPLE algorithm (Patankar and Spalding, 1972), whereby a staggered grid is
employed to reduce numerical instability. A detailed discussion on the appli
cation of this numerical technique to TFMs for gasuidized beds is presented
by Kuipers et al. (1992).
M.A. VAN DER HOEF ET AL. 120
In principle, this numerical solution method can be straightforwardly applied
to modern TFMs with closure laws according to the KTGF. However, when
doing so, the numerical stability of the TFM is highly affected by the value of
the coefcient of normal restitution. Problems that can be handled with ac
ceptable time steps of 10
4
s for ideal particles (e 1) require time steps of 10
5
s when the coefcient of normal restitution is taken to be 0.97, and unacceptably
small time steps of 10
6
s have to be taken when the coefcient of normal
restitution is reduced below 0.93. This extreme sensitivity to the value of the
coefcient of normal restitution is caused by the fact that particle volume frac
tions at the next time level are estimated without taking into account the strong
nonlinear dependence of the particle pressure on the particle volume fraction. A
new numerical algorithm, which estimates the new particle volume fraction
taking the compressibility of the particulate phase more directly into account, is
presented in this section.
1. Discretization of the Governing Equations
The set of conservation equations, supplemented with the constitutive equa
tions, boundary, and initial conditions cannot be solved analytically, and a
numerical method must be applied to obtain an approximate solution. There
fore, the domain of interest is divided into a number of xed Eulerian cells
through which the gassolid dispersion moves. A standard nite difference
technique is applied to discretize the governing equations.
4
The cells are labeled
by indices i, j, and k located at their centers, and a staggered grid conguration
is applied. According to this conguration the scalar variables are dened at the
cell centers, whereas the velocities are dened at the cell faces, as indicated in
Fig. 24. Furthermore, different control volumes have to be applied for mass and
granular energy conservation on one hand and the momentum conservation
equations on the other. The control volumes for mass and granular energy
conservation coincide with the Eulerian cells, whereas the control volumes for
momentum conservation in all three directions are shifted half a cell with re
spect to the Eulerian cells. Applying rstorder time differencing and fully im
plicit treatment of the convective uxes, the discretized form of continuity
equation for the solid phase, Eq. (57), becomes
s
r
s
n1
i;j;k
s
r
s
n
i;j;k
dt
dx
h
s
r
s
u
s;x
i
n1
i
1
2
;j;k
h
s
r
s
u
s;x
i
n1
i
1
2
;j;k
_ _
dt
dy
h
s
r
s
u
s;y
i
n1
i;j
1
2
;k
h
s
r
s
u
s;y
i
n1
i;j
1
2
;k
_ _
dt
dz
h
s
r
s
u
s;z
i
n1
i;j;k
1
2
h
s
r
s
u
s;z
i
n1
i;j;k
1
2
_ _
0 84
4
This part is based upon Chapter 2 of the thesis of Goldschmidt (2001).
MULTISCALE MODELING OF GASFLUIDIZED BEDS 121
where the superscripts n and n+1 indicate that the quantities are at the old and
the new time, respectively. For the discretization of all convective mass,
momentum, and uctuating kinetic energy uxes the secondorder accurate
Barton scheme (Centrella and Wilson, 1984; Hawley et al., 1984) is applied. A
schematic representation of this scheme for the convective transport of a
quantity D (e.g., er) by a velocity V
i+1/2
(e.g., u
x
) is given in Fig. 25. In the
discretization of the momentum Eq. (58), the terms associated with the gas and
solid pressure gradients are treated fully implicitly. The interphase momentum
transfer term is treated in a linear implicit fashion, and all other terms are
treated explicitly. The discretization of the solid phase momentum in Eq. (58)
for the xdirection is given by
s
r
s
u
s;x
n1
i
1
2
;j;k
A
n
i
1
2
;j;k
s
n1
i
1
2
;j;k
dt
dx
p
n1
i1;j;k
p
n1
i;j;k
_ _
dt
dx
p
s
n1
i1;j;k
p
s
n1
i;j;k
_ _
dtb
n
i
1
2
;j:k
u
x
u
s;x
n1
i
1
2
;j;k
85
where momentum convection, viscous interaction, and gravity have been
collected in the explicit term A
n
. The equation for the ydirection is obtained by
FIG. 24. Positions at which the key variables are evaluated for a typical computational cell in the
staggeredgrid conguration.
M.A. VAN DER HOEF ET AL. 122
substituting y for x, B for A, and a change of subscripts:
. . .
i
1
2
;j;k
) . . .
i;j
1
2
;k
. . .
i1;j;k
) . . .
i;j1;k
and the equation for the zdirection is obtained by the substituting z for x, C for
A, and a change of subscripts
. . .
i
1
2
;j;k
) . . .
i;j;k
1
2
. . .
i1;j;k
) . . .
i;j;k1
Note that the mass and momentum equations for the gas phase can simply be
obtained by replacing e
s
e, r
s
r, u
s
u in Eqs. (84) and (85), and dropping
the terms concerning the particlepressure gradient.
The granular energy equation is solved in a fully implicit manner. The
solution of the equation however proceeds through a separate iterative
procedure that solves the granular temperature equations for the whole
computational domain when this is required by the main solution procedure
discussed in the next paragraph. In this separate iterative procedure, the terms
regarding convective transport and generation of uctuating kinetic energy by
viscous shear are explicitly expressed in terms of the most recently obtained
granular temperature y*. The granular energy dissipation term is treated in a
semiimplicit manner, whereas all other terms are treated fully implicitly. The
FIG. 25. Schematic representation of the Barton scheme for the convective ux of a quantity D by
velocity V
i+1/2
in the xdirection.
MULTISCALE MODELING OF GASFLUIDIZED BEDS 123
applied discretization of the granular temperature equation is given by
3
2
s
r
s
y
n1
i;j;k
3
2
s
r
s
y
n
i;j
D
i;j;k
dt3b
n
i;j;k
y
n1
i;j;k
dt
g
y
_ _
i;j;k
y
n1
i;j;k
p
s
n1
i;j;k
dt
dx
u
s;x
i
1
2
;j;k
u
s;x
i
1
2
;j;k
_ _
p
s
n1
i;j;k
dt
dy
u
s;y
i;j
1
2
;k
u
k;y
i;j
1
2
;k
_ _
p
s
n1
i;j;k
dt
dx
u
s;z
i;j;k
1
2
u
k;y
i;j;k
1
2
_ _
dt
dx
k
s
i
1
2
;j;k
1
dx
y
n1
i1;j;k
y
n1
i;j;k
_ _
k
s
i
1
2
;j;k
1
dx
y
n1
i;j;k
y
n1
i1;j;k
_ _
_ _
dt
dx
k
s
i;j
1
2
;k
1
dy
y
n1
i;j1;k
y
n1
i;j;k
_ _
k
s
i;j
1
2
;k
1
dy
y
n1
i;j;k
y
n1
i;j1;k
_ _
_ _
dt
dz
k
s
i;j;k
1
2
1
dz
y
n1
i;j;k1
y
n1
i;j;k
_ _
k
s
i;j;k
1
2
1
dz
y
n1
i;j;k
y
n1
i;j;k1
_ _
_ _
86
In this equation, the superscript (*) indicates that a term is computed based
upon the most recent information, which complies with the (n+1)th time level
when all iterative loops have converged. Further, the convective transport and
viscous generation of uctuating kinetic energy have been collected in the
explicit term D*. The iterative solution procedure for the granular energy
equations continues until the convergence criteria
y
n1
i;j;k
y
i;j;k
oe
y
y
n1
i;j;k
(87)
are simultaneously satised for all cells within the computational domain. For a
typical value of e
y
10
6
, this takes only a couple of iterations per time step.
2. Solution Procedure of the Finite Difference Equations
The numerical solution of the discretized model equations evolves through a
sequence of computational cycles, or time steps, each of duration dt. For each
computational cycle, the advanced (n+1)level values at time t+dt of all key
variables have to be calculated for the entire computational domain. This
calculation requires the old nlevel values at time t, which are known from either
the previous computational cycle or the specied initial conditions. Then each
computational cycle consists of two distinct phases:
i;j;k
s
r
s
i;j;k
s
r
s
n
i;j;k
dt
dx
h
s
r
s
u
s;x
i
i
1
2
;j;k
h
s
r
s
u
s;x
i
i
1
2
;j;k
_ _
dt
dy
h
s
r
s
u
s;y
i
i;j
1
2
;k
h
s
r
s
u
s;y
i
i;j
1
2
;k
_ _
dt
dz
h
s
r
s
u
s;z
i
i;j;k
1
2
h
s
r
s
u
s;z
i
i;j;k
1
2
_ _
88
D
g
i;j;k
r
i;j;k
r
n
i;j;k
dt
dx
hru
x
i
i
1
2
;j;k
hru
x
i
i
1
2
;j;k
_ _
dt
dy
hru
y
i
i;j
1
2
;k
hru
y
i
i;j
1
2
;k
_ _
dt
dz
hru
z
i
i;j;k
1
2
hru
z
i
i;j;k
1
2
_ _
89
If the convergence criteria
D
g
i;j;k
oe
g
r
i;j;k
(90)
D
s
i;j;k
oe
s
s
r
s
i;j;k
(91)
are not satised for all computational cells (typically e
g
e
s
10
6
), a whole
eld pressure correction is calculated, satisfying
D
g
i;j;k
J
g
n
i;j;k
dp
i;j;k
J
g
n
i1;j;k
dp
i1;j;k
J
g
n
i1;j;k
dp
i1;j;k
J
g
n
i;j1;k
dp
i;j1;k
J
g
n
i;j1;k
dp
i;j1;k
J
g
n
i;j;k1
dp
i;j;k1
J
g
n
i;j;k1
dp
i;j;k1
92
where (J
g
)
n
represents the Jacobi matrix for the gas phase. This matrix contains
the derivatives of the defects D
g
with respect to the gas phase pressure, for which
explicit expressions can be obtained from the continuity equation for the gas
phase in combination with the momentum equations. To save computational
effort, the elements of the Jacobi matrix are evaluated at the old time level. The
banded matrix problem corresponding to Eq. (92) is solved using a standard
ICCG sparse matrix technique. Once new pressures have been obtained, the
corresponding new gas phase densities are calculated.
So far, the solution procedure has been exactly the same as the SIMPLE
procedure that is usually applied for the solution of the classic TFMs with
MULTISCALE MODELING OF GASFLUIDIZED BEDS 125
constant property closure equations. In the next step however, the standard
procedures continue with the computation of the new velocities from the
coupled momentum equations, after which the new volume fractions are
obtained from the solid phase mass balances, and only then the new solid
pressures are determined. This regularly leads to excessive compaction and
extremely high particle pressures in areas where the particle packing densities
are close to random close packing. Therefore, the new solution procedure
computes the particle volume fractions, taking the compressibility of the solid
phase more directly into account. Similar to the pressure correction for the gas
phase, a whole eld particle volume fraction correction is computed, satisfying
D
s
i;j;k
J
s
n
i;j;k
d
s
i;j;k
J
s
n
i1;j;k
d
s
i1;j;k
J
s
n
i1;j;k
d
s
i1;j;k
J
s
n
i;j1;k
d
s
i;j1;k
J
s
n
i;j1;k
d
s
i;j1;k
J
s
n
i;j;k1
d
s
i;j;k1
J
s
n
i;j;k1
d
s
i;j;k1
93
In this Eq. (J
s
)
n
is the Jacobi matrix for the solid phase, which contains the
derivatives of the mass residuals for the particulate phase to the solid volume
fraction. Explicit expressions for the elements of the Jacobi matrix can be
obtained from the continuity for the solid phase and the momentum equations.
For example for the central element, the following expression is obtained from
the solid phase continuity equation, in which the convective terms are evaluated
with central nite difference expressions:
J
s
n
i;j;k
@D
s
i;j;k
@
s
i;j;k
r
s
i;j;k
dt
dx
@
s
r
s
u
s;x
i
1
2
;j;k
@
s
i;j;k
@
s
r
s
u
s;x
i
1
2
;j;k
@
s
i;j;k
_ _
dt
dy
@
s
r
s
u
s;y
i;j
1
2
;k
@
s
i;j;k
@
s
r
s
u
s;y
i;j
1
2
;k
@
s
i;j;k
_ _
dt
dz
@
s
r
s
u
s;z
i;j;k
1
2
@
s
i;j;k
@
s
r
s
u
s;z
i;j;k
1
2
@
s
i;j;k
_ _
94
The derivatives of the mass uxes to the solid volume fractions can
subsequently be obtained from the solid phase momentum equations. From
Eq. (85), the discretized xmomentum equation, the derivatives of the mass
uxes in the xdirection can easily be obtained, e.g.,
@
s
r
s
u
s;x
i
1
2
;j;k
@
s
i;j;k
1
2
dt
dx
p
i1;j;k
p
i;j;k
_ _
dt
dx
@p
s
@
s
_ _
i;j;k
dtb
n
i
1
2
;j;k
@u
x
u
s;x
i
1
2
;j;k
@
s
i;j;k
95
M.A. VAN DER HOEF ET AL. 126
The second term on the RHS of this equation shows that the compressibility
of the solid phase is taken directly into account in the estimation of the new
particle volume fractions. Furthermore, the expression for the derivatives of the
velocities to the solid pressure can be obtained by combination with the x
momentum equation for the gas phase that results in
@ru
x
i
1
2
;j;k
@
s
i;j;k
1
2
dt
dx
p
i1;j;k
p
i;j;k
_ _
dtb
n
i
1
2
;j;k
@u
x
u
s;x
i
1
2
;j;k
@
s
i;j;k
(96)
Together with Eq. (93), this equation forms a set of equations from which
explicit expressions for the derivatives of the velocities can readily be obtained.
Expressions for the y and zdirection and for the other elements of the Jacobi
matrix are obtained in a similar manner.
After the new solid volume fractions have been obtained from Eq. (93), new
particle pressures are calculated, where after new velocities can be obtained
from the coupled momentum equations. Next, new granular temperatures are
calculated from the granular energy equations by an iterative procedure
described in Section IV.E.1. Finally, the new mass residuals (D
g
)
i,j,k
and (D
s
)
i,j,k
are computed and the convergence criteria are checked again.
Though this new algorithm still requires some time step renement for
computations with highly inelastic particles, it turns out that most computations
can be carried out with acceptable time steps of 10
5
s or larger. An alternative
numerical method that is also based on the compressibility of the dispersed
particulate phase is presented by Laux (1998). In this socalled compressible
dispersephase method the shear stresses in the momentum equations are
implicitly taken into account, which further enhances the stability of the code in
the quasistatic state near minimum uidization, especially when frictional shear
is taken into account. In theory, the stability of the numerical solution method
can be further enhanced by fully implicit discretization and simultaneous solution
of all governing equations. This latter is however not expected to result in faster
solution of the TFM equations since the numerical efforts per time step increase.
F. APPLICATION TO GELDART A PARTICLES
A great challenge in CFD modeling of gassolid twophase ows is to obtain
realistic predictions of the uidization behavior of small particles such as
Geldart A particles (Geldart, 1973), for which the standard TFM has so far
failed to predict even the bubbling uidization. Ferschneider and Mege (1996)
found a major overprediction of bed expansion in a bubbling bed of FCC
particles, and Bayle et al. (2001) obtained the same results in a turbulent bed of
FCC particles. Recently, Lettieri et al. (2003) used a particlebed model,
originally developed by Chen et al. (1999), to investigate the homogeneous
uidization of Geldart A particles. It has been demonstrated that a
MULTISCALE MODELING OF GASFLUIDIZED BEDS 127
homogeneous expansion can be obtained in this particlebed model. However in
this model, an articial particlephase elasticity force is required. McKeen and
Pugsley (2003) used the twouid CFD code MFIX to simulate a freely
bubbling bed of FCC catalyst for U
0
0.050.2 m/s and compared their
simulation results with ECT data. In accordance with ndings of Ferschneider
and Mege (1996), McKeen and Pugsley (2003) also found that the standard
CFD model greatly overpredicted bed expansion without any modications of
the drag closures. By using a scale factor of 0.25 for the commonly used
gassolid drag laws, they found that their simulation results are in accordance
with experimental observations. They argued that this is due to the formation of
clusters with a size smaller than the CFD grid size. Such smallscale clusters
have not been reported before, in particular for particles with a size of 75 mm.
Although the van der Waals force can play a role in the uidization of
Geldart A particles, it is not clear how this force affects the gassolid drag. The
inuence of the cohesion on the KTGF has not been carefully checked so far.
Recently, Kim and Arastoopour (2002) tried to extend the kinetic theory to
cohesive particles; however, their nal expression for the particular phase stress
is very complex. A simpler route would be to assume that the Enskog expres
sions in Eqs. (75)(76) still hold for cohesive particles, only with a modied
excess compressibility. However at present, it proves difcult to give an accurate
estimate of the deviation of y due to the cohesive force (see Fig. 22). Moreover,
as discussed in Section III.F, also the magnitude of the cohesive force itself (i.e.,
the Hamaker constant) is not known. For this reason, we will only study the
effect of the gasparticle drag in this section, where we use two different models:
(i) the abinitio drag model in Eq. (47) derived from detailed scale LB sim
ulations and (ii) the empirical drag model in Eq. (46). Note that for the latter
model, the literature values for the exponent n are extremely scattered (Morgan
et al., 1971). In Table IV, we show the results for n from different experiments
for Geldart A particles, which are clearly much higher than the value n 4.65,
originally obtained by Richardson and Zaki (1954). In this section, we show
results using the Wen and Yu expression with the commonly used value
n 4.65, and with the highest reported value n 9.6, from the experiments by
Lettieri et al. (2002).
For the simulations we use a 2D TFM as described in the previous sections.
The simulation conditions are specied in Table V. The gas ow enters at the
bottom through a porous distributor. The initial gas volume fraction in each
uid cell is set to an average value of 0.4 and with a random variation of 75%.
Also for the boundary condition at the bottom, we use a uniform gas velocity
with a superimposed random component (10%), following Goldschmidt et al.
(2004).
The simulations show that for low gas velocities (U
0
0.009 m/s), the com
monly used exponent n 4.65 does not yield a realistic bed expansion dynamics
for Geldart A particles. By using a large exponent (n 9.6), which was deter
mined by gas uidization of Geldart A particles, we can get a bed expansion
M.A. VAN DER HOEF ET AL. 128
around 31% of the initial bed height, which is much closer to the experimental
results (Geldart, 1973). Basically, a larger exponent n in Eq. (46) will lead to a
higher drag at the same gas velocity. It can thus be argued that at lowgas
velocities the drag force is underestimated by the commonly used drag models.
The question arises what the physical origin is of such large exponents. One
possibility is that they are caused by microstructures that form from smallscale
instabilities and perhaps other mechanisms. Also, the experiments by Lettieri et
al. (2002) showed a much larger apparent terminal velocity, which is indicative
of a much larger effective size. If such microstructures cannot be captured by the
CFD grid, then the use of a modied drag function, such that the experimental
bed expansion is obtained, is a possible way to go about. It should be stressed,
however, that this type of approach is rather ad hoc and not in the spirit of the
multiscale modeling strategy.
It has been reported by several researchers (Ferschneider and Mege, 1996;
Bayle et al., 2001; McKeen and Pugsley, 2003) that an overestimated bed ex
pansion was found at a highgas velocity ($0.2 m/s). We also carried out several
simulations for a high gas velocity, U
0
0.2 m/s. We still use the drag model
given by Eq. (46) with an exponent n 4.65. The simulation domain, however,
TABLE V
SIMULATION CONDITIONS
Parameters Value Parameter Value
Gas shear viscosity 1.8 10
5
Pa s CFD cells 30 45
Gas temperature 293 K Size of the cell 5 5 mm
2
Gas pressure 1.01 10
5
Pa Particle diameter 75 mm
Gas constant 8.314 J/(mol K) Particle density 1,500 kg/m
3
CFD time step 1.0 10
4
s Coefcient of restitution 0.97
TABLE IV
EXPONENT N FOR GELDART A PARTICLES
d
p
(mm) n
Lettieri et al. (2002), Newton and Gates (2002): Gasuidization
71 9.6
57 9.0
49 8.2
Massimilla et al. (1972): Gasuidization ()
60 7.12
53 6.86
45 6.1
Lewis and Bowerman (1952): Liquiduidization
8
6
8.3
Whitmore (1957): Liquidsedimentation
65 9.5
MULTISCALE MODELING OF GASFLUIDIZED BEDS 129
is enlarged so that a high bed expansion can be accommodated. The compu
tational domain is composed of 30 70 cells, and the size of each cell still
remains as 5 5 mm
2
. With such a high gas velocity the bed in fact is in the
turbulent uidization regime. In Fig. 26, we show the results obtained at differ
ent points in time when the bed reaches a dynamical equilibrium. Clearly, the
particle phase displays a turbulencelike ow pattern. Also, an overestimation
of bed height is found in the simulations, which is around 100% of the initial
bed height.
We also carried out a set of simulations using Eq. (47) as a drag model, which
was based on the data of LB simulations. The results are shown in Fig. 27. As
can be seen, no big differences can be observed compared to the results from the
drag model given by Eq. (46) with an exponent n 4.65.
A similar simulation was also carried out by McKeen and Pugsley (2003).
They also found an overestimation of the bed height, compared to their ex
perimental results. They argued that a factor should be used to scale down the
FIG. 26. The bed expansion dynamics of Geldart A particles from the TFM. The supercial gas
velocity U
0
is set to 0.2 m/s. The exponent n of the Wen and Yu equation is set to 4.65. No cohesion
is considered here. The results are, from the right to left, taken at t 9.6, 9.7, 9.8, 9.9, and 10.0 s.
M.A. VAN DER HOEF ET AL. 130
drag force in this regime in order to obtain a better agreement with the ex
periments. In Fig. 28, we show the results of our simulations with a drag force
(n 4.65) scaled down by a factor 0.15. A signicant decrease of the bed height
is found, with a bed expansion that is around 16% of the initial bed height, close
to the experimental observations (McKeen and Pugsley, 2003).
V. Towards IndustrialScale Models
In Section I, we mentioned that the TFM can simulate uidized beds at
engineering scales (height 12 m), and that the largescale industrial uidized
bed reactors (diameter 15 m, height 320 m) are still far beyond its capabilities.
Clearly, it would be highly desirable to predict the properties of gassolid ows
at the industrial scale; however at present, there is no fully evolved model
based on fundamental principleswhich is capable of this. In this section, we
outline some new ideas in this direction that have been developed both at the
FIG. 27. The same as in Fig. 26, but now using the LB drag model in Eq. (47) from Hill et al.
(2001b), with A
2
A
o
.
MULTISCALE MODELING OF GASFLUIDIZED BEDS 131
Princeton University and at the University of Twente. Before doing so; how
ever, it is rst important to understand why the current class of TFMs is not
suitable for describing largescale gassolid ows.
A. THE LIMITS OF THE TWOFLUID MODEL
Let us step back and examine the TFM and the closures we described thus far
in the chapter. Recall that the details of ow at the level of individual particles
are erased by the averaging process leading to the TFM equations, and that
their consequences appear in the averaged equations through terms which have
to be closed. The size of the averaging region was not explicitly considered
anywhere in the derivation of the TFM equations or the closures, and it was
implicitly assumed that a separation of scale existsnamely, the size of the
averaging region is much larger than the particle sizebut is much smaller than
the scale of the macroscopic ow structure that we wish to study by solving the
TFM equations. The assumption of such a separation of scales underlies the
very formulation of continuum models.
FIG. 28. The same as in Fig. 26, but with the drag force scaled by a factor of 0.15.
M.A. VAN DER HOEF ET AL. 132
Furthermore, the closures for the uidparticle drag and the particle
phase stresses that we discussed were all derived from data or analysis of
nearly homogeneous systems. In what follows, we refer to the TFM equa
tions with closures deduced from nearly homogeneous systems as the micro
scopic TFM equations. The kinetic theory based model equations fall in this
category.
We illustrated how these equations are discretized over an appropriate nu
merical grid and also showed some sample results. One can readily appreciate
that one must choose the grid sizes in the numerical solution of the TFM
equations to be smaller than the shortest length scale at which the TFM equa
tions afford inhomogeneities. This requirement leads to a practical difculty
when one tries to solve these microscopic TFM equations for gasparticle ows,
as discussed below.
Gasparticle ows in uidized beds and riser reactors are inherently unstable
and they manifest inhomogeneous structures over a wide range of length and
time scales. There is a substantial body of literature where researchers have
sought to capture these uctuations through numerical simulation of micro
scopic TFM equations, and it is now clear that TFMs for such ows do reveal
unstable modes whose length scale is as small as ten particle diameters (e.g., see
Agrawal et al., 2001; Andrews et al., 2005).
This is illustrated in Fig. 29. Transient simulations of a uidized suspension of
ambient air and typical uid catalytic cracking catalyst particles were performed
(using MFIX (Syamlal et al., 1993; Syamlal, 1998, which is an opendomain
code for solving multiphase ow problems) in a 2D periodic domain at different
grid resolutions. These simulations employed kinetic theorybased (microscopic)
TFM equations; see Agrawal et al. (2001) for a summary of the equations,
closures, and parameter values used in the simulations. Although there are some
slight differences between the closure expressions used by these authors and
those described (as illustrative examples) in this article, the differences are only
quantitative and not qualitative, so there is no need to present these closures
here. A pressure drop that is commensurate with the weight of the gasparticle
mixture in the periodic box was applied across the box in the vertical direction,
which provided the driving force for the upow of the uidizing gas. The sim
ulations revealed that an initially homogeneous suspension gave way to an
inhomogeneous state with persistent uctuations. Snapshots of the particle
volume fraction elds obtained in simulations with different number of spatial
grids are shown in Fig. 29.
It is readily apparent that ner and ner structures get resolved as the
number of spatial grids is increased. Statistical quantities, such as average slip
velocity between the gas and particle phases, obtained by averaging over the
whole domain, were found to depend on the grid resolution employed in the
simulations and they became nearly gridsize independent only when grid sizes
of the order of a few (E10) particle diameters were used. Thus, if one sets out
to solve microscopic TFM equations, grid sizes of the order of few particle
MULTISCALE MODELING OF GASFLUIDIZED BEDS 133
diameters are required; such ne spatial grids (and the fact that inhomoge
neous structures extend down to this ne scale) limit the time steps that can be
taken as well. For most devices of practical (commercial) interest, such ex
tremely ne spatial grids and small time steps are unaffordable (e.g., see Sun
daresan, 2000). Indeed, gasparticle ows in large uidized beds and risers are
often simulated by solving discretized versions of the TFM equations over
coarse spatial grids. Such coarsegrid simulations do not resolve the smallscale
(i.e., subgrid scale) spatial structures that, according to the microscopic TFM
equations, do indeed exist. The effect of these unresolved structures must be
brought to bear on the structures resolved in coarsegrid simulations through
appropriate modications to the closuresfor example, the effective drag
coefcient in the coarsegrid simulations will be smaller than that in the orig
inal TFM to reect the tendency of the gas to ow around the unresolved
clusters. Qualitatively, this is equivalent to an effectively larger apparent size
for the particles.
One can readily pursue this line of thought and examine the inuence of these
unresolved structures on the effective interphase transfer and dispersion coef
cients that should be used in coarsegrid simulations. Inhomogeneous distribu
tion of particles will promote by passing of the gas around the particlerich
regions and this will necessarily decrease the effective interphase mass and en
ergy transfer rates. Similarly, uctuations associated with the smallscale in
homogeneities will contribute to the rate of dispersion of the particles and the
gas, but they will be unaccounted for in the coarsegrid simulations of the
microscopic TFM equations.
FIG. 29. Snapshots of particle volume fraction elds obtained while solving a kinetic theorybased
TFM. 75 mm uid catalytic particles in ambient air. Simulations were done over a 16 32 cm pe
riodic domain. The average particle volume fraction in the domain is 0.05. Dark (light) color
indicates regions of high (low) particle volume fractions. (See Refs. Agrawal et al., 2001; Andrews et
al., 2005) for other parameter values.) Source: Andrews and Sundaresan (2005).
M.A. VAN DER HOEF ET AL. 134
B. STATEOFTHEART ON DEALING WITH UNRESOLVED STRUCTURES
Researchers have approached this problem of treating unresolved structures
through various approximate schemes. OBrien and Syamlal (1993) and Boemer
et al. (1994) pointed out the need to correct the drag coefcient to account for
the consequence of clustering and proposed a correction for the very dilute
limit. Some authors have used apparent cluster size in an effective dragcoef
cient closure as a tuning parameter; for example see McKeen and Pugsley
(2003), who attribute the larger apparent size to interparticle attractive forces,
and others have deduced corrections to the drag coefcient using energy min
imization multiscale approach (see Yang et al., 2004). The concept of particle
phase turbulence has also been explored to introduce the effect of the uctu
ations associated with clusters and streamers on the particlephase stresses
(Dasgupta et al., 1994; Hrenya and Sinclair, 1997). However, a systematic ap
proach that combines the inuence of the unresolved structures on the drag
coefcient and the stresses that can arise even when interparticle forces are not
important has not yet emerged.
One can summarize the multiscale character of TFM simulations using coarse
spatial grids as follows. When confronted with the task of performing simu
lation of gasparticle ows in large process vessels, one faces constraints on
affordable grid resolution; this can lead to unresolved subgrid structures that
would have been obtained if only the TFM equations were solved on a ne
spatial grid. The consequence of these subgrid structures on the ow pattern
resolved by the coarsegrid simulations should be brought in through appro
priate corrections to the closure relations. If one simply uses the closures in the
microscopic TFM without adding the corrections, then there is no guarantee
that the obtained solution is a true solution for the TFM equations that one sets
out to solve.
This is well known in other contexts, such as singlephase turbulence. Large
eddy simulations introduce corrections to the uidphase stress through subgrid
models; for example, Smagorinsky, in his pioneering work (Smagorinsky, 1963),
introduced a model for subgrid viscosity correction.
Agrawal et al. (2001) pointed out that, in gasparticle ows such as those
encountered in uidized beds and riser ows, one should include subgrid cor
rections for not only the effective particle and uidphase stresses but also the
effective drag. They showed that the effective drag law and the effective stresses
obtained by averaging (the results gathered in highly resolved simulations of a
set of microscopic TFM equations, such as that corresponding to the most
resolved snapshot in Fig. 29 over the whole (periodic) domain were very differ
ent from those used in the microscopic TFM and that they depended on size of
the domain over which simulations were carried out (Agrawal et al., 2001). They
also found that all the effects seen in the 2D simulations persisted when sim
ulations were repeated in three dimensions (3D) and that both 2D and 3D
simulations revealed the same qualitative trends.
MULTISCALE MODELING OF GASFLUIDIZED BEDS 135
Andrews et al. (2005) performed many highly resolved simulations of uid
ized gasparticle mixtures in a 2D periodic domain, whose size coincided with
that of the grid size in an anticipated largescale riser ow simulation. Through
such highly resolved simulations, they constructed ad hoc subgrid models for the
effects of the nescale ow structures on the drag force and the stresses, and
examined the consequence of these subgrid models on the outcome of the
coarsegrid simulations of gasparticle ow in a largescale vertical riser. They
have demonstrated that these subgrid scale corrections can affect the predicted
ow patterns profoundly.
Thus, there is no doubt that one must carefully examine whether the micro
scopic TFM equations must be modied to introduce the effects of unresolved
structures before embarking on coarsegrid simulations of gasparticle ows in
chemical reactors. At the same time, the ad hoc method employed by Andrews et
al. (2005), namely performing highly resolved simulations in periodic domains
whose linear dimensions are the same as those of the grids, is not a rigorous
approach to take either; for example, one can anticipate that the periodic
boundary conditions imposed in such highly resolved simulations would place
some restrictions (on the smallscale ow structure) that would be absent in the
real, largescale ow. Thus, alternate approaches to constructing closures suit
able for coarsegrid computations must be developed.
Adopting the approach pursued in large eddy simulations, one can start with
the TFM equations and perform a ltering operation, where the averaging is
done over a lter length scale that is somewhat larger than the grid size to be
used in the coarsegrid simulation of largescale process vessels and over high
(temporal) frequencies. The mathematical steps involved in ltering any version
of the microscopic TFM are conceptually straightforward (e.g., see Zhang and
VanderHeyden, 2002) and will not be presented here. We simply note that the
dominant terms in the ltered equations can be recast in exactly the same form
as the original TFM equations; however, effective stresses, interphase interac
tion force term, etc. will now involve additional contributions resulting from the
ltering process. (It is because of this similarity that one can use the same
platform such as MFIX to perform integration of the ltered equations as well.)
Insight into these closures for the additional contributions resulting from the
ltering process can be gained through analysis of computational data gathered
through highly resolved simulations in sufciently large domains, while ensuring
that the overall ow domain simulated is considerably larger than the region
over which the ltering operation is performed. This is illustrated below by
some results obtained by Andrews and Sundaresan (2005).
Consider a highly resolved simulation of a set of microscopic TFM equations
for a uidized suspension of particles in a large periodic domain. The ltering
operation does not require a periodic domain; however, as each location in a
periodic domain is statistically equivalent to any other location, statistical av
erages can be gathered much faster when simulations are done in periodic do
mains. After an initial transient period that depends on the initial conditions,
M.A. VAN DER HOEF ET AL. 136
persistent, timedependent, and spatially inhomogeneous structures develop.
Fig. 30 shows an instantaneous snapshot of the particle volume fraction eld in
one such 2D simulation (performed using MFIX) and the cells (i.e., ne grids)
used in the simulations. One can then zoom in any region of desired size and
average any quantity of interest over all the cells inside that region, and obtain
regionaveraged (ltered) values. Note that one can choose a large number of
regions inside the overall domain and thus several regionaveraged values can be
constructed for any quantity of interest from each instantaneous snapshot.
When the system is in a statistical steady state, one can construct tens of thou
sands of such averages by repeating the analysis at various time instants.
Returning to Fig. 30, note that the averages over different regions at any
given time are not equivalent; for example, at the given instant, different regions
(of the same size) will correspond to different regionaveraged particle volume
fractions, particle and uid velocities, and so on. Thus, one cannot simply lump
the results obtained over all the regions; instead, the results must be grouped
FIG. 30. Snapshot of particle volume fraction elds obtained while solving a kinetic theorybased
TFM. Fluid catalytic particles in air. Simulations were done over a 16 16 cm periodic domain.
128 128 cells (shown in the gure). The average particle volume fraction in the domain is 0.05.
Dark (light) color indicates regions of high (low) particle volume fractions. Squares of different sizes
illustrate regions (i.e., lters) of different sizes over which averaging over the cells is performed.
Source: Andrews and Sundaresan (2005).
MULTISCALE MODELING OF GASFLUIDIZED BEDS 137
into bins based on various markers and perform statistical averages within each
bin to get useful information. The 2D simulations of Andrews and Sundaresan
(2005) revealed that the single most important marker for regions is the average
particle volume fraction in that region. Therefore, in order to expose the effects
of particle volume fraction on the ltered (i.e., regionaveraged) quantities, they
classied the regionaveraged data into bins of particle volume fraction and
evaluated the ltered slip velocity, uidparticle interaction force, etc., and
averaged each of these quantities within each bin. From such bin statistics, they
calculated the ltered drag coefcient, ltered particlephase pressure, and
ltered particlephase viscosity as functions of ltered particle volume fraction.
Fig. 31 shows the variation of the ltered drag coefcient as a function of the
ltered (i.e., regionaverage) particle volume fraction for various lter sizes.
5
Each point represents the average of many realizations in a bin. (Here, the
ltered drag coefcient is dened as the regionaverage drag force divided by the
regionaverage slip velocity.) It is clear from Fig. 31 that the ltered drag
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
0
10
20
30
40
50
60
70
80
90
s
bin
e
f
f
e
c
t
i
v
e
d
r
a
g
c
o
e
f
f
i
c
i
e
n
t
1x1 cells = 0.125cm
2x2 cells = 0.250cm
4x4 cells = 0.50cm
8x8 cells = 1.0cm
16x16 cells = 2.0cm
FIG. 31. Filtered drag coefcient (in CGS units) extracted from simulations over 16 16 cm
domain using 128 128 cells. Source: Andrews and Sundaresan (2005).
5
Strictly speaking, one should use 2D bins involving particle volume fraction and a Reynolds
number based on slip velocity to classify the ltered drag coefcient; however in these simulations,
the Reynolds number effect was found to be weak and hence the data were collapsed to just volume
fraction bins.
M.A. VAN DER HOEF ET AL. 138
coefcient depends on the size of the lter used in the analysis. This gure
includes results obtained from three different simulations corresponding to
three different average particle volume fractions in the domain (0.05, 0.15, and
0.40). The larger the lter size the smaller is the drag coefcient , the reason
being that the averaging (i.e., ltering) is being performed over larger and larger
clustersthe larger the clusters, the greater is the bypassing of the gas around
the clusters and hence lower is the apparent drag coefcient.
Fig. 32 shows the variation of ltered particlephase pressure as a function of
the ltered particle volume fraction for various lter sizes. Here the ltered
particlephase pressure includes the pressure arising from the streaming and
collisional parts captured by the kinetic theory and the sublterscale Reyno
ldsstress like velocity uctuations (see Agrawal et al., 2001 for further details).
Indeed, the contributions resulting from the sublterscale velocity uctuations
swamp the kinetic theory pressure, indicating that at the coarsegrid scale one
can even ignore the kinetic theory contributions to the pressure! This gure
clearly shows that the ltered particlephase pressure increases with lter size,
and this is a direct consequence of the fact that the energy associated with the
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
0
20
40
60
80
100
120
140
160
180
200
s
bin
p
s
1x1 cells = 0.125cm
2x2 cells = 0.250cm
4x4 cells = 0.50cm
8x8 cells = 1.0cm
16x16 cells = 2.0cm
FIG. 32. Filtered particle phase pressure (in CGS units) extracted from simulations over 16
16 cm domain using 128 128 cells. Source: Andrews and Sundaresan (2005). The ltered particle
phase pressure includes the Reynolds stresslike uctuations and the kinetic theory pressure.
MULTISCALE MODELING OF GASFLUIDIZED BEDS 139
velocity uctuations increases with lter length (analogous to what one has in
singlephase turbulence).
The trends presented in Figs. 31 and 32 qualitatively similar to those pre
sented earlier by Agrawal et al. (2001) and Andrews et al. (2005) who, for the
sake of simplicity, did simulations on much smaller domains and let the lter
size be the same as the domain size. This shows clearly that the effects leading to
the type of results presented in Figs. 31 and 32 are robust.
Andrews and Sundaresan (2005) have also extracted the ltered particle
phase viscosity from these simulations and found that at low particle volume
fractions (0.00.25), the ltered viscosity varies nearly linearly with particle
volume, and that it increases monotonically (and nearly linearly) with lter size.
A nal piece of the proofofconcept calculations is to compare the predic
tions obtained by solving the ltered TFM equations with highly resolved sim
ulations of the microscopic TFM equations. For this purpose, Andrews and
Sundaresan (2005) performed simulations of the microscopic TFM equations in
a 16 32 cm periodic domain at various resolutions (e.g., see Fig. 29). From
these simulations, they extracted domainaverage quantities in the statistical
steady state (see Agrawal et al., 2001 for a discussion of how these data are
gathered). Fig. 33 shows the domainaverage slip velocity between the gas and
particle phases at various grid resolutions (shown by the squares connected by
0 10 20 30 40 50 60 70 80 90
0
20
40
60
80
100
grids in lateral direction
t
i
m
e
a
n
d
d
o
m
a
i
n
a
v
e
r
a
g
e
d
s
l
i
p
v
e
l
o
c
i
t
y
kinetic theory model
filter size = 2.0cm
FIG. 33. Comparison of the domainaverage slip velocity (in cm/s) determined by solving a mi
croscopic TFM and the corresponding ltered TFM. 16 32 cm periodic domain. Domainaverage
particle volume fraction 0.05. Number of grids in the vertical direction is twice that in the lateral
direction. Source: Andrews and Sundaresan (2005).
M.A. VAN DER HOEF ET AL. 140
the bold solid line in this gure). After sufcient grid resolution, this quantity
clearly levels off, indicating convergence in a statistical sense. They also per
formed computations with the ltered TFM, using the computationally gen
erated closures (e.g., drag and particlephase pressure closures shown in Figs. 31
and 32, and particlephase viscosity closure, not shown) for a 2 2 cm lter.
The domainaverage slip velocity obtained by solving the ltered equations at
different grid resolutions are shown in Fig. 33 as triangles (connected by the thin
solid line).
Fig. 33 reveals two important features. Firstly, at coarse resolutions, the
domainaverage slip velocity obtained by solving the microscopic TFM changes
appreciably with grid resolution; in contrast, the gridsize dependence of the slip
velocity computed by solving the ltered TFM is much weaker. Secondly, at
sufciently highgrid resolution, both approaches yield comparable predictions,
and this is an important rst step in validating the ltered TFM approach.
Another result that is not evident in Fig. 33 concerns the computational times
required for gathering the statistical steadystate values of various quantities
(such as the slip velocity shown in Fig. 33); at comparable grid resolutions, the
computational time required to solve the ltered equations is much smaller than
that for the microscopic equations. This can be attributed to the fact that the
structures obtained in the solution of the ltered equations are comparatively
coarser than those for the microscopic TFM equations.
C. A DIFFERENT APPROACH: THE DISCRETE BUBBLE MODEL
An alternative scheme to tackle the problem of largescale ow structures is
being pursued at Twente University. In this model the bubbles, as observed in
the DPM and TFM models of gasuidized beds, are considered as discrete
entities. This is the socalled discrete bubble model, which has been successfully
applied in the eld of gasliquid bubble columns (Delnoij et al., 1997). The idea
to apply this model to describe the largescale solids circulation that prevail in
gassolid reactors is new, however, and involves some slight modications of
the equivalent model for gasliquid systems (Bokkers et al., 2005a). To this end,
the emulsion phase is modeled as a continuumlike the liquid in a gasliquid
bubble columnand the larger bubbles are treated as discrete bubbles. Note
that granular systems have no surface tension, so in that respect there is a
pronounced difference with the bubbles present in gasliquid bubble columns.
For instance, the gas will be free to ow through a bubble in gassolid systems,
which is not the case for gasliquid systems. As far as the numerical part is
concerned, the DBM strongly resembles the DPM as outlined in Section III,
since it is also of the EulerLagrange type with the emulsion phase described by
the volumeaverage NavierStokes equations:
@
e
r
e
@t
r
e
r
e
u
e
0 (97)
MULTISCALE MODELING OF GASFLUIDIZED BEDS 141
with r
e
, e
e
, and u
e
the density, volume fraction, and ow velocity, respectively,
of the emulsion phase. Momentum conservation gives that
@
e
r
e
u
e
@t
r
e
r
e
u
e
u
e
e
rp S
E
r s
e
e
r
e
g (98)
where the symbols take their usual meaning, and the subscript e indicates
emulsion phase. The term S
E
accounts for the twoway coupling between the
dispersed phase and the continuous phase. The bubbles are considered as dis
crete elements that are tracked individually according to Newtons second law
of motion:
m
b
dv
b
dt
F
tot
(99)
where F
tot
is the sum of different forces acting on a single bubble:
F
tot
F
g
F
p
F
d
F
W
F
VM
(100)
As in the DPM model, the total force on the bubble has contributions from
gravity (F
g
), pressure gradients (F
p
), and drag from the interaction with emulsion
phase (F
d
). The sum of F
g
and F
p
is equal to (p
e
p) V
b
g, with V
b
the volume of
the bubble. For the drag force on a single bubble (diameter d
b
), the correlations
for the drag force on a single sphere are used, only with a modied drag co
efcient C
d
, such that it yields the relation v
br
0:711
gd
b
_
by Davies and Tay
lor (1950) for the rise velocity of a single bubble. Note that in Eq. (100), there are
two forces present that are not included in the DPM, namely the wake force F
W
and the virtual mass force F
VM
. The wake force, accounting for the acceleration
of a bubble in the wake of a leading bubble, is neglected in this application;
whereas for the virtual mass force, the relation by Auton (1983) is used:
F
VM
DI
Dt
I ru
_ _
; I 0:5r
e
V
b
v
b
u
e
~
S
2
q
(5)
with
~
S
2
1
2
r~ v r~ v
T
: r~ v r~ v
T
(6)
While the theoretical value (based on homogeneous, isotropic turbulence) of the
Smagorinsky coefcient c
s
amounts to 0.165 (Mason and Callen, 1986), in many
simulation studies lower values for c
s
proved to result in a better reproduction
of experimental data. This may have to do with the abundant presence of shear
ows in process equipment. Derksen (2003) reported that varying c
s
values in
the range 0.080.14 does not have a large impact on the simulation results. A
value of 0.12 is recommended.
At the basis of the Smagorinsky SGS model is the assumption of equilibrium
between production and dissipation of turbulent kinetic energy in the inertial
subrange of eddy sizes. In stirred tanks, however, there is hardly any position
where this equilibrium prevails. Furthermore, there is not a good reason why
the Smagorinsky coefcient should be constant across the ow domain. Other
more specic SGS models have therefore been proposed and also investigated
as to their impact on the resulting ow elds and turbulence characteristics.
Hartmann et al. (2004a) assessed the usability of an SGS model due to Voke
(more geared to lowReynolds number turbulence), while Derksen (2001) in
vestigated a socalled structure function SGS model for a turbulent viscosity that
depends on eddy size. Derksen (2006a) supplemented the standard Smagorinsky
SGS model with walldamping functions to bring the eddy viscosity explicitly to
zero at solid walls, since in physical reality velocity uctuations and subgrid
stresses are zero at walls. Recently, FLUENT 6.2 came with several new SGS
models. Further renements in SGS modeling may be expected to improve the
accuracy of LES.
An inherent property of the LES approach is that the simulated ow eld is
no longer steady, but exhibits a transient character due to the presence and
motion of largescale eddies. The LES methodology has proven to be a powerful
tool for studying and visualizing stirred tank ows (Eggels, 1996; Derksen et al.
1999; Bakker et al., 2000; Derksen, 2001; Bakker and Oshinowo, 2004), as it
inherently takes the unsteady and periodic behavior of the ow (around impeller
and bafes) into account.
HARRY E. A. VAN DEN AKKER 162
C. REYNOLDS AVERAGED NAVIER STOKES SIMULATIONS
The focus of RANS simulations is on the timeaveraged ow behavior
of turbulent ows. Yet, all turbulent eddies do contribute to redistributing
momentum within the ow domain and by doing so make up the inherently
transient character of a turbulentow eld. In RANS, these effects of the full
range of eddies are made visible via the socalled Reynolds decomposition of
the NS equations (see, e.g., Tennekes and Lumley, 1972, or Rodi, 1984) of the
ow variables into mean and uctuating components. To this end, a clear dis
tinction is required between the temporal and spatial scales of the mean ow
on the one hand and those associated with the turbulent uctuations on the
other hand.
Via this Reynolds decomposition and after subsequent averaging all terms of
the NS equations, the socalled turbulent or Reynolds stresses u
i
u
j
emerge in
the transport equations, where these stresses represent the additional averaged
momentum transport due to the eddies. These stresses may be resolved explicitly
from separate transport equations which in sufx notation (usual in the eld of
turbulence) look as follows:
@u
i
u
j
@t
u
k
@u
i
u
j
@x
k
P
ij
D
ij
ij
P
ij
(7)
in which the rst three terms of the rightterm side denote the production,
diffusion, and dissipation of the turbulent stresses, respectively, while the last
term is the socalled pressurestrain term that represents the redistribution of the
turbulent kinetic energy among the three coordinate directions that makes the
turbulence more isotropic (Rodi, 1984). Several of these terms need modeling
for which a gamut of choices is available. In principle, this approach implies
the need of solving nine more partial differential equations per grid cell. As a
result, CPU times required for computational simulations on the basis of some
Reynolds Stress Model (RSM) are relatively high.
A cure against these longer CPU times is the Algebraic Stress Model (ASM)
described by, e.g., Rodi (1984) and used and recommended by, e.g., Bakker
(1992) and Bakker (1996). Most commercial codes do no longer support an ASM.
Usually, however, the stresses are modeled with the help of a single turbulent
viscosity coefcient that presumes isotropic turbulent transport. In the RANS
approach, a turbulent or eddy viscosity coefcient, n
t
, covers the momentum
transport by the full spectrum of turbulent scales (eddies). Frisch (1995) recol
lects that as early as 1870 Boussinesq stressed turbulence greatly increases vis
cosity and proposed an expression for the eddy viscosity. The eventual set of
equations runs as
@V
@t
V rV
1
r
rP n n
t
r
2
V (8)
THE DETAILS OF TURBULENT MIXING PROCESS 163
In its turn, the turbulent viscosity may be position dependent and generally
may be modeled in terms of a model, very usually a ke model:
n
t
C
m
k
2
(9)
where k is the concentration of turbulent kinetic energy in J/kg (or m
2
/s
2
) and e
is the rate of dissipation (in W/kg, or m
2
/s
2
/s) of this turbulent kinetic energy.
These two concentrations k and e are generally conceived as the most important
parameters describing a turbulentow eld. In their turn, their spatial distri
butions within the turbulentow domain may be calculated from the following
transport equations for k and e, respectively:
@k
@t
u
i
@k
@x
i
P
k
D
k
P
k
(10)
the righthand side of which contains similar terms as the above transport
equations for the turbulent stresses, and
@
@t
u
i
@
@x
i
P
O (11)
in which the last term denotes the destruction of e. The interested reader is
referred to, e.g., Rodi (1984) for the meaning of all these righthand terms and
their modeling. The assumption often used in classical turbulence theory that
production and dissipation of turbulent kinetic energy balance locally, is found
by putting in Eq. (10) all terms but the rst and third at the righthand side
equal to zero.
These two transport equations for k and e form an inherent part of any ke
model of RANSsimulations. As the result of closing the turbulence modeling
such that no further unknown variables and equations are introduced, the
eequation does contain some terms that are still the result of modeling, albeit at
the very small scales (e.g., Rodi, 1984).
The (isotropic) eddy viscosity concept and the use of a ke model are known
to be inappropriate in rotating and/or strongly 3D ows (see, e.g., Wilcox,
1993). This issue will be addressed in more detail in Section IV. Some researchers
prefer different models for the eddy viscosity, such as the ko model (where o
denotes vorticity) that performs better in regions closer to walls. For this latter
reason, the ke model and the ko model are often blended into the socalled
ShearStressTransport (SST) model (Menter, 1994) with the view of using these
two models in those regions of the ow domain where they perform best. In spite
of these objections, however, RANS simulations mostly exploit the eddy vis
cosity concept rather than the more delicate and timeconsuming RSM turbu
lence model. They deliver simulation results ofin many casesreasonable or
sufcient accuracy in a costeffective way.
HARRY E. A. VAN DEN AKKER 164
RANSbased simulations exploiting the eddy viscosity concept just reproduce
the average ow eld and the spatial distribution of turbulence properties such
as k and e. As such, RANSbased simulations are excellently suited for iden
tifying dead zones, recirculatory ows, shortcircuiting between entrance
and exit, and further undesired ow features. Even in transient RANSbased
simulations, however, it is not a priori clear which part of the uctuations
is temporally resolved and which part is taken care of by the turbulence
model. This inherent property of RANSbased simulations especially raises
concerns in the case of ows exhibiting no clear spectral separation between
lowfrequency coherent motions (such as macroinstabilities, precessing vortices,
and trailing vortices) and turbulent uctuations (making part of the cascade of
eddies or whirls typical of turbulence); and: A mechanistic picture of turbu
lence cannot be treated on the average since such ows are dynamic. (Praturi
and Brodkey, 1978).
Yet (steady) RANSbased simulations are attractive as they relatively cheaply
deliver a quick impression of the overall ow eld in the vessel. Effects on the
overall ow eld of varying the position of impeller, feed pipe, withdrawal pipe,
and/or heat coil can easily be explored.
Note that the Eqs. (1), (2), and (8) are really and essentially different due to
the absence or presence of different turbulent transport terms. Only by incor
porating dedicated formulations for the SGS eddy viscosity can one attain that
LES yield the same ow eld as DNS. RANSbased simulations with their
turbulent viscosity coefcient, however, essentially deliver steady ow elds
and as such are never capable of delivering the same velocity elds as the
inherently transient LES or DNS, irrespectively of the renement of the com
putational grid!
D. THE SIMULATION OF PROCESSES IN A TURBULENT SINGLEPHASE FLOW
For simulating computationally the spatial and temporal evolution of both
physical and chemical processes in mixing devices operated in a turbulent single
phase mode, two essentially different approaches are available: the Lagrangian
approach and the Eulerian technique. These will be explained briey.
In the Lagrangian approach, individual parcels or blobs of (miscible) uid
added via some feed pipe or otherwise are tracked, while they may exhibit
properties (density, viscosity, concentrations, color, temperature, but also vorti
city) that distinguish them from the ambient uid. Their path through the
turbulentow eld in response to the local advection and further local forces (if
applicable) is calculated by means of Newtons law, usually under the assump
tion of oneway coupling that these parcels do not affect the ow eld. On
their way through the tank, these parcels or blobs may mix or exchange mass
and/or temperature with the ambient uid or may adapt shape or internal
velocity distributions in response to events in the surrounding uid.
THE DETAILS OF TURBULENT MIXING PROCESS 165
In real life, the parcels or blobs are also subjected to the turbulent uctuations
not resolved in the simulation. Depending on the type of simulation (DNS,
LES, or RANS), the wide range of eddies of the turbulentuidow eld is not
necessarily calculated completely. Parcels released in a LES ow eld feel both
the resolved part of the uid motion and the unresolved SGS part that, at
best, is known in statistical terms only. It is desirable that the forces exerted by
the uid ow on the particles are dominated by the known, resolved part of
the ow eld. This issue is discussed in greater detail in the next section in the
context of tracking real particles. With a RANS simulation, the turbulent
velocity uctuations remaining unresolved completely, the effect of the turbu
lence on the tracks is to be mimicked by some stochastic model. As a result,
particle tracking in a RANS context produces less realistic results than in an
LESbased ow eld.
An early example of tracking uid parcels in a stirred tank can be found in
Bouwmans (1992) and Bouwmans et al. (1997). Bakker (1996) used a tracking
routine with the view to provide a Lagrangian description of micromixing in a
stirred tank chemical reactor. Lapin et al. (2004) recently described a compu
tational strategy for travelling along the lifelines of single cells (i.e., tracking
them) in stirred bioreactors in order to characterize the dynamics of the hetero
geneous cell population and to study the impact of spatial and dynamic varia
tions in concentrations of substrates and products across the reactor.
The motions of the individual uid parcels may be overlooked in favor of a
more global, or Eulerian, description. In the case of singlephase systems, con
vective transport equations for scalar quantities are widely used for calculating
the spatial distributions in species concentrations and/or temperature. Chemical
reactions may be taken into account in these scalar transport equations by
means of source or sink terms comprising chemical rate expressions. The per
tinent transport equations run as
@T
@t
v rT kr
2
T q (12)
and
@c
@t
v rc Dr
2
c r (13)
In Eq. (13), r stands for the production (or consumption) of the species of
interest due to a chemical reaction, while in Eq. (12) q represents the heat
production, e.g., due to one of more chemical reactions. Equation (13) is often
referred to as the ConvectionDiffusionReaction (CDR) equation.
Since turbulent uctuations not only occur in the velocity (and pressure) eld
but also in species concentrations and temperature, the convection diffusion
equations for heat and species transport under turbulentow conditions also
comprise crosscorrelation terms, obtained by properly averaging products of
HARRY E. A. VAN DEN AKKER 166
velocityconcentration, velocitytemperature, concentrationtemperature, and
concentrationconcentration uctuations, on the analogy of the Reynolds
stresses in the NS equations (e.g., Patterson, 1985; Ranade, 2002). The challenge
is still to nd appropriate closure relations for these crosscorrelation terms:
these may be either phenomenological or mechanistic (micromixing models) or
probabilistic (exploiting probability density functions, PDFs).
In stirred chemical reactors, unlike in combustion and with other gasphase
reactions, these closure terms should take into account that for liquids the
Schmidt number (Sc n=D) is in the order 1001,000, and, hence, the role of
species diffusion at scales within the Kolmogorov eddies should explicitly be
taken into account (Kresta and Brodkey, 2004). Essential is that diffusion of
chemical species is governed by the Batchelor length scale Z
B
which obeys to
Z
B
Z
K
Sc
1=2
(14)
which for large Sc numbers is much smaller than the Kolmogorov length
scale Z
K
indeed. Such socalled micromixing processes have to be described by
means of micromixing models which will be dealt with in some greater detail in
Section VIII.
These convective transport equations for heat and species have a similar
structure as the NS equations and therefore can easily be solved by the same
solver simultaneously with the velocity eld. As a matter of fact, they are much
simpler to solve than the NS equations since they are linear and do not involve
the solution of a pressure term via the continuity equation. In addition, the
usual assumption is that spatial or temporal variations in species concentration
and temperature do not affect the turbulentow eld (another example of one
way coupling).
E. THE COMPUTATIONAL FLUID DYNAMICS OF TWOPHASE FLOWS
On the analogy of simulating the process of adding blobs of a miscible
liquid, twophase ow in stirred tanks in a RANS context may be treated in
two ways: EulerLagrange or EulerEuler, with the second, dispersed phase
treated according to a Lagrangian approach and from a Eulerian point of view,
respectively.
1. Euler Lagrangian Approach
The EulerLagrangian approach is very common in the eld of dilute dis
persed twophase ow. Already in the mid 1980s, a particle tracking routine
was available in the commercial CFDcode FLUENT. In the EulerLagrangian
approach, the dispersed phase is conceived as a collection of individual particles
(solid particles, droplets, bubbles) for which the equations of motion can be
solved individually. The particles are conceived as point particles which move
THE DETAILS OF TURBULENT MIXING PROCESS 167
across the ow domain in response to the turbulentow eld of the carrier
phase. The consequence of treating particles as point particles is that the detailed
ow between the particles in response to the presence and motion of the par
ticles is not resolved.
For the hydrodynamics forces acting on the particles, mostly singleparticle
expressions are used; this implies that hydrodynamic interactions between par
ticles are ignored completely. In many cases, direct interactions of particles
owing to collisions are ignored as well. All thisalong with the computational
burden that increases linearly with the number of particles being trackedmay
limit the practical applicability of the method to dilute systems with relatively
low volume fractions of dispersed phase and/or to ow domains of small size.
By feeding back the reactive forces exerted by the particles on the continuous
carrier phase, twoway coupling between the two phases is obtained.
Without much discussion, one may anticipate that particle inertia, gravity,
and drag force need to be part of the equations of motion describing the motion
and paths of the particles. Since a stirred tank is very inhomogeneous and
exhibits strong gradients in velocity, pressure, and stress elds, it is difcult to
estimate a priori if more exotic uidparticle forces such as the Saffman lift
force, the Magnus force, and the history force may play a role of signicance
either globally or locally. For a concise summary about these forces and for
expressions for these forces, the reader is referred to, e.g., Derksen (2003).
Added mass may be important for bubbly ows. It is obvious that in such a
Lagrangian approach distributions in particle size may easily be taken into
account.
As discussed earlier in the context of tracking miscible parcels or blobs,
particles travel through the resolved average or uctuating velocity eld as well
as feel the unresolved velocity uctuations. Since the major uidparticle force
may be the drag force, the uids velocity eld is of primary importance, the
turbulent velocity uctuations inclusively, whether or not they are resolved in
the simulation; uctuations in pressure and stresses may be secondary. Supply
ing stochastic variations on top of the resolved velocity eld mimics the
unresolved uctuations and brings the expected seemingly erratic paths of the
particles about.
Of course, the role of the articially introduced stochastics for mimicking the
effect of all eddies in a RANSbased particle tracking is much more pronounced
than that for mimicking the effect of just the SGS eddies in a LESbased
tracking procedure. In addition, the random variations may suffer from lacking
the spatial or temporal correlations the turbulent uctuations exhibit in real life.
In RANSbased simulations, these correlations are not contained in the steady
spatial distributions of k and e and (if applicable) the Reynolds stresses from
which a typical turbulent time scale such as k/e may be derived. One may
try and cure the problem of missing the temporal coherence in the velocity
uctuations by picking a new random value for the uids velocity only after a
certain period of time has lapsed.
HARRY E. A. VAN DEN AKKER 168
In LESbased simulations, just the SGS part of the turbulence spectrum needs
to be mimicked by stochastics. The idea is that the resolved eddies have the
biggest impact on the paths of the particles indeed. This requires not only that
the resolved velocity uctuations should be stronger than the (estimated) SGS
uctuations, but also that the particle relaxation time should be larger than
the timestep applied in the LES. Meeting the latter criterion implies that the
time step of the LES is capable of keeping up with the time scale the particle
needs to respond to changes in the ow eld of the surrounding uid. In the
context of LES, picking a new random velocity only after (part of) a time k
sgs
/e
has lapsed may cure the problem of missing the coherence in the SGS eddies
when mimicking the effect of the SGS eddies on the particle tracks. Here, k
sgs
stands for the kinetic energy associated with the SGS eddies only and has to be
estimated. All these issues have been extensively discussed by, e.g., Derksen
(2003, 2006a).
2. Euler Euler Approach
In the complete Eulerian description of multiphase ows, the dispersed phase
may well be conceived as a second continuous phase that interpenetrates the
real continuous phase, the carrier phase; this approach is often referred to as
twouid formulation. The resulting simultaneous presence of two continua is
taken into account by their respective volume fractions. All other variables such
as velocities need to be averaged, in some way, in proportion to their presence;
various techniques have been proposed to that purpose leading, however, to
different formulations of the continuum equations. The method of ensemble
averaging (based on a statistical average of individual realizations) is now gene
rally accepted as most appropriate.
In the twouid formulation, the motion or velocity eld of each of the two
continuous phases is described by its own momentum balances or NS equations
(see, e.g., Rietema and Van den Akker, 1983 or Van den Akker, 1986). In both
momentum balances, a phase interaction force between the two continuous
phases occurs predominantly, of course with opposite sign. Twouid models
therefore belong to the class of twoway coupling approaches. The continuum
formulation of the phase interaction force should reect the same effects as
experienced by the individual particles and discussed above in the context of the
Lagrangian description of dispersed twophase ow.
One therefore has to decide here which components of the phase inter
action force (drag, virtual mass, Saffman lift, Magnus, history, stress gradients)
are relevant and should be incorporated in the two sets of NS equations.
The reader is referred to more specic literature, such as Oey et al. (2003), for
reports on the effects of ignoring certain components of the interaction force
in the twouid approach. The question how to model in the twouid formu
lation (lateral) dispersion of bubbles, drops, and particles in swarms is relevant
THE DETAILS OF TURBULENT MIXING PROCESS 169
as well: various models are available. See also the discussion on page 204 as
to (Eq. (19)).
Another important issue in twouid models is about modeling the turbulent
stresses under twophase conditions (e.g., Van den Akker, 1998). At this mo
ment, there is still no consensus on a universal twophase turbulence model.
Generally, turbulence in the continuous phase may be generated by shear due to
largescale velocity gradients felt by the continuous phase itself (just like in
singlephase ows) as well as by the presence and relative motion of the dis
persed phase particles. The ratio at which these two mechanisms contribute to
the generation of turbulence may be an important factor in drafting a universal
model. In addition, the dispersed phase may exhibit a turbulentow behavior in
response to the turbulent motions of the continuous phase in which it is em
bedded; this response may depend on several time scales and their interaction
(Oey et al., 2003). Lance et al. (1991) suggested that the motion of bubbles
promotes a return to isotropy (see also Van den Akker, 1998). A universal
model, however, is not available right now.
In dense systems such as encountered in solids suspension, particleparticle
interaction may be important as well. Then, the closure of solidphase stresses is
an important issue for which kinetic theory models and solids phase viscosity
may be instrumental (see, e.g., Curtis and Van Wachem, 2004).
As a matter of fact, in comparison with the EulerLagrangian approach, the
complete Eulerian (or EulerEuler) approach may better comply with denser
twophase ows, i.e., with higher volume fractions of the dispersed phase, when
tracking individual particles is no longer doable in view of the computational
times involved and the computer memory required, and when the physical in
teractions become too dominating to be ignored. Under these circumstances, the
motion of individual particles may be overlooked and it is wiser to opt for a more
supercial strategy that, however, still has to take the proper physics into account.
Precisely owing to the continuum description of the dispersed phase, in
EulerEuler models, particle size is not an issue in relation to selecting grid cell
size. Particle size only occurs in the constitutive relations used for modeling the
phase interaction force and the dispersedphase turbulent stresses.
In the case of droplets and bubbles, particle size and number density may
respond to variations in shear or energy dissipation rate. Such variations are
abundantly present in turbulentstirred vessels. In fact, the explicit role of the
revolving impeller is to produce small bubbles or drops, while in substantial
parts of the vessel bubble or drop size may increase again due to locally lower
turbulence levels. Particle size distributions and their spatial variations are
therefore commonplace and unavoidable in industrial mixing equipment. This
seriously limits the applicability of common EulerEuler models exploiting just
a single value for particle size. A way out is to adopt a multiuid or multiphase
approach in which various particle size classes are distinguished, with mutual
transition paths due to particle breakup and coalescence. Such models will be
discussed further on.
HARRY E. A. VAN DEN AKKER 170
III. Computational Aspects
As the continuity equation, the NS equations, and the transport equations for
the turbulent variables are highly nonlinear, any CFDcalculation is essentially
iterative. Generally, the convergence rate of simulations depends on the number
of grid points and on the number of equations to be solved.
The number of grid points is associated with the desired or required degree of
detail and accuracy of the simulations given the type of simulation selected. In
running a DNS (provided it is doable) one is interested in a fully resolved ow
eld and the grid should be sufciently ne to catch all motions. In running a
LES, the grid should be sufciently ne for the subgrid scales to become in
dependent of the macroow.
The number of equations to be solved is, among other things, related to the
turbulence model chosen (in comparison with the k e model, the RSM involves
ve more differential equations). The number of equations further depends on
the character of the simulation: whether it is 3D, 2
1/2
D, or just 2D (see below,
under The domain and the grid). In the case of twophase ow simulations, the
use of twouid models implies doubling the number of NS equations required
for singlephase ow. All this may urge the development of more efcient so
lution algorithms. Recent developments in computer hardware (faster proces
sors, parallel platforms) make this possible indeed.
Various numerical techniques are available for discretizing the set of partial
differential equations to be solved. Discretizing essentially is the method of con
verting the (partial) differential into algebraic equations by transforming (partial)
derivatives into nitedifference formulations. First of all, most (commercial) ow
simulation codes exploit the nitevolume (or nitedifference) method that has
been discussed extensively by Shyy (1994). An introduction to the concept can be
found with, e.g., Abbott and Basco (1989) and Shaw (1992). Many more tech
niques are available for discretization, such as niteelement, spectral, arclength,
and lterscheme methods, which are beyond the scope of the present review. The
result is a (large) set of algebraic equations anyway: one algebraic equation per grid
point for each ow variable that connects the value of a particular ow variable at
a particular grid point with those at a number of neighboring grid points.
A. FINITE VOLUME TECHNIQUES
Most commercial CFDcodes are based upon the ideas and numerical
methods developed back in the 1970s at Imperial College London by Spalding,
Patankar, Gosman, and others:
the staggered grid concept (velocity components and scalar quantities such as
pressure are not dened at the same mesh points),
The hydrodynamic quantities such as the velocities and velocity gradients are
determined locally in a LB code. As a result, the local character of the LB
technique is strongly in favor of (massive) parallellization of the computa
tional job via domain decomposition: the communication between CPUs
relates to grid cells very near to the subdomain boundaries only. On the
contrary, in solving the NS equations iteratively with a FV solver the eld
properties result in longrange variations propagating across the boundaries
of subdomains and, hence, requiring intense communication between CPUs
dealing with neighboring subdomains.
More basically, LB with its collision rules is intrinsically simpler than most
FV schemes, since the LB equation is a fully explicit rstorder discretized
scheme (though secondorder accurate in space and time), while temporal
discretization in FV often exploits the CrankNicolson or some other mixed
(i.e., implicit) scheme (see, e.g., Patankar, 1980) and the numerical accuracy in
FV provided by rstorder approximations is usually insufcient (Abbott and
Basco, 1989). Note that fully explicit means that the value of any variable at
a particular moment in time is calculated from the values of variables at the
previous moment in time only; this calculation is much simpler than that with
any other implicit scheme.
A comprehensive and more extensive overview of the pros and cons of LB
with respect to applications can be found in Succi (2001). By the way, LB
methods are continuously improved to increase speed and accuracy, particularly
by introducing grid renement techniques and advanced techniques for arbi
trarily shaped boundaries (e.g., Rohde et al., 2002, 2003, 2006; Rohde, 2004).
It makes sense to compare the implications (in terms of simulation times) of
using FV vs. LB in simulating turbulentow elds in process devices. Hoekstra
(2000) demonstrated the numerical implications of applying different numerical
schemes in an industrial application. He compared the outcome of his RANS
simulation for a gas cyclone with that of a LES carried out by Derksen and Van
den Akker (2000). Table I presents a number of numerical features of the two
types of simulations.
THE DETAILS OF TURBULENT MIXING PROCESS 177
Not only did the results of the LES agree much better with the experimental
data than those of the RANS simulation, the LB code was also roughly 30 times
faster than the FV code in terms of simulation time per grid node per step. To
put it in a more practical sense: performing this LES on, e.g., 18 processors with
the FV code would take 2 months, whereas a simulation as such with LB would
take 2 days only.
Another comparison is due to Van Wageningen et al. (2004) who performed a
similar study (in terms of the numerical scheme used) on unsteady laminar ow
in a Kenics
s
static mixer. They found that the LB code was 500600 times
faster than FLUENT in terms of simulation time per grid node per time step and
that FLUENT used about 5 times more memory than LB.
The difference in speed between a LB code and a FV code in the above studies
partly originates from their different character: the FV code in a general
purpose commercial CFD code should be robust and suited in many applica
tions, while the LB codes used are of a research type and usually strongly
dedicated and geared to a specic job.
IV. Boundary Conditions
Solving sets of (partial) differential equations inherently requires the spec
ication of boundary conditions and, in case of transient simulations, also in
itial conditions. This is not as simple as it looks like, especially for turbulent
ows in complex process equipment.
Whenever a free surface is present at some (mean) xed position, most CFD
codes assume it to be strictly at, while in the direction normal to the free
surface velocities and gradients of most variables are taken zero. Usually, this is
accomplished by dening mirror cells at the free surface. It is not clear what
the effect is of the use of such mirror cells on the ow eld in the upper part of
the vessel in comparison with real life where the surface is not necessarily at.
A. MOVING BOUNDARIES
An aspect of CFD in stirred vessels that needs separate discussion is the issue
of the revolving impeller and the way its motion is dealt with in the simulations.
In the early days, see, e.g., Bakker and Van den Akker (1994a), a black box
representing the impeller swept volume was often used in RANS simulations,
with boundary conditions in the outow of the impeller which were derived
from experimental data. The idea behind this approach was that such near
impeller data are hardly affected by the rest of the vessel and therefore could be
used throughout. Generally, this is not the case of course. Furthermore, this
approach necessitates the availability of accurate experimental data, not only
HARRY E. A. VAN DEN AKKER 178
with respect to the average velocity components, but also for k and e. The latter
variable in particular can hardly be measured directly. Nowadays, this approach
is no longer used, also due to the steep increase in computer power which no
longer urges for such drastic simplications.
In those early days, when computer power was limited, often use was made of
a symmetry assumption: each quarter of the vessel containing one of the four
bafes at the vessel wall was supposed to behave identically; hence, a steady
ow in the RANS approach was simulated in just a quarter vessel. Such strong
simplications are no longer in use. Precessing vortices moving around the
vessel centerline contribute to ow unsteadiness and, therefore, exclude models
that just assume ow steadiness or allow for domain reductions through geo
metrical symmetries. The most correct response to this ow unsteadiness is the
concept of LES.
Later on, in 1994, novel options were introduced such as Sliding Mesh (SM)
and Multiple Frames of Reference (MFR) in which the ow domain is divided
into two parts, each with their own meshing; one mesh is connected to the
stationary vessel wall and the other one to the revolving impeller. The inter
action of the ow elds at either side of the interface between these two meshes
requires delicate bookkeeping of uxes and forces among cells moving with
respect to one another at the interface. Of course, these methods are a drastic
improvement over the black box description of the impeller swept area applied
in the early days of stirred vessel CFD. The mesh associated with the impeller in
SM is perfectly capable of simulating the unsteady ow around the impeller
blades including the trailing vortices.
The MFR technique introduced by Luo et al. (1993, 1994) starts from a
steadystate description of the ow eld and therefore ts in RANSbased simu
lations only. The SM approach (e.g., Murthy et al., 1994; Bakker et al., 1997) is
a fully transient method, also applicable in LES (see, e.g., Bakker et al., 2000;
Jahoda et al., 2006; Gao and Min, 2006; Gao et al., 2006). Yeoh et al. (2004a, b)
adopted a Sliding and Deforming Mesh (SDM) technique in which the two
grids do not only slide with respect to each other but also feel shear resulting
in deforming interface cells. Keeping mass and momentum conserved in this
technique requires bookkeeping somewhat different from that with the SM
technique.
Generally, however, it is unclear what (with SM) the effect is of the position
of the interface between stationary and moving meshes on the simulated un
steadiness of the overall ow in the vessel. In addition, simulations making use
of SM and MFR may suffer from slower or poor convergence. Of course, the
transient SM technique is more accurate, though at the cost of larger computer
time consumption. Montante et al. (2006) even reported that MFR yielded
unphysical ndings, viz. a region of opposite swirl.
Harvey et al. (1995) and Harvey and Rogers (1996) proposed a multiblock
impellertted grid structure for dealing with the exact geometry of the impeller.
The rst of these two papers introduces an approximate steadystate method
THE DETAILS OF TURBULENT MIXING PROCESS 179
that solves the viscous ow with the impeller at one position with respect to the
bafes ignoring its relative motion, while the second paper is about an unsteady
moving grid approach that now takes the relative motion of the impeller with
respect to the bafes into account. This approach, however, never made it into
to the turbulentow regime.
Ranade and Van den Akker (1994) and Ranade and Dommeti (1996) intro
duced a snapshot approach in which the impeller was put in a standstill and the
revolving ow in the simulation was obtained by imposing velocity jets at one or
more xed positions of the impeller blades. At a rst glance, realistic ow elds
were obtained, energy dissipation levels and the total amount of energy dis
sipated being in the right order of magnitude. A more detailed assessment of this
approach, however, reveals that the equations used are not invariant to the type
coordinate transformation used. Furthermore, elementary turbulence theory
(e.g., Tennekes and Lumley, 1972) indicates that, with a view to the ow eld in
the impeller swept volume, a jet issuing from the face of an impeller blade may
not be equivalent to a wake behind an impeller blade, as jets and wakes obey
different laws for their expansion in downstream direction. As a result, the
shape of the impeller connected zones of high k and e values may not be pre
dicted condently. Ranades snapshot approach (still in use, see, e.g., Khopkar
et al., 2006) discussed further on page 207 should therefore be abandoned, in
spite of all explanations devoted to it (Ranade, 2002).
Applying Immersed or Embedded Boundary Methods (Mittal and Iaccarino,
2005) circumvents the whole issue of the friction between the more or less steady
overall ow in the bulk of the vessel and the strongly transient character of
the ow in the zone of the impeller. These methods are introduced below. In the
context of a LES, Derksen and Van den Akker (1999) introduced a forcing
technique for both the stationary vessel wall and the revolving impeller. They
imposed noslip boundary conditions at the revolving impeller and at the sta
tionary tank wall (including bafes). To this purpose, they developed a specic
control algorithm.
B. CURVED BOUNDARIES
In stirred vessels and static mixers the ow domain is bounded by complex
boundaries due to the curvature of containing walls, the revolving impeller axis
and/or static mixing elements.
While in the early days a staircase representation of a curved boundary in
a cubic grid was quite common, commercial CFD software nowadays ex
ploits boundary tted meshing. The staircase representation usually was not a
problemin terms of mass conservation or of introducing artefacts such as
additional small eddiesas long as the steps did not involve more than a single
grid cell. The currently widely adopted boundary tted meshing, i.e., generating
bodyconformal either structured or unstructured grids providing adequate
HARRY E. A. VAN DEN AKKER 180
local resolution with a minimum but large number of grid cells, requires either
commercial grid generating software or extensive coding (usually far beyond the
reach of academic groups). The impact of a poor grid on accuracy and on
convergence properties of the solver may not be underestimated.
Whenever a cubic grid is mandatoryeither due to coding limitations from
the part of academic groups or due to the inherent properties of, e.g., LB
techniquesand a staircase approach is to be avoided (e.g., for a revolving
impeller axis) one can take refuge to some immersed boundary method (see,
e.g., Mittal and Iaccarino, 2005). One may distinguish between
embedded boundary methods using cut cells: these methods adjust cell volume
and face areas to the geometry of the body in the ow domain; while in a FV
approach this method guarantees global and local conservation of mass and
momentum, it creates problems in an explicit formulation of the FV scheme;
immersed boundary methods using ghost cells: ghost cells are boundary cells
the centers of which are lying outside the ow domain; in this approach,
values of variables in these ghost cells are required to satisfy, e.g., a Neumann
boundary condition (velocity gradients normal to a wall zero, velocity com
ponent along a wall zero at the wall).
Which of the various immersed or embedded boundary methods is best
generally or for a particular caseis still an open question. Thornock and Smith
(2005) introduced a Cell Adjusted Boundary Force Method for a stirred vessel.
All methods proposed so far have their own pros and cons. Immersed boundary
methods are also exploited in LB techniques (e.g., Derksen and Van den Akker,
1999). Rohde (2004) investigated the use of triangular facets for representing a
spherical particle.
C. THE DOMAIN AND THE GRID
With a view to any simulation, a few important items have to be addressed.
First of all, it has to be decided whether the ow to be simulated is 2D, 2
1/2
D,
or 3D. When the ow is, e.g., axissymmetrical and steady, a 2D simulation
may sufce. For a ow eld in which all variables, including the azimuthal
velocity component, may not depend on the azimuthal coordinate, a 2
1/2
D
simulation may be most appropriate. Most other cases may require a full 3D
simulation. It is tempting to reduce the computational job by casting the
3D ow eld into a 2D mode. The experience, however, is that in 2D simu
lations the turbulent viscosity tends to be overestimated; in this way, the ow
THE DETAILS OF TURBULENT MIXING PROCESS 181
may become more or less diffusion dominated and less capable of sustaining
a transient behavior. Furthermore, between 2D and 3D, the wall to volume
ratio is different, turbulence intensity may go down, and hydrodynamic stability
may be affected. Of course, the above remarks apply to RANSbased simu
lations only, as LES are inherently transient and 3D.
A second choice to be made relates to the size of the ow domain. It may
be worthwhile to limit the calculational job by reducing the size of the ow
domain, e.g., by identifying an axis or plane of symmetry, or, in a cylindrical
vessel with bafes mounted on the wall, due to periodicity in the azimuthal
direction. Commercial software accomplishes these choices by means of sym
metry cells and cyclic cells, respectively; although such choices reduce the size
of the simulation, they may eliminate the possibility of nding the real (asym
metric, unstable, or transient) 3D ow eld. The presence of feed pipes or drain
or withdrawal pipes may also make the use of symmetry or cyclic cells impos
sible. Again, this issue only plays a role in RANStype simulations.
A third issue is how many grid cells should be used for the domain size
selected. In general, using more computational cells implies more detailed in
sight in the ow eld as a result of the simulation, at the cost of longer CPU
times (although sometimes convergence rate may increase as a result of in
creasing the number of grid cells.) This is a tradeoff that is to be decided upon,
each time a new ow eld is to be simulated. In some cases, a stepwise approach
may be pursued to zoom into a particular zone within a ow device (see, e.g.,
Stekelenburg et al., 1994).
A fourth issue is the use of local grid renement: many commercial codes
offer this modality, often with unstructured grids. The rationale behind the idea
of introducing local grid renement techniques is that the grid is only rened in
those parts of the calculation domain in which the ow exhibits strong spatial or
temporal gradients. By doing so, one does not waste grid cells in parts of the
domain where the ow does not vary signicantly. Rohde et al. (2006) explored
the use of a generic, massconservative local grid renement technique within a
LB technique. He successfully attempted this technique for various rather sim
ple cases; so far, it has not been exploited in a cylindrical stirred vessel in which
the ow would require a ne grid not only in the impeller swept region but also
near to the bafes.
Anyhow, the result of any ow simulation should be grid independent, i.e.,
the ow eld should not be different if a ner grid is used for the simulation.
This test is to be carried out always and everywhere, and may really be described
as a conditio sine qua non.
Finally, it is good to quote Shaw (1992, p. 227): the most important asset in a
CFD analysis process is the analyst, who actually translates the engineering
problem into a computational simulation, runs the CFD solver and analyzes the
results. It is the skill of this person, or set of persons, that will determine whether
all the hardware and software will be utilized in the best possible way and
produce good quality results. CFD is certainly not a panacea that may solve all
HARRY E. A. VAN DEN AKKER 182
possible design and optimization problems. It rather is a tool that in the hands
of a welltrained professional may provide valuable insights in the local phe
nomena and processes, which take place in (bio)reactors and all sorts of process
equipment including mixing devices. CFDsimulations may provide a good
starting point for a mechanistic description of operations such as blending,
suspending solids, dispersing gas, and carrying out (bio)chemical reactions.
V. Simulations of Turbulent Flows in Stirred Vessels
When exploiting computational techniques for studying the mixing perform
ance of stirred tanks and static mixers, the rst question is how well the
turbulentow eld and the characteristic turbulence properties are predicted
by the various forms of CFD discussed above. The different forms of CFD
(RANS, LES, DNS) and the various strategies, turbulence models, and sub
models used by them may suffer from a different degree of validity in the
various mixing devices considered. The impact of a limited validity of the
models used may vary from case to case. That is why validation of simulation
results is still an urgent issue. Sometimes, the results of computational simu
lations just look qualitatively correct. It is important to check code performance
by means of quantitative experimental data as to average quantities as well as
local and transient uctuations.
The next question then is whether the processes taking place inside this tur
bulentow eld can be modeled with condence. We will now rst consider the
rst question.
A. TURBULENCE PROPERTIES
It is worthwhile to verify whether or not some basic assumptions of turbu
lence theory (local equilibrium between production and dissipation of turbu
lence, isotropy, homogeneous turbulence) which are used in modeling certain
aspects of momentum transport via turbulent eddies are met with in stirred
tanks. To be honest, at hardly any position in a stirred tank the production rate
of turbulent kinetic energy balances the rate of its dissipation. Production of
turbulence mainly takes place in the impeller swept volume, while a much larger
part of the vessel takes part in dissipating the turbulence. In addition, the action
of a revolving impeller and its interaction with nearby bafes turns the ow
intrinsically unsteady. As a result, the turbulent ow in a stirred vessel is cer
tainly not an equilibrium ow, as presumed in using, e.g., the Smagorinsky SGS
model in LES.
Most RANSbased simulations make use of the ke model for taking into
account the momentum transport due to the turbulent eddies. This model is an
THE DETAILS OF TURBULENT MIXING PROCESS 183
eddyviscosity model and as such it assumes isotropic turbulent transport. The
question is whether everywhere in a stirred tank the turbulent ow is locally
isotropic. This issue might have to be explained a bit further as local isotropy
should be distinguished from just isotropy. Local isotropy means that the
uctuations can be modeled with an (isotropic) eddy viscosity, while isotropy
has been dened as having the same uctuation levels in the three coordinate
directions. Local isotropy does not imply isotropy of the uctuations: a ke
model can predict nonisotropic uctuations.
The question whether or not stirred tank ow is locally isotropic, may be
investigated with the help of a LES which resolves a great deal of the Reynolds
stresses. To this end, the Reynolds stress data are best presented in terms of the
socalled anisotropy tensor a
ij
and its invariants A
1
, A
2,
and A
3
.
The anisotropy tensor is related to the turbulent stresses, of course, and is
dened as
a
ij
u
i
u
j
k
2
3
d
ij
(15)
Its rst invariant A
1
is equal to zero by denition. The second and third in
variants of this tensor are A
2
a
ij
a
ji
and A
3
a
ij
a
jk
a
ki
, respectively. The range
of physically allowed values of A
2
and A
3
is bounded and represented by the so
called Lumley triangle in the (A
3
, A
2
) plane (Lumley, 1978). The dis
tanceA O(A
2
2
+A
3
2
) from the isotropic state, i.e., from the origin (A
2
0,
A
3
0), is a measure of the degree of anisotropy. See also Escudie and Line
(2006) for a more extensive discussion as to how to quantify and visualize how
different from isotropic turbulence a stirred vessel is.
Phaseaveraged values ofAin a plane midway between two bafes of a
stirred tank have been plotted in Fig. 1 (from Hartmann et al., 2004a) for two
different SGS models (Smagorinsky and Voke, respectively) in LES carried out
in a LB approach. The highest values, i.e., the strongest deviations from isot
ropy, occur in the impeller zone, in the boundary layers along wall and bottom
of the tank, and at the separation points at the vessel wall from which the
anisotropy is advected into the bulk ow. In the recirculation loops, the tur
bulent ow is more or less isotropic.
Fig. 2 (also from Hartmann et al., 2004a) shows how the values of the in
variants A
2
and A
3
found in a LES using a Smagorinsky SGS model are dis
tributed within the Lumley triangle. Most but not all of the points are clustered
in the lower part of this triangle. This implies that in a large part of a stirred
tank the turbulent ow is more or less isotropic and the use of a turbulent
viscosity and a ke model are permitted. On the whole, however, turbulent ow
is not really isotropic. This may also explain why in RANSbased simulations
turbulent kinetic energy levels generally are underestimated.
As far as e is concerned: it is extremely difcult to measure its spatial dis
tribution and its wide range of values experimentally. LES may provide an
HARRY E. A. VAN DEN AKKER 184
FIG. 1. Phaseaveraged plots of the anisotropy distanceAin a plane midway between two bafes
in a stirred vessel provided with a Rushton turbine, as obtained by means of LES, with two different
SGS models: (a) the Smagorinsky model; (b) the Voke model. Reproduced with permission from
Hartmann et al. (2004a).
FIG. 2. This plot shows to which degree, according to a LES, the turbulence in a plane midway
between two bafes in a stirred vessel provided with a Rushton turbine can be typied. For clarity,
not all grid points in such a plane have been used for this plot. According to Lumley (1978), the
borders represent different types of turbulent ows: 3D isotropic turbulence, 2d axissymmetric
turbulence, 2D turbulence, and 1D turbulence. Most but not all points are concentrated in the
lower part of this Lumley triangle. Reproduced with permission from Hartmann et al. (2004a).
THE DETAILS OF TURBULENT MIXING PROCESS 185
attractive alternative. Micheletti et al. (2004) present a valuable discussion on
this issue.
B. VALIDATION OF TURBULENT FLOW SIMULATIONS
In view of the different requirements as to computer power, it is very worth
while to compare the outcome of RANS and LES simulations mutually and/
or with quantitative experimental data. Several authors have done this, e.g.,
Derksen and Van den Akker (1998, 1999), Derksen (2001), Lu et al. (2002),
Ranade (2002), Derksen (2002b), and Yeoh et al. (2004a,b). In most cases, the
experimental data have been obtained by means of LaserDoppler Anemometry
(LDA, or LDV): see, e.g., Yianneskis et al. (1987), Wu and Patterson (1989),
Scha fer et al. (1997, 1998), and Derksen et al. (1999). In this review, we will
mainly refer to the validation study due to Hartmann et al. (2004a).
In comparing RANS results and LES results with LDA data, the focus is rst
on the global, phaseaveraged ow eld: see Fig. 3 (from Hartmann et al.,
2004a) that relates to a plane midway between two bafes. The two types of
simulations capture the dominant ow feature, viz. the two large circulation
loops, more or less to the same extent. The RANS simulation better predicts the
position of the point where the upper loop separates from the vessel wall. As far
as turbulent kinetic energy (k) levels are concerned, kvalues obtained in a LES
relate to the resolved turbulent eddies. Hartmann et al. (2004a) argued that the
SGS uctuations hardly contribute to the klevels. It is evident from Fig. 4 that
the RANS simulation underestimates the kinetic energy levels created by the
blades of the Rushton impeller. The same conclusion applies to phaseresolved
kinetic energy data (see Fig. 5). The RANS simulation is hardly capable of
catching the remainders of the trailing vortices created by the preceding impeller
blade, while LES nicely reproduces this succession of vortices. Whenever one is
interested in details of the turbulentow eld because they may affect the
performance of the mixing device, one should really consider carrying out a
LES. The ndings due to Yeoh et al. (2004a) are completely in line with this.
As long as the interest is in elds of averaged velocity components and in
overall mixing patterns, RANSbased simulations may sufce. Examples of
such satisfactory simulation results are plentiful, e.g., Marshall and Bakker
(2004) and Montante et al. (2006). When, however, the interest is in the details
of the turbulentow eld and in processes affected by these details, LES is
the option to be preferred. From the ndings reproduced above and from the
validation studies of Derksen and Van den Akker (1999) and Derksen (2001) the
general conclusion is that, as long as the spatial resolution is sufcient, LB LES
deliver results in excellent agreement with experimental turbulence data.
Large eddy simulations explicitly resolves the inherently unsteady character
of the turbulent ow in a stirred tank into account, including the periodic
phenomena associated with the motion of the impeller and their interaction with
HARRY E. A. VAN DEN AKKER 186
FIG. 3. Velocity vector elds and levels of turbulent kinetic energy in a plane midway between two
bafes in a stirred vessel, according to LDA data (a); a RANSbased simulation (b); and two LES (c)
and (d). Reproduced with permission from Hartmann et al. (2004a).
THE DETAILS OF TURBULENT MIXING PROCESS 187
the bafes. Global as well as subtle ow features are in quantitative agreement
with experimental data. Typical examples of such a quantitative agreement
arefor sixbladed Rushton turbinesthe path along which the trailing vor
tices developing at the impeller blades are swept into the bulk of the tank as well
as the turbulent kinetic energy levels in the wakes of the impeller blades; in a
vessel equipped with a pitchedblade turbine, we could mention the primary and
secondary recirculations in the phaseaveraged ow eld.
Derksen (2001) and Hartmann et al. (2004a) demonstrated that various
choices in SGS modeling did not have a big impact on the quality of the ow
eld predictions. In addition, all these studies did not reveal a signicant effect
of the value of the Smagorinsky coefcient within the range 0.080.14.
One of the complications in stirred tank ows is the presence of macroin
stabilities (i.e., lowfrequency mean ow variations) that may affect the mixing
performance. Various authors have distinguished between various types and in
vestigated their occurrence and their frequencies under varying operating con
ditions and with several types of vessels and impellers (Yianneskis et al., 1987;
Haam et al., 1992; Myers et al., 1997; Hasal et al., 2000; Nikiforaki et al., 2002).
FIG. 4. Phaseaveraged plots of turbulent kinetic energy in the vicinity of the Rushton impeller as
found in different types of simulations as indicated. Reproduced with permission from Hartmann
et al. (2004a).
HARRY E. A. VAN DEN AKKER 188
Roussinova et al. (2003) and Hartmann et al. (2004b) found that the LES
methodology is excellently capable of reproducing various types of macroin
stabilities. While Nikiforaki et al. (2002) experimentally found precessing fre
quencies in the range 0.0130.018 N, the LES of Hartmann et al. (2004b) arrived
at 0.0228 and 0.0255 N as the dominant frequencies, where N stands for the
impeller speed (in number of revolutions/s). The discrepancy between experi
mental and numerical frequencies is a challenge for further improving parti
cularly the SGSmodel and some numerical settings.
VI. Operations and Processes in Stirred Vessels
In simulating physical operations carried out in stirred vessels, generally one
has the choice between a Lagrangian approach and a Eulerian description.
While the former approach is based on tracking the paths of many individual
uid elements or dispersedphase particles, the latter exploits the continuum
concept. The two approaches offer different vistas on the operations and require
different computational capabilities. Which of the two approaches is most
FIG. 5. Phase resolved plots of velocity vector eld and turbulent kinetic energy in a plane 15
o
behind an impeller blade (obtained by sampling data only when the measuring point is at the
specied position with respect to the impeller blades). Not all vectors have been plotted for clarity.
Reproduced with permission from Hartmann et al. (2004a).
THE DETAILS OF TURBULENT MIXING PROCESS 189
suited is hard to say: it depends on the details of the issue of interest and on the
computational power available.
In whichever approach, the common denominator of most operations in
stirred vessels is the common notion that the rate e of dissipation of turbulent
kinetic energy is a reliable measure for the effect of the turbulentow charac
teristics on the operations of interest such as carrying out chemical reactions,
suspending solids, or dispersing bubbles. As this e may be conceived as a con
centration of a passive tracer, i.e., in terms of W/kg rather than of m
2
/s
3
, the
spatial variations in e may be calculated by means of a usual transport equation.
In the context of the RANSmethodology, this e is also required for solving
the spatial distributions of the velocity components, while in LES e just serves
the purpose of providing a basis for modeling the operation(s) of interest.
Even when the number of grid cells in a LB LES simulation of a stirred vessel
1.1 m
3
in size amounts to some 36 10
6
grid cells, this implies a cell size, or grid
spacing, of 5 mm only. Even a cell size of just a few millimeters makes clear that
substantial parts of the transport of heat and species as well as all chemical
reactions take place at scales smaller than those resolved by the ow simula
tion. In other words: concentrations of species and temperature still vary and
uctuate within a cell size. The description of chemical reactions and the trans
port of heat and species therefore ask for subtle approaches to these SGS
uctuations.
A. MIXING AND BLENDING
One of the most crucial (design) parameters in blending two miscible liquids
or distributing a particular miscible addition over a heel of liquid is the socalled
mixing time, i.e., the time needed to achieve complete homogenization or a
predetermined degree of homogeneity (see, e.g., Grenville and Nienow, 2004).
In this review, the focus is on blending operations carried out with low viscosity
uids in the turbulent regime.
Kramers et al. (1953) were among the rst to study mixing times as a function
of bafe position and impeller rotational speed. Results of several experimental
studies have been combined into empirical correlations relevant to industrial
applications (Procha zka and Landau, 1961; Hoogendoorn and Den Hartog,
1967; Sano and Usui, 1985; Grenville, 1992; Ruszkowski, 1994; Nienow, 1997).
Grenville and Nienow (2004) present a concise review as to such correlations.
Bouwmans (1992; see also Bouwmans et al., 1997) used a particle tracking
technique in a RANS ow eld to estimate trajectories of neutral and buoyant
additions, to construct Poincare sections of additions crossing specic horizon
tal crosssectional planes, to predict probabilities of surfacing for buoyant
additions, and to mimic the temporal response of conductivity probes.
Exploiting the Eulerian point of view, Ranade et al. (1991) was one of the rst
to simulate mixing times for a vessel provided with a pitched blade turbine; in
HARRY E. A. VAN DEN AKKER 190
this early work, the impeller swept area was still modeled as a black box just
delivering inlet conditions for the turbulent ow in the remainder of the vessel.
Various authors (Osman and Varley, 1999; Jaworski et al., 2000; Bujalski et al.,
2002) used SM or MFR techniques in RANSbased simulations for Rushton
turbines driven ow combined with species transport equations to predict mix
ing times, but arrived at values 23 times higher than the experimental values.
This is in line with the ndings for Rushton turbines that in such RANSbased
simulations turbulence levels are underpredicted and that mixing across the
central plane of the discharge plane is poorly reproduced. Montante and
Magelli (2004) studied a homogenization process in a bafed vessel stirred with
various sets of Rushton turbines; while effects of varying impeller number and
spacing were correctly forecasted, a very low turbulent Schmidt number had
to be adopted for obtaining good quantitative agreement in terms of tracer
response curves. RANSbased simulations of mixing in tanks provided with
pitched blade turbines prove to underpredict mixing times, probably owing to
mesoscale concentration uctuations not really reproduced by the simulation.
All these ndings of disappointing quantitative agreement with experimental
data stem from the inherent drawback of the RANSapproach that there is no
clear distinction between the turbulent uctuations modeled by the Reynolds
stresses and (mesoscale) uctuations. In LES, however, the distinction between
resolved and unresolved turbulence is clear and relates to the cell size of the
computational grid chosen.
The LES methodology has recently been applied in a number of studies
simulating in a stirred tank the mixing in response to the addition of a tracer.
Yeoh et al. (2004a) carried out an FVsimulation that matched the experimental
setup of Lee (1995); the focus of their study is on mixing patterns, on traces of
concentrations at certain monitoring points, and on a comparison of predicted
mixing time with correlations from literature. Hartmann (2005) and Hartmann
et al. (2006) coupled an FV solver for the species transport to an LB ow solver.
In his study, different from Yeoh et al. (2005), the passive scalar is injected at
zero speed to avoid an effect of jet mixing on mixing times. Hartmann aimed at
reproducing the experiments of Distelhoff et al. (1997) and focused on the effect
of impeller size and injection position on mixing time. A sample of his results is
presented in Fig. 6. For the time being, the results due to Hartmann still suffer
from an unphysical mass increase caused by the current implementation of his
novel immersed boundary method using ghost cells.
Comparative studies on simulating mixing times by means of the traditional
RANS approach and the more sophisticated LES are due to Gao and Min
(2006), Gao et al. (2006), and Jahoda et al. (2006). They all show that
RANSbased simulations fail in reproducing the transient responses of probes
monitoring the local tracer concentrations, while LES is able to mimic
the experimental traces quite accurately (see Fig. 7, from Jahoda et al., 2006).
The latter traces strongly resemble those presented by Hartmann (2005) and
Hartmann et al. (2006).
THE DETAILS OF TURBULENT MIXING PROCESS 191
B. SUSPENDING SOLIDS
An important class of stirred tank applications involves suspending discrete
solid particles in a continuous liquid phase where the turbulentliquidow
FIG. 6. A sample of Hartmanns results (2005). The tracer is injected (with zero speed) at the top
of the tank in a plane midway between two bafes (black dot in the plots). The concentration elds
shown are also in the midway bafe plane. T/D 3, c
N
denotes the nal uniform concentration.
FIG. 7. Time traces of normalized concentration as seen by a probe in the lower part of a vessel
in simulations of a mixing time experiment. The vessel is provided with two Pitched Blade Turbines.
Three different types of simulations are shown, where ske stands for a standard ke simulation and
sm, mrf, and les have the usual meaning. Reproduced with permission from Jahoda et al. (2006).
HARRY E. A. VAN DEN AKKER 192
generated by the impeller induces particle motion and should prevent sedimen
tation. Such suspensions are for instance encountered in industrial crystalliza
tion and in catalytic slurry reactors. The usual procedure in designing such
processes is making use of the Zwietering correlation (1956). Note that in the
NAMF Handbook of Mixing, Chapter 10 on SolidLiquid Mixing does not
refer to any CFD results.
Although the Zwietering correlation provides valuable guidance to chemical
engineers in, e.g., solving practical engineering problems and dealing with scal
ingup issues, at a more fundamental level there are many unresolved issues
such as:
What is actually going on at the particle scale (in terms of, e.g., heat and mass
transfer, or mechanical load on particles as a result of particleparticle and
particleimpeller collisions) and how are these microscale events affected by
the largerscale phenomena?
On the reverse, how does the presence of particles affect local and global ow
features in the vessel such as the vortex structure in the vicinity of the impeller,
power consumption, circulation and mixing times, and the spatial distribution
of turbulence quantities; more specically: colliding particles have an impact
on the liquids turbulence (Ten Cate et al., 2004) while local particle con
centrations affect the effective (slurry) viscosity which may be useful in the
macroow simulations?
Most of these issues may best be studied by DNS, while other can better be
tackled by LES. Anyhow, RANSbased simulations are not very suited to this
purpose as the turbulence in the RANSapproach is not resolved at all but just
modeled. Below, typical DNS, LES, and RANSbased simulations of solids
suspensions will be reviewed in succession.
1. A Direct Numerical Simulations Approach
In view of secondary nucleation in crystallizers, Ten Cate et al. (2004) were
interested in nding out locally about the frequencies of particle collisions in a
suspension under the action of the turbulence of the liquid. To this end, they
performed a DNS of a particle suspension in a periodic box subject to forced
turbulentow conditions. In their DNS, the ow eld around and between
the interacting and colliding particles is fully resolved, while the particles are
allowed to rotate in response to the surrounding turbulentow eld.
Ten Cate et al. (2004) were able to learn from their DNS about the mutual
effect of microscale (particle scale) events and phenomena at the macroscale: the
particle collisions are brought about by the turbulence, and the particles affect
the turbulence. Energy spectra conrmed that the particles generate uid
motion at length scales of the order of the particle size. This results in a strong
increase in the rate of energy dissipation at these length scales and in a decrease
THE DETAILS OF TURBULENT MIXING PROCESS 193
of the turbulence at larger length scales. All these details were obtained just by
restricting the DNS to a small periodic box. Details of their approach will be
dealt with further on; a typical result is Fig. 12.
2. Results of Large Eddy Simulations
Derksen (2003), on the contrary, was interested in simulating the process of
solids suspension in a stirred tank; to this end, he tracked particles in the whole
tank in a Lagrangian sense, considering the particles as point particles and not
resolving the detailed ow eld between the particles. In other words, Derksen
applied a more supercial view on the particle suspension by dropping details,
and was rewarded with a picture of the full tank: see Fig. 8. Yet, Derksen was
able to track just over 6.7 million particles, to include effects such as particle
rotation, particleparticle collisions, particleimpeller collisions and even
twoway coupling, and to include uidparticle interaction forces such as the
Saffman force, the Magnus force, and forces due to stress gradients. Tracking all
these particles was done in a turbulentow eld obtained via an LB LES.
Derksen (2006a) continued along this line of approach andby means of a
clever strategymimicked the longtime behavior of solids suspension in an
unbafed tall stirred tank equipped with four hydrofoil impellers (Lightnin
s
A310). The time span covered by his LES amounted to some 20,000 impeller
revolutions (some 20 min). Running a LES for a Reynolds number of 1.6 10
5
over the entire time span is not an option, and for that reason a particular ow
FIG. 8. This is a snapshot of a spatial particle distribution. The plane shown is the horizontal
crosssection just below the disc of a Rushton turbine in a atbottomed stirred tank. The impeller
revolves in the counter clockwise direction. Particle size is some 0.468 mm; Re 1.5? 10
5
;
volume fraction amounts to 3.6%; number of particles tracked in the simulation just over 6.7
million. Reproduced with permission from Derksen (2003).
HARRY E. A. VAN DEN AKKER 194
time series of sufcient length is stored and repeatedly played for tracking the
paths of 20,000 particles 0.33 mm in diameter and with a density of 2,450 kg/m
3
.
The advantage of this approach is that use is made of the highly resolved ow
information of the LES. The simulations of the particle response to the typical
ow eld created by the four impellers showed the counterintuitive behavior
observed by Pinelli et al. (2001): almost all solid particles rise to the top of
the tank.
3. Results of Reynolds Averaged Navier Stokes Simulations
A very different type of CFDsimulation results are those due to, e.g., Montante
and Magelli (2005) who studied solids suspension in stirred tanks by means of
twouid CFD simulations, i.e., a RANStype of EulerEuler simulations. These
authors tested two commercial CFD codes, viz. FLUENT 6.0 and CFX4, for
various formulations of the twouid model, of the uidparticle interaction force,
and of the ke turbulence model for multiphase ow. Montante and Magelli just
focused on predicting axial proles of the solids concentration for three bafed
stirred tanks agitated with single and multiple impellers; they evaluated the effect
of the various model formulations on these proles and compared their various
predicted curves with experimental data. This type of simulation delivers data
relevant for engineering purposes of limited scope and depth only.
In this context, one component of the phase interaction force may need sepa
rate discussion: the drag force. While most authors use the SchillerNauman
equation
C
D
24
Re
p
1:0 0:15Re
0:687
p
(16)
for the relation between particle drag coefcient and the particle Reynolds
number (see, e.g., Ranade, 2002, or Derksen, 2003). Brucato et al. (1998),
however, reported experimental data showing that free stream turbulence may
signicantly increase particle drag coefcients. They proposed a novel corre
lation for predicting the effect of free stream turbulence on the particle drag
coefcient:
C
Dt
C
D
C
D
8:76 10
4
d
p
Z
K
3
(17)
in which C
D t
denotes the particle drag coefcient in a turbulent ow and C
D
stands for the usual particle drag coefcient as given by Eq. (16). Equation (17)
implies that the drag coefcient for particles the size of which is smaller than
or comparable with that of the smallest turbulent eddies (represented by the
local Kolmogorov length scale Z
K
) is hardly affected by freestream turbulence,
while the uidmechanical interaction of larger particles with turbulent eddies
becomes signicant and leads to an increase of particle drag.
THE DETAILS OF TURBULENT MIXING PROCESS 195
Montante et al. (2000), Micale et al. (2000), and Montante and Magelli (2005)
incorporated this novel correlation into their twouid simulations of solids
suspension in stirred vessels and arrived at pretty good agreement with experi
mental data. These authors claim that this correlation due to Brucato et al.
(1998) plays an essential role in arriving at this good agreement: introducing
the turbulence effect reduces the tendency of the larger particles to collect in the
bottom part of the vessel. One may argue, however, that the introduction of
a higher particle drag coefcient is rather articial and just compensates for
shortcomings in the twouid formulation used, such as ignoring all other
components of the phase interaction force. By the way, Micale et al. (2004) in
simulating particle suspension height ignored this effect of free stream turbu
lence and identied it as a secondorder effect only.
In some way, introducing an increased particle drag by means of Eq. (17)
resembles the earlier proposal raised by Bakker and Van den Akker (1994b) to
increase viscosity in the particle Reynolds number due to turbulence (in agree
ment with the very old conclusion due to Boussinesq, see Frisch, 1995) with the
view of increasing the particle drag coefcient and eventually the bubble hold
up in the vessel. Lane et al. (2000) compared the two approaches for an aerated
stirred vessel and found neither proposal to yield a correct spatial gas distri
bution.
In addition, it is dubious whether this new correlation due to Brucato et al.
(1998) should be used in any EulerLagrangian approach and in LES which
take at least part of the effect of the turbulence on the particle motion into
account in a different way. So far, the LES due to Derksen (2003, 2006a) did not
need a modied particle drag coefcient to attain agreement with experimental
data. Anyhow, the need of modifying particle drag coefcient in some way
illustrates the shortcomings of the current RANSbased twouid approach of
twophase ow in stirred vessels.
The present author wonders whether, due to its very nature (particularly, the
various assumptions as to averaging and the various modeling uncertainties), a
RANSbased twouid approach is suited for reproducing the details of solids
suspension in a stirred vessel. May be we should be satised with the gross
predictions of the current RANS methods and turn to LES for the details
of those processes which are dominated by the spatially distributed turbulence.
It is really a valid question how long we should keep trying and improving the
RANS twouid approach as now the increased computer power brings the
much more sophisticated LES within reach.
C. DISSOLVING SOLIDS
Properly simulating a dissolution process of solid particles in a stirred vessel
operated in the turbulentow regime urges for a very detailed simulation of the
turbulentow eld itself. Just reproducing the overall ow pattern by means of
HARRY E. A. VAN DEN AKKER 196
a RANStype of simulation is not sufcient! Our estimate is that a twouid
simulation (on the basis of RANS) yields too rough an approximation of the
process. The dissolution process may too strongly depend on the heavily uc
tuating ow and concentration elds around particles whichas the result of
the presence and action of eddies of various scalesmay move chaotically with
respect to one another while being advected through the vessel.
One really may need an inherently transient LES to capture all these details.
The ner the grid for such a LES, the more reliably the local transient con
ditions may be taken into account in reproducing this turbulent mass transfer
process (while ignoring the issue of supplying the heat for the dissolution which
may also depend on a proper representation of the turbulentow eld).
An additional important issue is how many particles have to be tracked for a
proper representation of the transient spatial distribution of the particles over
the vessel.
Hartmann et al. (2006) reported very detailed simulation results (see also
Hartmann, 2005) (Fig. 9). Their LB simulation was restricted to a labscale
vessel 10 L in size only for which 240
3
lattice cells a bit smaller than 1 mm
2
were
used; the temporal resolution was 25 ms only. A set of 7 million monodisperse
spherical particles 0.3 mm in size was released in the upper 10% part of the
vessel. At the moment of release, the local volume fraction amounted to 10%.
The particle properties were those of calcium chloride. The simulation was
carried out on 30 parallel processors of an SGI Altrix 3700 system and required
for 6 weeks for 100 impeller revolutions.
Figure 9 presents some typical results for the spatial distribution of the dis
solving particles, their sizes, and their size distribution. Initially, the particles
respond to the centrifugal forces in the vortices and collect at the outer regions
of the vortices, giving rise to streaky patterns. The dissolution process results in
decreasing particle sizes and, hence, in decreasing inertia to the effect that
gradually they start behaving like uid tracers. Overall, the solids and scalar
concentrations become more homogeneous in the course of the dissolution
process. The simulation even predicts that a plot of the Sauter mean diameter
d
32
vs. time may exhibit a minimum (not shown here).
D. PRECIPITATION AND CRYSTALLIZATION
The counterparts of dissolving particles are the processes of precipitation and
crystallization the description and simulation of which involve several addi
tional aspects however. First of all, the interest in commercial operations often
relates to the average particle size and the particle size distribution at the com
pletion of the (batch) operation. In precipitation reactors, particle sizes strongly
depend on the (variations in the) local concentrations of the reactants, this
dependence being quite complicated because of the nonlinear interactions of
uctuations in velocities, reactant concentrations, and temperature.
THE DETAILS OF TURBULENT MIXING PROCESS 197
The same applies to crystallizers, in which particle sizes and particle number
concentrations not only depend on nucleation and growth from supersaturated
mother liquid, but are also affected by sheardominated agglomeration and
by secondary nucleation as a result of particleparticle and particleimpeller
collisions. Some of the subprocesses involved may be limited to specic and
different parts of the vessel: e.g., nucleation may be restricted to a amelike
region around the outlet of a feed pipe (Van Leeuwen, 1998). In addition, in
FIG. 9. Snapshots of particle sizes and their spatial distribution in a vertical midway bafe plane
at two moments in time, along with the pertinent respective overall particle size distributions. The
diameter of the particles is enlarged by a factor of 10 for reasons of clarity. Grey colors represent
particle size with respect to the original particle size. Reproduced with permission from Hartmann
(2005).
HARRY E. A. VAN DEN AKKER 198
many cases, many more product properties are relevant, such as color, texture,
and particle morphology.
Gradually, the insight is growing that all these properties are not only
affected by the averaged process conditions in reactor or crystallizer, but also
by their spatial and temporal variations felt by the particles during their stay in
and on their pathways through the vessel (see, e.g., Rielly and Marquis, 2001).
The rst response of modelers was to introduce compartmental modeling
(Van Leeuwen, 1998; Bermingham et al., 1998; Ten Cate et al., 2000): the vessel
is divided into several compartments, each compartment being considered as
(more or less) ideally mixed and described in terms of averaged values of the
process parameters such as temperature, concentrations, and rate of turbulent
kinetic energy dissipation; these averaged values may vary strongly from com
partment to compartment. Although this approach is an improvement over
the traditional method of lumping all variations into a single average value, an
important problem still is how to decide on the number and size of the com
partments. Carrying out RANStype simulations would be an option for
selecting proper compartmentalization on a casetocase basis.
Seckler et al. (1995), Van Leeuwen et al. (1996), and Wei and Garside (1997)
were among the rst to exploit commercial CFD codes (FLUENT and
PHOENICS) for simulating precipitation reactors (of a particular simplied
design) by adding to their codes some simple precipitation kinetics, i.e., relations
for nucleation and particle growth. Van Leeuwen et al. (1996) included the rst
four moments of the crystal size distribution in their simulations. Van Leeuwen
(1998) was the rst to study precipitation in a stirred vessel; among other things,
he explored the option for extending his RANStype simulations with a routine
involving the use of PDFs to account for variations in the species concentrations
(see one of the next sections on chemical reactors). He did not arrive at a
satisfactory agreement with experimental data.
As discussed in several of the above sections, LES is much better suited to
represent the spatial and temporal variations in a turbulentow eld. Of
course, this is very relevant in precipitation reactors and crystallizers where the
particle formation is a highly nonlinear process. Progress is being made in
developing LES including detailed models (e.g., PDFs, or methods of moments)
for specic reactive systems (e.g., combustion, polymerization) under specic
conditions in simple geometries (e.g., tubular reactors): see also one of the next
sections on chemical reactors. In turbulent agitated precipitation reactors and
crystallizers, however, mixing is so intense and complex and so heavily dom
inated by the revolving impeller that so far no one has succeeded in simulating
the full process of nucleation, particle growth, agglomeration, and particle
breakup and in arriving at a reasonable prediction of the eventual particle size
distribution.
Certain isolated aspects of precipitation and crystallization which have suc
cessfully been studied by means of LES, are discussed below to illustrate the
progress being made in the eld of precipitation and crystallization. The LB
THE DETAILS OF TURBULENT MIXING PROCESS 199
technique exploited in these studies is very helpful, since LB easily allows par
allel computing and the related computational acceleration. These detailed
studies may make a similar simulation possible as presented above for a dis
solution process.
1. Agglomeration
Hollander (2002) and Hollander et al. (2001a,b, 2003) studied agglomeration
in a stirred vessel by adding a single transport equation for the particle number
concentration m
0
(actually, the rst moment of the particle size distribution)
@m
0
@t
r ~ vm
0
1
2
b
0
m
2
0
(18)
Note that the particle diffusion term is ignored, just like particle dispersion
due to SGS motions (this was found justied in a separate simulation). The
shape of the sink term in the righthand term of this equation is due to Von
Smoluchowski (1917) while the local value of the agglomeration kernel b
0
is
assumed to depend on the local 3D shear rate according to a proposition due to
Mumtaz et al. (1997).
This additional Eq. (18) was discretized at the same resolution as the ow
equations, typical grids comprising 120
3
and 180
3
nodes. At every time step,
the local particle concentration is transported within the resolved ow eld.
Furthermore, the local ow conditions yield an effective 3D shear rate that can
be used for estimating the local agglomeration rate constant b
0
. Fig. 10 (from
Hollander et al., 2003) presents both instantaneous and timeaveraged spatial
distributions of b
0
in vessels agitated by two different impellers; color versions
of these plots can be found in Hollander (2002) and in Hollander et al. (2003).
The results of these simulations conrm the suspicion of a large spread in
b
0
values across the vessel. Furthermore, the simulations show that agglomer
ation does not occur in the impeller region, as the hydrodynamic conditions
(shear!) are too severe for agglomerates to survive. A bulk region can be dened
in which agglomeration conditions are benecial. Due to the typical structure of
the ow (i.e., trailing vortices, largescale turbulent motion), the vessel contains
large gradients in agglomeration rates. It is this effect that causes large differ
ences in the local particle consumption rate and leaves the vessel with a very
uneven distribution in particle concentration.
Fig. 11 presents the results of some 30 simulations for various conditions and
two impeller types in terms of the mean agglomeration rate constant observed in
the various simulations vs. the vesselaveraged shear rate found in the simu
lations. The simulations all started from the dotted curve for relating local
agglomeration rate constant to local shear rate. A clear decrease in the maxi
mum of b
0
as well as a shift toward higher average shear rates was found which
are caused by the local nature of the nonlinear ow interactions only. These
HARRY E. A. VAN DEN AKKER 200
FIG. 11. The discrepancy between the original kinetic relation due to Mumtaz et al. (1997) and the
observed relation between mean agglomeration rate constant b
0
and volumeaveraged shear rate.
Symbols refer to individual numerical simulations (LES). RT stands for Rushton Turbine, PBT for
Pitched Blade Turbine. Reproduced with permission from Hollander et al. (2001b).
FIG. 10. Results of LESbased simulations of an agglomeration process in two vessels: one ag
itated by a Rushton turbine (left) and one agitated by a Pitched Blade Turbine (right). The two plots
show the agglomeration rate constant b
0
normalized by the maximum value, in a vertical cross
sectional plane midway between two bafes and through the center of the vessel. Each of the two
plots consists of two parts: the righthand parts present instantaneous snapshots; the lefthand parts
present spatial distributions of timeaveraged values after 50 impeller revolutions. Reproduced with
permission from Hollander et al. (2003).
THE DETAILS OF TURBULENT MIXING PROCESS 201
nonlinear effects may play an important role in our mediocre understanding of
the scaleup behavior of agglomeration processes in stirred vessels.
The computational demands for Hollanders simulations (in 2000) were typ
ically of the order of 1 CPU week on a PentiumIII/500 MHz at 400 MB of
memory per run. The computer code was fully parallelized and was run on a 12
CPU Beowulf cluster.
2. Colliding Particles
Finally, Ten Cate et al. (2001) studied secondary nucleation as a result
of crystalcrystal collisions by means of a twostep approach. The rst step
involved a LES of the complete crystallizer in which the liquid phase (typically
containing 1020%v of solids) is treated as a single phase with a homogeneous
density and viscosity. (In principle, density and viscosity may be allowed to vary
locally in the simulation, by invoking the help of a particle dispersion routine
(e.g., Liu, 1999) and coupling the resulting particle concentrations back to a
SGS model to modify the local values of density and viscosity.) The spatial grid
resolution (some 5.0 mm) used by Ten Cate typically was at least 10 times the
crystal size; the grid comprised some 36 million cells.
The second step in Ten Cates twostep approach was to focus on crystal
crystal interaction by means of an explicit twophase DNS of the turbulent
suspension that completely resolves the translational and rotational motions
and collisions of the spherical particles plus the turbulence of the liquid between
the particles. The particle motions are driven by the turbulent ow and the
particles affect the turbulent ow of the liquid in between. When particles
approach one another down to a distance smaller than the grid spacing, lubri
cation theory is exploited to bridge the gap between them.
At the start of a simulation, the particles (up to 3,900) are placed randomly,
without mutual contact and with zero velocity, into a fully developed turbulent
singlephase ow eld in a periodic box. This starting situation is subjected to a
precalculated force eld; this forcing involves a divergencefree whitenoise sig
nal, distributed over the wave number domain as a Gaussian distribution about
a desired wave number, with a characteristic root mean square velocity and a
characteristic length scale. In this way, turbulent conditions are generated which
accurately recover a priori set values such as the Kolmogorov length scale, the
integral length scale, and the integral time scale, derived from the LES at some
position somewhere in the crystallizer. For details, the reader is referred to Ten
Cate (2002) and Ten Cate et al. (2004).
The evolution of the twophase turbulence depends on the initial random
position of the particles, the motion of which modies the turbulentow eld
directly. These DNS are therefore a nice example of twoway coupling between
the two phases: see Fig. 12. From these DNS, detailed knowledge can be derived
as to the frequency of the particleparticle collisions and the forces involved
HARRY E. A. VAN DEN AKKER 202
(as a function of local conditions in a crystallizer) which may yield quantitative
data on secondary nucleation (fragmentation of crystals due to collisions).
VII. Stirred GasLiquid and LiquidLiquid Dispersions
Similar problems as encountered in simulating agitated suspensions also
play a role in agitated gasliquid and liquidliquid systems. Again, there is
the dilemma whether to use a Lagrangian approach or to apply a twouid,
EulerEuler, model. Then, there is again the question of oneway coupling vs.
twoway coupling. Thirdly, the issue of taking the particle size distribution into
account (yes or no) should be addressed. Further, particleparticle interaction
may not be ignored in many cases, and then the additional problem is coa
lescence of bubbles or droplets, and their breakup. All these issues have not
been tackled simultaneously in the past, not in the least because of computa
tional limitations.
Furthermore, the physics of the interaction between turbulence and bubbles
in the complex ow of a stirred vessel, with its implications for coalescence and
breakup of bubbles and drops, is still far from being understood. Up to now,
simple correlations are available for scaleup of industrial processes; generally,
these correlations have been derived in experimental investigations focusing on
the eventual mean drop diameter and the drop size distributions as brought
1.210
1
1.210
2
0.4
FIG. 12. Snapshot from a twophase DNS of colliding particles in an originally fully developed
turbulent ow of liquid in a periodic 3D box with spectral forcing of the turbulence. The particles
(in blue) have been plotted at their position and are intersected by the plane of view. The arrows
denote the instantaneous ow eld, the colors relate to the logarithmic value of the nondimensional
rate of energy dissipation.
THE DETAILS OF TURBULENT MIXING PROCESS 203
about by the power input (mainly) via the impeller given the uid properties
(e.g., Colenbrander, 2000).
In the 1980s, Issa and Gosman (1981), Pericleous and Patel (1987), and
Tra ga rdh (1988) made the rst attempts to simulate aerated stirred vessels
computationally. Their results were rather approximate indeed, and were not
validated by means of experimental data.
A. BAKKERS GHOST! CODE
In the early 1990s, Bakker and Van den Akker (1991, 1994) introduced an
approximate but effective EulerEuler approach (see also A. Bakkers PhD
Thesis, 1992): on the basis of a singlephase RANS ow eld calculated by
FLUENT, a code named GHOST! calculated local and averaged values of
bubble size d
b
, gas holdup a, and specic mass transfer rate k
l
a.
Their GHOST!code essentially consisted of a mass balance for the gas, a
transport equation for the bubble number density n
b
, and a force balance for a
single bubble, respectively, which run as
r a u r r D a
S
g
(19)
@n
b
@t
r n
b
u o n
b1
n
b
S
g
V
b;in
(20)
r
l
V
b
w
2
r
^ r r
l
gV
b
^ z C
D
1
2
r
l
u
s
j ju
s
p
4
d
2
b
(21)
respectively. Of course, the variables n
b
, d
b,
and a are interrelated. The second
term in the mass balance, Eq. (19) stands for the extra transport of the bubbles
due to the larger turbulent eddies; usually, this turbulent transport is taken into
account by a separate term in the momentum balance for the dispersed phase
which, however, in Bakkers approach is replaced by the simple force balance,
Eq. (21), for a single bubble. The coefcient o in the transport equation for n
b
is
just a relaxation parameter that stands for the rate at which n
b
responds to the
local turbulence, i.e., adaptseither by coalescence (in the bulk of the stirred
vessel) or by bubble breakup (in the impeller swept domain of the vessel)to
the local equilibrium number density n
bN
The latter variable corresponds to the
local maximum stable bubble size d
bN
whichaccording to Hinze (1955)
depends on the local value of the specic rate of energy dissipation e:
d
b1
C
b1
12
s
r
l
3=5
2=5
(22)
Bakker increased the local values of e as obtained with FLUENT by a con
tribution related to the slip velocity of the bubbles. Working with a single
HARRY E. A. VAN DEN AKKER 204
relaxation parameter o is a phenomenological description avoiding detailed
relations for coalescence and bubble breakup.
The Eulerian gas velocity eld required in both the mass balance and the
above transport equation for n
b
is found by an approximate method: rst, the
complete eld of liquid velocities obtained with FLUENT is adapted downward
because the power draw is smaller under gassed conditions; next, in a very
simple way of oneway coupling, the bubble velocity calculated from the above
force balance is just added to this adapted liquid velocity eld. This procedure
makes a momentum balance for the bubble phase redundant; this saves a lot of
computational effort.
Finally, Bakker and Van den Akker calculated local values for the specic
mass transfer rate k
l
a, by estimating local k
l
values from local values of the
Kolmogorov time scale O(n/e) and by deriving local values of the specic in
terfacial area a from local values for bubble size and bubble holdup.
In spite of all the simplications Bakker and Van den Akker applied and
given the black box approach for the impeller swept domain, their simulations
resulted in values for the bubble size just below the liquid surface, overall hold
up, and average k
l
a values which are in good agreement with their experimental
data (see Table II). The major step forward they made was the acquisition of the
different spatial distributions of average bubble size (see Fig. 13), bubble hold
up and k
l
a as effected by three common impeller types. As a matter of fact, their
approach may be restricted to low values of the gas holdup.
B. VENNEKERS DAWN CODE
Venneker et al. (2002) extended the GHOST! approach due to Bakker and
Van den Akker (1994b) by replacing the single transport equation for the
TABLE II
COMPUTATIONAL AND EXPERIMENTAL RESULTS FOR OVERALL HOLDUP A, SIZE /D
B,OUT
S OF THE BUBBLES
LEAVING THE LIQUID HEEL, AND OVERALL SPECIFIC MASS TRANSFER RATE K
L
A (FROM: BAKKER, 1992)
S. no. a (%) /d
b,out
S (mm) k
l
a (l/s)
Exp. Sim Exp Sim Exp Sim
1 DT 4.770.2 4.9 3.25 2.91 0.038 0.038
2 A315 4.670.2 4.2 3.76 3.59 0.035 0.036
3 A315 4.870.2 4.3 3.82 0.038 0.036
4 PBT 4.170.3 4.1 3.44 3.39 0.036 0.037
5 PBT 1.170.3 1.0 2.00 0.011 0.013
Note: The respective impellers used are a classical Rushton turbine (DT), a hydrofoil impeller
(A315) manufactured by Lightnin, and a Pitched Blade impeller (PBT). The cases 1 through 4 all
relate to a supercial gas rate of 3.6 mm/s only, with impeller speeds varying between 5 and 10/s (gas
ow numbers between 0.01 and 0.02); cases 2 and 3 differ in sparger size and position.
THE DETAILS OF TURBULENT MIXING PROCESS 205
bubble number density including the effective relaxation parameter o by a
number of population balance equations. Actually, these population balance
equations are convectivediffusive transport equations for the bubble number
density in a specic bubble size class and include separate birth and death terms
which take coalescence and breakup into account. In addition, the procedure
for adapting the liquid ow eld to the lower power draw under gassed con
ditions has been improved on the basis of experimental ndings due to Rous ar
and Van den Akker (1994). For the rest, the same approximations are made as
in GHOST! with respect to gas velocities, rate of energy dissipation, and specic
mass transfer rate.
Venneker et al. (2002) used as many as 20 bubble size classes in the bubble
size range from 0.25 to some 20 mm. Just like GHOST!, their inhouse code
named DAWN builds upon a liquidonly velocity eld obtained with FLUENT,
now with an anisotropic Reynolds Stress Model (RSM) for the turbulent
momentum transport. To allow for the drastic increase in computational bur
den associated with using 20 population balance equations, the 3D FLUENT
ow eld is averaged azimuthally into a 2D ow eld (Venneker, 1999, used a
less elegant simplication!)
The agreement between simulation results and experimental data is encour
aging (see Fig. 14), although the simulation gives higher holdup values in the
upper part of the vessel while the overall holdup is lower in the simulation than
FIG. 13. Spatial distributions of bubble size in three vessels agitated by different impellers: a
classical Rushton turbine (DT), a hydrofoil impeller (A315) manufactured by Lightnin, and a
Pitched Blade Impeller (PBT). The gas ow numbers in these simulations are in the range 0.010.02.
These simulation results have been obtained by using GHOST! Reproduced with permission from
Bakker (1992).
HARRY E. A. VAN DEN AKKER 206
in the experiment. The latter discrepancy may be due to the use of optical probes
overlooking bubbles smaller than, say, 1 mm.Venneker et al. (2002) present
some more comparisons between computational and experimental results. One
should realize, however, that for validation purposes hardly any detailed ex
perimental data as to bubble size distributions in stirred vessels are available.
This same shortage of experimental data hampers the assessment of the socalled
MUltipleSIze Group approach MUSIG due to Lo (2000) as incorporated in the
commercial CFD code CFX.
C. FURTHER SIMULATIONS
In comparison with Bakker and Van den Akker (1994b) and Venneker et al.
(2002), Khopkar et al. (2005) applied a more sophisticated twouid approach
including a standard ke turbulence model. Using the incorrect snapshot ap
proach due to Ranade (2002), their simulation results (for gas ow numbers
being 4 times higher than those of Bakker and Van den Akker, 1994b) still
exhibit major discrepancies with respect to experimental data. One of the
alpha (%)
10.0
9.0
8.0
7.0
6.0
5.0
4.0
3.0
2.0
1.0
0.0
FIG. 14. Spatial distribution of the gas holdup in a turbulent stirred vessel lled with a 0.075%
Keltrol solution in water and agitated by a Rushton turbine: (a) experimental data obtained by
means of an optical probe; (b) computational result from DAWN. Overall holdup amounts to some
3.1% in the simulation and 3.7% in the experiment. Reproduced with permission from Venneker
(1999), improved.
THE DETAILS OF TURBULENT MIXING PROCESS 207
striking features is that their liquid velocities even in the outow of the Rushton
impeller are pretty much overestimated in the simulation, may be due to the use
of the snapshot approach (cf. the dicussion on page 180). The spatial distri
butions of gas holdup found in the simulations are compared with experimental
data obtained by means of computed tomography (CT); there is substantial
room for improvement (see Fig. 15) in spite of the much more sophisticated
type of simulation. The paper due to Gentric et al. (2005) exploits several fea
tures and options of the twouid mode of the commercial code STARCD and
illustrates its capabilities in comparing the mixing performance of two industrial
mixing vessels, but does not present validation by means of experimental data.
The combination of twophase ow, turbulence, and a revolving impeller
poses tremendous simulation problems and still requires excessive computer
time or power. While in bubble columns and gas lift loops population balances
are used with some success (see, e.g., Wang et al., 2006), the ow eld in a
stirred tank is so much more complicated and turbulent and dominated by the
revolving impeller that implementing them in stirred vessel simulations still
causes serious convergence problems. Laakkonen et al. (2006) reduced such
problems by restricting his simulation to a multiblock approach subdividing the
stirred vessel into just 23 ideally mixed subregions: this approach actually is a
kind of network of zone approach (see also, e.g., Hristov et al., 2004, who even
used 36,000 zones) extended with population balances and may not be named
CFD indeed. The use of the socalled Quadrature Method of Moments
(QMOM) has been suggested as a proper tool for combining population bal
ances with CFD (Marchisio et al., 2003), but so far has not been used for
simulating gassed stirred tanks.
FIG. 15. Comparison of simulated and experimental gas holdup distribution in a horizontal plane
10 cm above the bottom of the vessel. The gas ow number amounts to 0.084. Reproduced with
permission from Khopkar et al. (2005).
HARRY E. A. VAN DEN AKKER 208
D. A PROMISING PROSPECT
Just like in the context of simulating solids suspension, one may wonder
whether much may be expected from just sticking to the twouid approach
combined with population balances. A better way ahead might rather be to
combine population balances with LES, while proper relations for the various
kernels used for describing coalescence and breakup processes could be deter
mined from DNS of periodic boxes comprising a certain number of bubbles (or
drops). The latter simulations would serve to study the detailed response of
bubbles or drops to the ambient turbulent ow.
An attractive framework for investigating these phenomena is provided by
Derksen (2006b) who carried out DNS of liquidliquid dispersions in a 3D
periodic box. Derksen investigated the response of a turbulent dispersion of
droplets to a history comprising rst a rapidly increasing turbulent activity, then
a quasisteady situation of high turbulence intensity and nally a rapid decay in
turbulence intensity; this history may be equivalent to what a uid package
experiences during its passage through the impeller stream. Again, Derksen
applied a particular LB method for mimicking the two phases. The drop
size distribution and the Sauter mean diameter were tracked in time. Further
more, the presence of the droplets affected the turbulence spectrum because of
the smallscale uid motions induced by the droplets, on the analogy of the
interaction between solid particles and turbulence in the work of Portela and
Oliemans (2003) and Ten Cate et al. (2004).
VIII. Chemical Reactors
So far, most (stirred) chemical reactors have been designed and scaled up by
traditional methods exploiting simple conceptssuch as continuous stirred
tank reactor and residence time distributionand scaleup rules involving
usually a single dimensionless number such as the (mixing) Damko hler number
being the ratio of the turbulent macrotime scale to the characteristic reaction
time scale. In addition, various types of local mixing times and their ratios
are used to characterize or categorize the interaction of mixing and chemical
reactions at scaleup (Patterson et al., 2004).
These methods hardly take spatial distributions of velocity eld and chemical
species or transient phenomena into account, although most chemical reactors
are operated in the turbulent regime and/or a multiphase ow mode. As a result,
yield and selectivity of commercial chemical reactors often deviate from the
values at their laboratory or pilotscale prototypes. Scaleup of many chemical
reactors, in particular the multiphase types, is still surrounded by a fame of
mystery indeed. Another problem relates to the occurrence of thermal runaways
due to hot spots as a result of poor local mixing effects.
THE DETAILS OF TURBULENT MIXING PROCESS 209
Patterson (1985) presented a concise review of the early developments in
computational modeling of secondorder chemical reactions and of more com
plicated and multiple reaction sets which are affected by an intermediate rate of
local turbulent mixing. At that moment in time, closing the crosscorrelation
terms stemming from the turbulent uctuations by means of micromixing
models was still in its infancy. He also just hinted on the use of PDFs. Fur
thermore, the limited computer power of those days kept detailed simulations
and their assessment impossible. Stirred vessels in particular were too difcult a
type of ow devices to allow for application of rigorous CFD techniques, al
though some attempts were made with a very small number of zones or mixing
segments only.
A. MECHANISTIC MICROMIXING MODELS
In the 1980s, Bourne along with a long series of coworkers at ETH Zurich
developed a mechanistic micromixing approach in which lamellar structures
were central. His lamellar structures represent the small ow structures of the
size of the Kolmogorov length scale within which molecular diffusion is the
mechanism bringing the chemical species into the intimate contact required for a
chemical reaction. The best reference might be two papers due to Baldyga and
Bourne (1984a, b).
Such lamellar structures have also been described and modeled by Ranz
(1979) and Ottino (1980) in the context of chemical reactions in laminar ows.
In Bournes micromixing models for chemical reactors operated in the turbulent
ow regime, various assumptions are raised as to the engulfment, the defor
mation, and the lifetime of these lamellar structures which, along with the
diffusion of the reacting species, all affect the yield of the chemical reactions
taking place within these structures. Actually, a CDR equationsee Eq. (13)
is solved explicitly for a chemical species within a single Kolmogorov eddy. The
most appealing models Baldyga and Bourne proposed for the evolution of
such eddies are the Strain (St) Model, the Shear (Sh) Model, the Engulfment
DeformationDiffusion (EDD) model, and then the simpler Engulfment (E)
model. Bourne applied his technique to various sets of competing parallel or
consecutive model reactions each carried out in a fed batch reactor.
In the 1990s, Bakker and Van den Akker (1994, 1996)see also R.A.
Bakkers PhD thesis (1996)continued this mechanistic modeling approach by
attempting a completely deterministic description of the 3D smallscale ow
eld in which the chemical reactions take place at the pace the various species
meet. Starting point is a lamellar structure of layers intermittently containing
the species involved in the reaction. These authors conceived such smallscale
structures as Cylindrical Stretched Vortex (CSV) tubes being strained in the
direction of their axis andas a resultshrinking in size in a plane normal to
HARRY E. A. VAN DEN AKKER 210
their axis. Such CSV tubes showed up around 1990 in several studies exploiting
DNS of turbulent ows in a periodic box. The evolution of a CSV tube during
its lifetime can be found by means of an analytical solution of the vorticity
equation. For the parameters typical of the turbulentow eld in a stirred
reactor, however, the vorticity distribution does not result in substantial wind
ing of material lines during the pertinent short time scales. As a result, the main
effect of the stretching of the CSV is just the exponential shrinking rate of more
or less at, only slightly curling material layers.
Consequently, Bakker (1996) described the concentration evolution of a
single layer subjected to the vorticity eld of a single CSV by means of a one
dimensional differential equation where both the nondimensional time and the
nondimensional spatial coordinate contain the exponential shrinking rate. In
this respect, the CSV approach differs from the various Bourne models in which
the successive generation of several multiplelayer stacks is required and vortex
age is a crucial element.
B. A LAGRANGIAN APPROACH
In addition, Bakker and Van den Akker (1994, 1996) were the rst to track
the path such structures follow in the turbulentow eld of a fed batch reactor
computationally. This is extremely relevant as both vortex age (in Bournes
multiplelayer models) and Kolmogorov length scale strongly depend on the
spatially strongly varying e. Precisely this latter variable exhibits a very inhomo
geneous spatial distribution that only can be estimated by means of CFD. The
idea is that during the microscale process of mixing and reaction the macroow
eld advects the reaction zone throughout the reactor, thereby exposing the
zone to regions of varying e. The ow eld and the spatial edistribution were
obtained via a RANStype of simulation (FLUENT), while the tracking was
done by means of a Discrete Random Walk approach. (It should be kept in
mind that at the time of their simulations LES was not really an option yet!) In
addition to their own CSV model, Bakker and Van den Akker also validated
some of Bournes micromixing models.
Some typical results from their simulations are presented in Fig. 16 in which
the yield X
Q
of the product Q from the slow reaction of a set of two competitive
reactions in a fed batch reactor has been plotted vs. impeller speed for two
micromixing models, viz. their own CSV model and Bournes EDD model; their
simulation results are compared with experimental data from Bourne and Yu
(1991). For the cases shown, the CSV model may perform better than Bournes
EDD model, in particular when A is fed near to the impeller where mixing is
most intense.
An alternative but similar approach (Akiti and Armenante, 2004) is to dene
the reaction zones (or blobs) as a separate phase distinct for the ambient uid
THE DETAILS OF TURBULENT MIXING PROCESS 211
and to track these reaction zones by means of a VolumeofFluid (VOF) tech
nique, which may be conceived being a pseudomultiphase model, originally
designed by Hirt and Nichols (1981). Rather than a ke model, Akiti and
Armenante used a RSM model to reproduce the turbulence characteristics of
the ow eld needed for the tracking procedure.
The most important drawback of using a Lagrangian approach for simu
lating (micro) mixing in chemical reactors is that some model is required
for describing the formation of the Kolmogorovscale ow structures at
some (which?) distance from the mouth of the feed tube. This socalled feed
discretization, aimed at dening the starting conditions (size, number, concen
trations) for the lamellar structures, may have an unknown impact on the
eventual yield.
FIG. 16. The yield X
Q
of the product Q of the slower reaction of a set of two competitive parallel
reactions in a fed batch reactor plotted vs. impeller speed (in /s). The experimental data are due to
Bourne and Yu (1991); the crosses refer to feeding reactant A at the top of the vessel, while the
diamonds refer to feeding more closely to the impeller. The various types of lines refer to simulations
as specied in the legend. Reproduced with permission from R. A. Bakker (1996).
HARRY E. A. VAN DEN AKKER 212
C. A EULERIAN PROBABILISTIC APPROACH
An alternative approach (e.g., Patterson, 1985; Ranade, 2002) is the Eulerian
type of simulation that makes use of a CDR equationsee Eq. (13)for each
of the chemical species involved. While resolution of the turbulent ow down to
the Kolmogorov length scale already is far beyond computational capabilities,
one certainly has to revert to modeling the species transport in liquid systems in
which the Batchelor length scale is smaller than the Kolmogorov length scale by
at least one order of magnitude: see Eq. (14). Hence, both in RANS simulations
and in LES, species concentrations and temperature still uctuate within a
computational cell. Consequently, the description of chemical reactions and the
transport of heat and species in a chemical reactor ask for subtle approaches as
to the SGS uctuations.
In order to obtain a realistic estimate for the reaction rate, the joint distri
bution of the reactants at the smallest turbulent scales is required. Any model
disregarding this joint distribution may lead to an erroneous estimate of the
reaction rate. For instance, ltering the reaction term, on the analogy of the
LES lter for the uid ow, would result in an overprediction of the reaction
rate due to the segregation at the subgrid scales. It may be just due to peculiar
operating or mixing conditions when, such as in the FLUENT simulations re
ported on page 845 of Patterson et al. (2004), CFD simulations ignoring SGS
uctuations result in yield predictions close to experimental data. The value of
the Damkohler number, denoting the ratio of the turbulent macrotime scale to
the characteristic reaction time scale, plays an important role as well.
During the years, quite some proposals have been raised as to closure equa
tions in the CDR equations for the spatial species distributions. These closure
equations relate to the correlation terms in general and to the SGS uctuations
in species concentrations in particular. These proposals substantially differ in
degree of sophistication. Patterson et al. (2004) present several examples of
closure models which were reasonably successful in reproducing a particular set
of experimental data. This, however, does not necessarily say something about
their universal applicability. Bakker and Fasano (1993) applied the socalled
Magnussen model and arrived at reasonable yield predictions for a competitive
consecutive reaction system in a stirred reactor (see also Marshall and Bakker,
2004). This Magnussen model, originally derived for combustion, locally cal
culates several reaction rates as a function of both mean concentrations and
turbulence levels and then selects the lower rate for the source term in the CDR
equation.
Particular attention is to be paid to closure models exploiting various types
of PDFs such as beta, presumed, or full PDFs (e.g., Baldyga, 1994; Fox, 1996,
2003; Ranade, 2002). While PDFs have successfully been exploited for descri
bing chemical reactions in turbulent ames, tubular reactors (Baldyga and
Henczka, 1997), and a TaylorCouette reactor (Marchisio and Barresi, 2003),
they have never been used successfully in stirred reactors so far.
THE DETAILS OF TURBULENT MIXING PROCESS 213
D. A PROMISING PROSPECT
At the end of this review on chemical reactors, special room is reserved for
a very promising approach, although this approach, too, has not yet been
applied for simulating a stirred chemical reactor. Van Vliet et al. (2001,
2005) exploited an elegant probabilistic approach (see also Van Vliet, 2003),
where the PDF methodology incorporates the joint scalar information by solv
ing the transport equation for the full joint scalar PDF (Pope, 1985). In this
way, the secondorder and thus nonlinear reaction terms in the CDR equation
are kept in closed form, making further modeling of the chemical reaction term
redundant.
In order to implement the PDF equations into a LES context, a ltered
version of the PDF equation is required, usually denoted as ltered density
function (FDF). Although the LES ltering operation implies that SGS mode
ling has to be taken into account in order to capture micromixing effects, the
reaction term remains closed in the FDF formulation. Van Vliet et al. (2001)
showed that the sensitivity to the Damkohler number of the yield of competitive
parallel reactions in isotropic homogeneous turbulence is qualitatively well
predicted by FDF/LES. They applied the method for calculating the selectivity
for a set of competing reactions in a tubular reactor at Re 4,000.
Although this LES/FDF methodology is a promising technique, the (current)
drawback is the high computational costs involved to obtain a numerical so
lution of the FDF transport equation. In the above study due to Van Vliet
(2003), the LES uid transport was computed with the help of an LB solver on a
5 10
6
computational grid. Solving a transport equation for the joint PDF of
the chemical species is most effectively done in a Lagrangian MonteCarlo (MC)
manner: the chemical composition of the ow as a function of time and space is
represented by a collection of ctitious particles that are randomly released in
the ow domain and that carry with them the full chemical composition. The
assembly of MC particles is tracked through physical and chemical space by a
set of stochastic ordinary differential equations, where the random term rep
resents diffusion. These equations need closure as to the way the particles in
teract with their direct chemical environment, more specically for the scalar
energy dissipation rate. The model used is the rather common Interaction by
Exchange with the Mean (IEM) model in which a mixing frequency describes
the mixing at SGS.
Van Vliet et al. (2005) tracked 1 10
8
computational nodes to obtain a stoc
hastic solution of their FDF equations. In order to deal with the high com
putational costs, the code was run in parallel on a Linux cluster of 11 dual
AMD Athlon (TM) MP 1800+processors. In this way, about one turbulent
macro time scale (or 8,000 computational steps) per 2 days was computed.
Van Vliet et al. (2004, 2006) investigated the formation of hot spots and
reactor efciency in various geometrical congurations of a tubular reactor
for manufacturing LowDensity Polyethylene (LDPE) by means of the above
HARRY E. A. VAN DEN AKKER 214
LES/FDFapproach. An In situ Adaptive Tabulation (ISAT) technique (due
to Pope) was used to greatly reduce (by a factor of 5) the CPU time needed to
solve the set of stiff differential equations describing the fast LDPE kinetics.
Fig. 17 shows some of the results of interest: the occurrence of hot spots in the
tubular LDPE reactor provided with some feed pipe through which the initiator
(peroxide) is supplied. The 2004simulations were carried out on 34 CPUs
(3 GHz) with 34 GB shared memory, but still required 34 h per macroow time
scale; they served as a demo of the method. The 2006simulations then dem
onstrated the impact of installing mixing promoters and of varying the inlet
temperature of the initiator added.
The above simulations as to the occurrence of hot spots once more illustrate
the power and promises of LES over RANStype simulations. The hot spots can
never be found by means of a RANStype of simulation. The same technique
was used by Van Vliet et al. (2006) to study the inuence of the injector geo
metry and inlet temperature on product quality and process efciency in
the LDPE reactor. On the contrary, the RANSbased simulations due to
R. A. Bakker and Van den Akker (1994, 1996) were pretty much suited to arrive
at yield predictions for a fed batch reactor as a whole.
So far, to the best of our knowledge, the above LES/FDFapproach has not
been applied to stirred chemical reactors in which the turbulentow eld is far
more complex than in a tubular reactor. This LES/FDFapproach, however,
may be the way to go, as it provides highly detailed information on turbulent
reactive ows with the usage of a minimum of modeling assumptions. Although
the high computational demands make LES/FDF simulations currently acces
sible to academic research groups only, the continued exponential growth of
FIG. 17. Three different representations (using increasing temperature thresholds) of hot spots in
a tubular LDPE reactor as found by the LES/FDFmethodology due to Van Vliet et al. (2004).
THE DETAILS OF TURBULENT MIXING PROCESS 215
computer resources will make them a versatile tool for process and geometry
optimization of turbulent reactive ows in the process industries.
IX. Summary and Outlook
The above review has shown that 20 years of developing CFDtechniques has
yielded us substantial simulation capabilities for studying and predicting mixing
under turbulent conditions.
The start in the 1980s was slow and with much trial and error. In the early
1990s, we had to be content with RANSbased simulations onspeaking af
terwardscoarse grids of limited size only. With increasing computer power
and memory becoming available at lower cost, ner and larger grids offered the
potential of getting more detailed pictures, among other things via DNS. LES
entered the mixing scene, although their proliferation suffered from slow con
vergence of the FV solvers. In the late 1990s, however, LB solvers entered the
mixing eld and, owing to being faster and better geared to parallellization,
made LES much more attractive and viable.
A. THE VARIOUS COMPUTATIONAL FLUID DYNAMICS OPTIONS
Nowadays, it is therefore essential to distinguish between the various main
CFD options for dealing with turbulent mixing issues, viz.
LES: with some model of the SGS ow and transport phenomena, suited for
reproducingat the level of the grid cell sizerather detailed transient elds
of velocities and other transport variables in fullscale process equipment
operated under turbulentow conditions.
Commercial CFD software has become a reliable tool for carrying out simu
lations for laminar ows andbased on RANSfor turbulent ows. Practising
engineers gradually have become convinced about the usefulness of RANS
based simulations. This review, however, emphasizes that CFD now has much
more to offer. For practicing engineers confronted with mixing problems, it is
HARRY E. A. VAN DEN AKKER 216
important to realize that CFD is not inherently restricted to just the average
singlephase ow eld, but gradually is becoming more and more capable of
dealing with the details of turbulent eddies and twophase ows. The perform
ance of many physical operations and the yield and selectivity of many chemi
cally reacting systems strongly depend on nonlinear interactions at the small
scales of turbulent ows.
An example: Hollander et al. (2001a) nicely demonstrated how the strong
inhomogeneities in stirredtank ow result in unpredictable scaleup behav
iour and that the impact of the detailed hydrodynamics and of the non
uniform spatial particle distribution on agglomeration rate is larger and
more complex than usually assumed; their study once more illustrated the
risks of scaleup on the basis of keeping a single nondimensional number.
Sophisticated CFD, especially on the basis of LES, offers an attractive
alternative indeed.
Compared to RANS simulations, DNS and LES are much better geared to
reproducing these smallscale processes. RANS simulations focus on the aver
age ow only and by their nature just model the small scales rather rudiment
arily. On the contrary, a DNS resolves all uid motions and a LES resolves
most part of the turbulence spectrum, i.e., all eddies larger than the grid cell size.
While DNS nowadays can be used for turbulent ows at Reynolds numbers up
to say 10,000 in simple geometries (channels, curved tubes) only, LES are quite
feasible for complex geometries, certainly when LB techniques are adopted.
B. THE PROMISES OF DIRECT NUMERICAL SIMULATIONS AND LARGE EDDY
SIMULATIONS
At the moment, DNS and LES for turbulent ows are still the playground
of academic research groups. These groups are making substantial progress,
however, in developing dedicated software forand building up competence
insimulating multiphase ows, transport phenomena, many types of physical
operations, and chemical reactions. Such dedicated software makes it possible
to dig into the details of the mechanisms of a variety of ow and transport
phenomenaoften beyond the current capabilities of experimental techniques.
That is why this review paper is anadmittedly provocativeplea for starting
the exploitation of the advantages of DNS and LES.
An example: rather than linking average bubble size to just or essentially the
(overall) power input of a particular vesselimpeller combination, dedicated
CFD (preferably DNS and LES) allows for studying (tracking) the response
of bubble size to local and spatial variations in the turbulence levels in a
stirred vessel. In this way, the validity of certain modeling assumptions may
be afrmed or disproved. Particularly, effects of spatial variations in e which
THE DETAILS OF TURBULENT MIXING PROCESS 217
remain hidden in the traditional engineering techniques, may surface as a
result of such dedicated CFD approaches. This type of dedicated CFD sim
ulations offer a better and closer look into the details of ow and transport
phenomena than experimental techniques which, e.g., still are not capable of
delivering reliable highresolution evalues.
The advantage of LES over RANSbased simulations is that in the former
approach modeling the effect of the unresolved scales of the ow is easier and
more straightforward, just because the SGS eddies are distinctly separated from
the vessel boundary conditions andas a resulttheir behavior is closer to the
ideal of isotropy rendering universal turbulence modeling feasible. This makes the
outcome of simulations less sensitive to deciencies in turbulence modeling.
Thisalong with the inherently transient character and the degree of detail of the
simulationsturns LES highly suited as a base for simulating physical operations
and chemical reactions carried out in stirred vessels. Whenever the performance
of these processes is strongly dependent on turbulent mixing, the degree to which
CFD simulations can be trusted depends on the ability to reproduce the com
plicated nonlinear interactions of ow and transport phenomena across the vari
ous turbulence scales. LES is then the CFD option to be recommended.
The present author even wonders whether we should not be satised with the
gross predictions of the current RANS methods and turn to LES for the details of
those singlephase and multiphase mixing processes which are dominated by the
spatially distributed turbulence. It is really a valid question how long we should
keep trying and improving the various RANS methods now the increased com
puter power brings the much more sophisticated LES within reach. The very
nature of the RANS approach itselfparticularly the basic assumptions as to
averaging and the various modeling uncertainties as to turbulence and multiphase
owmay really set limits to its exploitation. The modest demands on computer
resources RANSbased simulations require these days are no excuse in this respect.
In addition, DNS of turbulent ow in a periodic box offer interesting oppor
tunities for studying in a fully resolved mode the intimate details of the ow eld,
its interaction with particles and the mutual interaction between particles (in
cluding particleparticle collisions and coalescence). Such simulations may yield
new insights; see, e.g., Ten Cate et al. (2004) and Derksen (2006b). The same can
be said about our understanding of particleturbulence interactions in wall
bounded ows: this has increased due to Portela and Oliemans (2003) exploiting
both DNS and LES and due to Ten Cate et al. (2004).
C. AN OUTLOOK
Nowadays, CFD research at academia is heavily engaged in attempts to in
clude microscale transport phenomena and microscale processes in the dedicated
codes under development with a view to reproduce such divergent processes as
HARRY E. A. VAN DEN AKKER 218
blending, dissolution, crystallization, precipitation, coalescence and redispersion
of bubbles and droplets, suspending solids, and chemical reactions. Essential
physical challenges are in nding proper models for the details of the ow. In
singlephase ows, we need better models for the unresolved contribution of
microscale transport phenomena such as micromixing, while multiphase ow
CFD looks for better models for the mutual interaction of turbulence and dis
persed phase particles and for the interaction force(s) between the dispersed
phase particles and the ambient continuous phase. The devil is in the detail here
fully applies.
In developing multiphase ow CFD and in combining CFD with population
balances and various types of PDF approaches, one needs to keep the size of the
computational job under controlin spite of the overwhelming growth in com
putational power (processor speed, memory, communication tools). This requires
on the one hand efcient and effective numerical tools and on the other hand
clever strategies for handling the enormous amounts of data. Local grid rene
ment techniques may be of great help in avoiding an unnecessary degree of detail.
The development of the above more dedicated LES and DNS is promoted by
the introduction of LB techniques into the world of turbulent mixing simula
tions. LB techniques provide a viable alternative for the more classical FV
solvers of the commercial CFD software, in particular in the context of parallel
simulations on multiple processors. LB techniques are also inherently faster
than FV techniques due to the locality of their operations. In addition, complex
boundaries are easier to implement in the LB approach than with FV solvers.
Substantial improvements in LB techniques have been effectedin terms of
immersed or embedded boundary methods for dealing with moving and curved
boundaries (impeller blades, solid particles) and of grid renement techniques
which have had a positive impact on the fast proliferation of dedicated CFD
tools. Here, too, the details of the computational techniques do matter.
Finally, the large number of processors used in many of the parallel simu
lations cited is striking. It illustrates the enormous progress made in the size of
the simulations academic groups have realized. The falling prices of such pro
cessors and the ease at which these can be combined into platforms for parallel
simulations may have the effect thatjust like in the past decade with RANS
based simulationspretty soon industrial users can afford such dedicated and
detailed simulations, both LES and DNS, and can benet from their outcome in
dealing with their commercial targets.
NOTATION
a specic interface area
a
ij
anisotropy tensor, comprising, essentially, the turbulent stresses
made nondimensional with the turbulent kinetic energy k
THE DETAILS OF TURBULENT MIXING PROCESS 219
A distance to origin in (A
3
, A
2
) plane
A
1
, A
2
, invariants of anisotropy tensor a
ij
A
3
c concentration
c
s
Smagorinsky constant (in SGS modeling)
C
b1
coefcient in relation for local maximum bubble size d
bN
C
D
drag coefcient
C
Dt
drag coefcient in a free stream turbulence
C
m
coefcient in model equation for n
t
(in RANS models)
d
b
bubble size
d
bN
local maximum stable bubble size
D diffusion (or dispersion) coefcient
D
ij
specic rate of production of turbulent stresses
D
k
specic rate of production of turbulent kinetic energy
D
e
specic rate of production of e (Kolmogorov eddies)
g gravitational acceleration constant
I unity tensor
k concentration of turbulent kinetic energy
k
l
mass transfer coefcient
k
sgs
turbulent kinetic energy contained in the SGS eddies
m
0
particle number concentration
n
b
particle number density
n
b1
local equilibrium number density
N impeller speed (number of revolutions per unit of time)
p pressure
~ p pressure as resolved in LES
P average pressure (in RANS context)
P
ij
specic rate of production of turbulent stresses
P
k
specic rate of production of turbulent kinetic energy
Pe specic rate of production of e (Kolmogorov eddies)
q specic heat production rate
r specic rate of chemical reaction producing or consuming a
particular species
^ r radial vector component
~
S local resolved deformation rate (in LES)
S
g
source term in mass balance for gas phase (due to gas supply)
t time
T temperature
U
i
U
j
components of velocity vector v (in sufx notation)
U
k
u
i
u
j
average turbulent, or Reynolds, stresses
u gas velocity vector
U
s
slip (or: relative) velocity vector
v uid velocity vector
HARRY E. A. VAN DEN AKKER 220
~ v uid velocity vector as resolved in LES
V average uid velocity vector (in RANS context)
V
b
bubble volume
V
b,in
bubble volume at position of gas supply
w azimuthal component of velocity vector
x
i
, x
k
spatial coordinate
X
Q
yield of product Q
^ z vertical vector component
GREEK SYMBOLS
a volume fraction of gas
b
0
agglomeration coefcient
d
ij
Kronecker delta
D grid spacing
e specic rate at which turbulent kinetic energy is dissipated (in the
Kolmogorov eddies)
ij
specic rate at which turbulent stresses are dissipated
Z
B
Batchelor length scale (proportional to penetration depth for
diffusion)
Z
K
Kolmogorov length scale (smallest scale in turbulent ow)
k thermal conductivity coefcient
n kinematic viscosity coefcient
n
e
effective SGS viscosity coefcient (in LES)
n
t
turbulent viscosity coefcient (in RANS)
P
ij
specic rate of production of turbulent stresses
P
k
specic rate of production of turbulent kinetic energy
r uid density
r
l
liquid density
s interfacial tension
s shear stress tensor as resolved in LES
s
0
part of s, see Eq. (3)
o effective breakup/agglomeration coefcient (a kind of relaxation
parameter)
o vorticity
O specic rate of destruction of e
DIMENSIONLESS NUMBERS
Re Reynolds number
Sc Schmidt number
THE DETAILS OF TURBULENT MIXING PROCESS 221
ABBREVIATIONS
ASM Algebraic Stress Model
CDR ConvectionDiffusionReaction
CFD Computational Fluid Dynamics
CSV Cylindrical Stretched Vortex
CT Computed Tomography
DNS Direct Numerical Simulation
E Engulfment
EDD EngulfmentDeformationDiffusion
FDF Filtered Density Function
FV Finite Volume
IEM Interaction by Exchange with the Mean
ISAT In situ Adaptive Tabulation
LB LatticeBoltzmann
LDA Laser Doppler Anemometry
LDV Laser Doppler Velocimetry
LDPE Low Density Poly Ethylene
LES Large Eddy Simulation
LGA Lattice Gas Automaton
MC Monte Carlo
MFR Multiple Frames of Reference
NS NavierStokes
PDF Probability Density Function
QMOM Quadrature Method of Moments
RANS Reynolds Averaged Navier Stokes
RSM Reynolds Stress Model
SGS Sub Grid Scale
SDM Sliding and Deforming Mesh
SM Sliding Mesh
Sh Shear
St Strain
TVD Total Variation Diminishing
VOF Volume of Fluid
ACKNOWLEDGEMENTS
First of all, Dr. Jos J. Derksen of the Department of MultiScale Physics at
Delft University of Technology is gratefully acknowledged for a fruitful long
time collaboration and for his critical review of the draft paper. The author is
also indebted to all former PhD students of the Kramers Laboratorium voor
HARRY E. A. VAN DEN AKKER 222
Fysische Technologie of Delft University of Technology for contributing
through their PhD projects and theses to the development of the views and
capabilities described in this chapter.
REFERENCES
Abbott, M. B., and Basco, D. R., Computational Fluid Dynamics: An Introduction for
Engineers. Longman Scientic & Technical, Harlow (UK) (1989).
Akiti, O., and Armenante, P. M. AIChE J 50, 566577 (2004).
Artoli, A. M., Hoekstra, A. G., and Sloot, P. M. A. J. Mod. Phys. B 17(12), 9598 (2003).
Bakker, A., Hydrodynamics of stirred gasliquid dispersions, Ph.D. Thesis, Delft University of
Technology, Delft, Netherlands (1992).
Bakker, R. A., Micromixing in chemical reactors: models, experiments and simulations, Ph.D.
Thesis, Delft University of Technology, Delft, Netherlands (1996).
Bakker, A., and Fasano, J. B., TimeDependent, Turbulent Mixing and Chemical Reaction in
Stirred Tanks, AIChE Symposium. Series No 299 90 7178 (1993).
Bakker, A., LaRoche, R. D., Wang, M. H., and Calabrese, R. V. Chem. Eng. Res. Des. 75A, 4244
(1997).
Bakker, A., Oshinowo, L. M., and Marshall, E. M., The Use of Large Eddy Simulation to Study
Stirred Vessel Hydrodynamics. Proceedings of the 10th European Conference on Mixing,
Delft, Netherlands, 247254 (2000).
Bakker, A., and Oshinowo, L. M. Chem. Eng. Res. Des. 82(A9), 11691178 (2004).
Bakker, A., and Van den Akker, H. E. A., A Computational Study on Dispersing Gas in a Stirred
Reactor. Proceedings of the 7th European Conference on Mixing, Brugue, Belgium
199207. Also in: Fluid mechanics of mixing: modelling, operations and experimental tech
niques, (R. King, Ed.) Fluid Mechanics and its Applications, 10, 3746. Kluwer
Academic Publishers (1991).
Bakker, R. A., and Van den Akker, H. E. A. Chem. Eng. Des. 72A, 733738 (1994).
Bakker, A., and Van den Akker, H. E. A. Chem. Eng. Res. Des. 72A, 583593 (1994a).
Bakker, A., and Van den Akker, H. E. A. Chem. Eng. Res. Des. 72A, 594605 (1994b).
Bakker, R. A., and Van den Akker, H. E. A. Chem. Eng. Sci. 51, 26432648 (1996).
Baldyga, J. Chem. Eng. Sci. 49, 19852003 (1994).
Baldyga, J., and Bourne, J. R. Chem. Eng. Commun. 28, 243258 (1984a).
Baldyga, J., and Bourne, J. R. Chem. Eng. Commun. 28, 259281 (1984b).
Baldyga, J., and Henczka, M., Turbulent mixing and parallel chemical reactions in a pipe: appli
cation of a closure model, Recents Progre`s en Genie des Procedes 11, 341348 (1997).
Bermingham, S. K., Kramer, H. J. M., and Van Rosmalen, G. M. Comp. Chem. Eng. 22, 355362
(1998).
Bourne, J. R., and Yu, S., An Experimental Study of Micromixing Using Two Parallel Reactions.
Proceedings of the 7th European Conference on Mixing, Brugues, Belgium, 6775 (1991).
Bouwmans, I., The blending of liquids in stirred vessels, Ph.D. Thesis, Delft University of Tech
nology, Delft, Netherlands (1992).
Bouwmans, I., Bakker, A., and Van den Akker, H. E. A. Chem. Eng. Res. Des. 75A, 777783 (1997).
Brucato, A., Grisa, F., and Montante, G. Chem. Eng. Sci. 53, 32953314 (1998).
Bujalski, W., Jaworski, Z., and Nienow, A. W. Chem. Eng. Res. Des. 80, 97104 (2002).
Chen, S., and Doolen, G. D. Ann. Rev. Fluid Mech 30, 329364 (1998).
Colenbrander, G. W., Experimental Findings on the ScaleUp Behaviour of the Drop Size Dis
tribution of LiquidLiquid Dispersions in Stirred Vessels. Proceedings of the 10th European
Conference on Mixing, Delft, Netherlands, 173180 (2000).
THE DETAILS OF TURBULENT MIXING PROCESS 223
Curtis, J. S., and Van Wachem, B. AIChE J 50, 26382645 (2004).
Derksen, J. J. Chem. Eng. Res. Des. 79A, 824830 (2001).
Derksen, J. J. Lecture Notes Comput. Sci. 2329, 713722 (2002a).
Derksen, J. J. Flow Turbulence Combustion 69, 333 (2002b).
Derksen, J. J. AIChE J 49, 27002714 (2003).
Derksen, J. J. Chem. Eng. Res. Des. 84(A1), 3846 (2006a).
Derksen, J. J., Multiscale simulations of stirred liquidliquid dispersions. 12th European Con
ference on Mixing, Bologna, Italy, pp. 447454 (2006b).
Derksen, J. J., Doelman, M. S., and Van den Akker, H. E. A. Exp. Fluids 27, 522532 (1999).
Derksen, J. J., Kooman, J. L., and Van den Akker, H. E. A., A parallel DNS implementation for
conned swirling ow, in HPCN Challenges in Telecomp and Telecom: Parallel Simulation
of Complex Systems and LargeScale Applications (L. Dekker, et al., Eds.), pp. 237244.
Elsevier, Amsterdam (1996).
Derksen, J. J., Kooman, J. L., and Van den Akker, H. E. A., Parallel ow simulations by means of
a latticeBoltzmann scheme, In: B. Hertzberger, P. Sloot (Eds.), HighPerformance Com
puting and Networking, Lecture Notes in Computer Science 1225, 524530 (1997).
Derksen, J. J., and Van den Akker, H. E. A., Large eddy simulation of stirred tank ow by means
of a latticeBoltzmann scheme, In: C. R. Kleijn, S. Kawano (Eds.), ASME Proceedings
Volume PVP3772, ASME, 1116 (1998).
Derksen, J. J., and Van den Akker, H. E. A. AIChE J 45, 209221 (1999).
Derksen, J. J., and Van den Akker, H. E. A. AIChE J 46, 13171331 (2000).
Dimotakis, P. E. Annu. Rev. Fluid Mech. 37, 329356 (2005).
Distelhoff, M. F. W., Marquis, A. J., Nouri, J. M., and Whitelaw, J. H. Can. J. Chem. Eng. 75,
641652 (1997).
Ditl, P., and Rieger, F. Chem. Eng. Progr. 102(1), 2230 (2006).
Ducci, A., and Yianneskis, M. Chem. Eng. Sci. 61, 27802790 (2006).
Eggels, J. G. M. Int. J. Heat Fluid Flow 17(3), 307323 (1996).
Eggels, J. G. M., and Somers, J. A. Int. J. Heat Fluid Flow 16(5), 357364 (1995).
Escudie , R., and Line , A. Chem. Eng. Sci. 61, 27712779 (2006).
Fox, R. O. Rev. Inst. Franc . du Petrole 51(2), 215243 (1996).
Fox, R. O., Computational Models for Turbulent reacting Flows. Cambridge University Press,
Cambridge, UK (2003).
Frisch, U., Turbulence, the Legacy of A. N. Kolmogorov. Cambridge University Press,
Cambridge, UK (1995).
Frisch, U., Hasslacher, B., and Pomeau, Y. Phys. Rev. Lett. 56, 15051508 (1986).
Gao, Z., and Min, J. Chinese J. Chem. Eng. 14, 17 (2006).
Gao, Z., Min, J., Smith J. M., and Thorpe, R. B., Large Eddy Simulation of Mixing Time in a
Stirred Tank with Duals Rushton Turbines. 12th European Conference on Mixing,
Bologna, Italy, pp. 431438 (2006).
Gentric, C., Mignon, D., Bousquet, J., and Tanguy, P. A. Chem. Eng. Sci. 60, 22532272 (2005).
Grenville, R. K., Blending of viscous and pseudoplastic uids, Ph.D. Thesis, Craneld Institute
of Technology, Craneld (UK) (1992).
Grenville, R. K., and Nienow, A. W., Blending of miscible liquids, In: NAMF Handbook of
Industrial Mixing (E. L. Paul, V. A. AtiemoObeng, and S. M. Kresta, Eds.), Wiley,
Hoboken (NJ, USA) (2004).
Haam, S., Brodkey, R. S., and Fasano, J. B. Ind. Eng. Chem. Res. 31, 13841391 (1992).
Hartmann, H., Detailed simulations of liquid and liquidsolid mixingturbulent agitated ow and
mass transfer, Ph.D. Thesis, Delft University of Technology, Delft, Netherlands (2005).
Hartmann, H., Derksen, J. J., Montavon, C., Pearson, J., Hamill, I. S., and Van den Akker, H. E. A.
Chem. Eng. Sci. 59, 24192432 (2004a).
Hartmann, H., Derksen, J. J., and Van den Akker, H. E. A. AIChE J 50, 23832393 (2004b).
Hartmann, H., Derksen, J. J., and Van den Akker, H. E. A. Chem. Eng. Sci. 61, 30253032 (2006).
HARRY E. A. VAN DEN AKKER 224
Harvey, P. S., and Greaves, M. Trans. IChemE 60, 201210 (1982).
Harvey, A. D., Lee, C. K., and Rogers, S. E. AIChE J 41, 21772186 (1995).
Harvey, A. D., and Rogers, S. E. AIChE J 42, 27012712 (1996).
Hasal, P., Montes, J. L., Boisson, H. C., and Fort, I. Chem. Eng. Sci. 55, 391401 (2000).
Hinze, J. O. AIChE J 1, 289295 (1955).
Hirt, C. W., and Nichols, B. D. J. Comput. Phys. 39, 201225 (1981).
Hoekstra, A. J., Gas ow eld and collection efciency of cyclone separators, Ph.D. Thesis, Delft
University of Technology, Delft, Netherlands (2000).
Hollander, E. D., Shear induced agglomeration and mixing, Ph.D. Thesis, Delft University of
Technology, Delft, Netherlands (2002).
Hollander, E. D, Derksen, J. J., Bruinsma, O. S. L., Van Rosmalen, G. M., and Van den Akker, H.
E. A., A Numerical Investigation into the Inuence of Mixing on Orthokinetic Agglom
eration. Proceedings of the 10th European Conference on Mixing, Delft, Netherlands,
221230 (2000).
Hollander, E. D., Derksen, J. J., Portela, L. M., and Van den Akker, H. E. A. AIChE J 47,
24252440 (2001a).
Hollander, E. D., Derksen, J. J., Bruinsma, O. S. L., Van den Akker, H. E. A., and Van Rosmalen,
G. M. Chem. Eng. Sci. 56, 25312541 (2001b).
Hollander, E. D., Derksen, J. J., Kramer, H. M. J., Van Rosmalen, G. M., and Van den Akker, H.
E. A. Powder Technol. 130, 169173 (2003).
Holmes, D. B., Voncken, R. M., and Dekker, J. A. Chem. Eng. Sci. 19, 201208 (1964).
Hoogendoorn, C. J., and Den Hartog, A. P. Chem. Eng. Sci. 22, 16891699 (1967).
Hristov, H. V., Mann, R., Lossev, V., and Vlaev, S. D. Trans. IChemE, Food Bioproducts Process
82(C1), 2134 (2004).
Issa, R., and Gosman, A. D., The Computation of ThreeDimensional Turbulent TwoPhase Flow
in Mixer Vessels. Proc. 2nd Int. Conf. Num. Meth. Lam. Turb. Flows, Venice, Italy (1981).
Jahoda, M., Mos te k, M., Kukukova , A., and Machon , V., CFD Modelling of Liquid Homoge
nisation in Stirred Tanks With One and Two Impellers Using Large Eddy Simulation. 12th
European Conference on Mixing, Bologna, Italy, pp. 455462 (2006).
Jaworski, Z., Bujalski, W., Otomo, N., and Nienow, A. W. Chem. Eng. Res. Des. 78, 327333 (2000).
Kandhai, D., Derksen, J. J., and Van den Akker, H. E. A. AIChE J 49, 10601065 (2003).
Khopkar, A. R., Rammohan, A. R., Ranade, V. V., and Dudukovic, M. P. Chem. Eng. Sci. 60,
22152229 (2005).
Khopkar, A. R., Kasat, G. R., Pandit, A. B., and Ranade, V. V. Chem. Eng. Sci. 61, 29212929
(2006).
Kramers, H., Baars, G. M., and Knoll, W. H. Chem. Eng. Sci. 2, 3542 (1953).
Kresta, S. M., and Brodkey, R. S., Turbulence in mixing applications, Ch.2, In: Paul, E. L., Atiemo
Obeng, V. A., Kresta, S. M. (Eds.),NAMF Handbook of Industrial Mixing, Wiley,
Hobken (NJ, USA) (2004).
Laakkonen, M., Alopeus, V., and Aittamaa, J. Chem. Eng. Sci. 61, 218228 (2006).
Lance, M., Marie , J. L., and Bataille, J. J. Fluids Eng. 113, 295300 (1991).
Lane, G. L., Schwarz, M. P., and Evans, G. M., Modelling of the Interaction Between Gas and
Liquid in Stirred Vessels. Proceedings of the 10th European Conference on Mixing, Delft,
Netherlands, 197204 (2000).
Lapin, A., Mu ller, D., and Reuss, M. Ind. Eng. Chem. Res. 43, 46474656 (2004).
Lathouwers, D., Modelling and simulation of turbulent bubbly ow, Ph.D. Thesis, Delft
University of Technology, Delft, Netherlands (1999).
Lathouwers, D., and Van den Akker, H. E. A., A numerical method for the solution of twouid
model equations. Proceedings of the Fluids Eng. Div. 1996 Summer Meeting, San Diego
(CA, USA), ASME, New York (USA), Vol. 1, 121126 (1996).
Lee, K. C., An experimental investigation of the trailing vortex structure and mixing characteristics
of mixing vessels, Ph.D. Thesis, Kings College, London, (UK) (1995).
THE DETAILS OF TURBULENT MIXING PROCESS 225
Liu, S. Chem. Eng. Sci. 54, 873891 (1999).
Lo, S., Application of population balance to CFD modelling of gasliquid reactors. Conference
on Trends in Numerical and Physical Modelling for Industrial Multiphase Flows, Carge` se,
Corse 2729 September (2000).
Lu, Z., Liao, Y., Qian, D., McLaughlin, J. B., Derksen, J. J., and Kontomaris, K. J. Comput.
Physics 181, 675704 (2002).
Lumley, J. Adv. Appl. Mech. 24, 123176 (1978).
Luo, J. Y., Gosman, A. D., Issa, R. I., Middleton, J. C., and Fitzgerald, M. K. Trans. IChemE. 71A,
342344 (1993).
Luo, J. Y., Issa, R. I., and Gosman, A. D. IChemE Symp. Ser. 136, 549556 (1994).
Luo, H., and Svendsen, H. F. AIChE J 42, 12251233 (1996).
Marchisio, D. L., and Barresi, A. A. Chem. Eng. Sci. 58, 35793587 (2003).
Marchisio, D. L., Pikturna, J. T., Fox, R. O., Vigil, R. D., and Barresi, A. A. AIChE J 49, 12661276
(2003).
Marshall, E., and Bakker, A., Computational Fluid Mixing. Fluent Inc. Lebanon, NH; also as
Ch. 5 In: Paul, E. L., AtiemoObeng, V. A., Kresta, S. M. (Eds.), NAMF Handbook of
Industrial Mixing, Wiley, Hoboken (NJ, USA) (2004).
Mason, P. J., and Callen, N. S. J. Fluid Mech. 162, 439462 (1986).
McNamara, G., and Zanetti, G. Phys. Rev. Lett. 61, 23322335 (1988).
Menter, F. R. AIAA J 32(8), 269289 (1994).
Micale, G., Montante, G., Grisa, F., Brucato, A., and Godfrey, J. Chem. Eng. Res. Des. 78,
435444 (2000).
Micale, G., Grisa, F., Rizzuti, L., and Brucato, A. Chem. Eng. Res. Des. 82, 12041213 (2004).
Micheletti, M., Baldi, S., Yeoh, S. L., Ducci, A., Papadakis, G., Lee, K. C., and Yianneskis, M.
Chem. Eng. Res. Des. 82, 11881198 (2004).
Middleton, J. C., Pierce, F., and Lynch, P. M. Chem. Eng. Res. Des. 64, 1822 (1986).
Mittal, R., and Iaccarino, G. Annu. Rev. Fluid Mech. 37, 239261 (2005).
Moin, P., and Kim, J., Sci. Am., January 4652 (1997).
Montante, G., Micale, G., Brucato, A., and Magelli, F., CFD Simulation of Particle Distribution
in a MultipleImpeller HighAspectRatio Stirred Vessel. Proceedings of the 10th European
Conference on Mixing, Delft, Netherlands, 125132 (2000).
Montante, G., and Magelli, F. Chem. Eng. Res. Des. 82, 11791187 (2004).
Montante, G., and Magelli, F. Int. J. Comp. Fluid Dynam. 19, 253262 (2005).
Montante, G., Bakker, A., Paglianti, A., and Magelli, F. Chem. Eng. Sci. 61, 28072814 (2006).
Mumtaz, H. S., Hounslow, M. J., Seaton, M. J., and Paterson, W. R. Chem. Eng. Res. Des. 75,
152159 (1997).
Murthy, J. Y., Mathur, S. R., and Choudhury, D. IChemE Symp.Ser. 136, 341345 (1994).
Myers, K. J., Ward, R. W., and Bakker, A. ASME J. Fluids Eng. 119, 623632 (1997).
Nienow, A. W. Chem. Eng. Sci. 52, 25572565 (1997).
Nikiforaki, L., Montante, G., Lee, K. C., and Yianneskis, M. Chem. Eng. Sci. 58, 29372949 (2002).
Oey, R. S., Mudde, R. F., and Van den Akker, H. E. A. AIChE J 49, 16211636 (2003).
Osman, J. J., and Varley, J. IChemE Symp.Ser. 146, 1522 (1999).
Ottino, J. M. Chem. Eng. Sci. 35, 13771391 (1980).
Patankar, S. V., Numerical Heat Transfer and Fluid Flow. Hemisphere Publishing Corporation,
New York (USA) (1980).
Patterson, G. K., Modelling of turbulent reactors, Ch. 3, In: Mixing of Liquids by Mechanical
Agitation. (J. J. Ulbrecht and G. K. Patterson, Eds.), Gordon and Breach Science Pub
lishers, New York (USA) (1985).
Patterson, G. K., Paul, E. L., Kresta, S. M., and Etchells, A. W. III., Mixing and chemical
reactions, Ch. 13, In: NAMF Handbook of Industrial Mixing (E. L. Paul, V.A. Atiemo
Obeng, and S. M. Kresta, Eds.), Wiley, Hoboken (USA) (2004).
HARRY E. A. VAN DEN AKKER 226
Paul, E. L., AtiemoObeng, V. A., and Kresta, S. M., NAMF Hand book of Industrial Mixing.
Wiley, Hoboken (USA) (2004).
Pericleous, K. A., and Patel, M. K. Physico Chem. Hydrodynam 8, 105123 (1987).
Pinelli, D., Nocentini, M., and Magelli, F. Chem. Eng. Commun. 188, 91107 (2001).
Placek, J., and Tavlarides, L. AIChE J 31, 11131120 (1985).
Placek, J., Tavlarides, L., Smith, G. W., and Fort, I. AIChE J 32, 17711785 (1986).
Pope, S. B. Prog. Energy Combust. Sci. 11, 119192 (1985).
Portela, L. M., and Oliemans, R. V. A. Int. J. Numer. Meth. Fluids 43, 10451065 (2003).
Praturi, A. K., and Brodkey, R. S. J. Fluid Mech. 89, 251272 (1978).
Procha zka, J., and Landau, J. Coll. Czech. Chem. Commun 26, 29612973 (1961).
Ranade, V. V., Computational Flow Modeling for Chemical Reactor Engineering, Volume 5 of
Process Systems Engineering (G. Stephanopoulos and J. Perkins, Eds.), Academic Press, San
Diego (CA, USA) (2002).
Ranade, V. V., Bourne, J. R., and Joshi, J. B. Chem. Eng. Sci. 46, 18831893 (1991).
Ranade, V. V., and Dommeti, S. M. S. Chem. Eng. Res. Des. 74A, 476484 (1996).
Ranade, V. V., Joshi, J. B., and Marathe, A. G. Chem. Eng. Commun. 81, 225248 (1989).
Ranade, V. V., and Van den Akker, H. E. A. Chem. Eng. Sci. 49, 51755192 (1994).
Ranz, W. E. AIChE J 25, 4147 (1979).
Revstedt, J., Fuchs, L., and Tra ga rdh, Ch. Chem. Eng. Sci. 53, 40414053 (1998).
Revstedt, J., Fuchs, L., Kova cs, T., and Tra ga rdh, Ch. AIChE J 46, 23732382 (2000).
Revstedt, J., and Fuchs, L. Chem. Eng. Technol. 25, 443446 (2002).
Rielly, C. D., and Marquis, A. J. Chem. Eng. Sci. 56, 24752493 (2001).
Rietema, K., and Van den Akker, H. E. A. Int. J. Multiphase Flow 9, 2136 (1983).
Rodi, W.,Turbulence models and their application in hydraulicsa state of the art review,
International Association for Hydraulic Research, Delft (NL), reprinted in 1984 (1984).
Rohde, M., Extending the LatticeBoltzmann methodnovel techniques for local grid renement
and boundary conditions, Ph.D. Thesis, Delft University of Technology, Delft, Netherlands
(2004).
Rohde, M., Derksen, J. J., and Van den Akker, H. E. A., Phys. Rev. E 65, Paper No. 056701 (2002).
Rohde, M., Kandhai, D., Derksen, J. J., and Van den Akker, H. E. A., Phys. Rev. E 67, Paper No.
066703 (2003).
Rohde, M., Kandhai, D., Derksen, J. J., and Van den Akker, H. E. A. Int. J. Numer. Meth. Fluids
51(7), 439468 (2006).
Rous ar, I., and Van den Akker, H. E. A., LDA Measurements of Liquid Velocities in Sparged
Agitated Tanks with Single and Multiple Rushton Turbines. 8th European Conference on
Mixing, Cambridge, UK IChemE Symp. Ser., 136, 8996 (1994).
Roussinova, V., Kresta, S. M., and Weetman, R. Chem. Eng. Sci. 58, 22972311 (2003).
Rushton, J. H., Costich, E. W., and Everett, H. J. Chem. Eng. Progr 46, 395404 467476 (1950).
Ruszkowski, S., A Rational Method for Measuring Blending Performance, and Comparison of
Different Impeller Types. Proceedings of the 8th European Conference on Mixing,
Cambridge, UK, pp. 283291 (1994).
Sano, Y., and Usui, H. J. Chem. Eng Japan 18, 4752 (1985).
Scha fer, M., Ho fken, M., and Durst, F. Chem. Eng. Res. Des. 75A, 729736 (1997).
Scha fer, M., Yianneskis, M., Wa chter, P., and Durst, F. AIChE J 44, 12331246 (1998).
Schulze, K., Ritter, J., and Kraume, M., Investigations of Local Drop Size Distributions and Scale
Up in Stirred LiquidLiquid Dispersions. Proceedings of the 10th European Conference on
Mixing, Delft, Netherlands, 181188 (2000).
Seckler, M. M., Bruinsma, O. S. L., and Van Rosmalen, G. M. Chem. Eng. Commun. 135, 113131
(1995).
Shaw, C. T., Using Computational Fluid Dynamics. Prentice Hall International Ltd, Hemel
Hempstead (UK) (1992).
THE DETAILS OF TURBULENT MIXING PROCESS 227
Shyy, W. Adv. Heat Transfer 24, 191275 (1994).
Smagorinsky, J. Monthly Weather Rev. 91, 99164 (1963).
Somers, J. A. Appl. Sci. Res. 51, 127133 (1993).
Sommerfeld, M., and Decker, S. Chem. Eng. Technol. 27(3), 215224 (2004).
Stekelenburg, A. J. C., Van der Hagen, T. H. J. J., and Van den Akker, H. E. A. Int. J. Num. Meth.
Heat Fluid Flow 4, 143158 (1994).
Succi, S., The Lattice Boltzmann Equation for Fluid Dynamics and Beyond. Oxford University
Press, New York (USA) (2001).
Ten Cate, A., Turbulence and particle dynamics in dense crystal slurriesa numerical study by
means of latticeBoltzmann simulations, Ph.D. Thesis, Delft University of Technology,
Delft, Netherlands (2002).
Ten Cate, A., Bermingham, S. K., Derksen, J. J., and Kramer, H. M. J., Compartmental Modeling
of a 1,100 L Crystallizer Based on Large Eddy Flow Simulation. Proceedings of the 10th
European Conference on Mixing, Delft, Netherlands, 255264 (2000).
Ten Cate, A., Derksen, J. J., Kramer, H. J. M., Van Rosmalen, G. M., and Van den Akker, H. E. A.
Chem. Eng. Sci. 56, 24952509 (2001).
Ten Cate, A., Derksen, J. J., Portela, L. M., and Van den Akker, H. E. A. J. Fluid Mech. 519,
233271 (2004).
Tennekes, H., and Lumley, J. L., A First Course in Turbulence. MIT Press, Cambridge (MA,
USA) (1972).
Thornock, J. N., and Smith, P. J. WIT Trans. Built Environ 84, 573583 (2005).
Tra ga rdh, Ch., A Hydrodynamic Model for the Simulation of an Aerated Agitated FedBatch
Fermentor. Proceedings of the 2nd International Conference on Bioreactor Fluid Dynamics,
Cambridge, UK 117134 (1988).
Tsouris, C., and Tavlarides, L. L. AIChE J 40, 395406 (1994).
Van den Akker, H. E. A.,On Status and Merits of Computational Fluid Dynamics. In: Nienow,
A. W. (Ed.), Proceedings of the 4th International Conference on Bioreactor and Bioprocess
Fluid Dynamics, BHR, Edinburgh, UK 407432 (1997).
Van den Akker, H. E. A. ERCOFTAC Bull 36, 3033 (1998).
Van den Akker, H.E.A., Momentum Equations in Dispersed TwoPhase Flows. In:
Cheremisinoff, N.P. (Ed.), Encyclopedia of Fluid Mechanics, Gulf Publishing Company,
Houston (TX, USA), Vol. 3, Chapter 15, 371400 (1986).
Van den Akker, H. E. A., Computational uid dynamics: more than a promise to chemical reaction
engineering. Plenary paper presented at CHISA, Prague, CZ Paper #1270 (2000).
Van Leeuwen, M. L. J., Precipitation and mixing, Ph.D. Thesis, Delft University of Technology,
Delft, Netherlands (1998).
Van Leeuwen, M. L. J., Bruinsma, O. S. L., and Van Rosmalen, G. M. Chem. Eng. Sci. 51,
25952600 (1996).
Van Santen, H., Lathouwers, D., Kleijn, C. R., and Van den Akker, H. E. A., Inuence of
segregation on the efciency of nite volume methods for the incompressible NavierStokes
equations. Proceeding of the Fluids Eng. Div. 1996 Summer Meeting, San Diego (CA,
USA), ASME, New York (USA), Vol. 3, 151158 (1996).
Van Vliet, E., Turbulent reactive mixing in process equipment, Ph.D. Thesis, Delft University of
Technology, Delft, Netherlands (2003).
Van Vliet, E., Derksen, J. J., and Van den Akker, H. E. A., Modelling of Parallel Competitive
Reactions in Isotropic Homogeneous Turbulence Using a Filtered Density Function Ap
proach for Large Eddy Simulations. Proc. PVP01 3rd Int. Symp. on Comput. Techn. for
Fluid/Thermal/Chemical Systems with Industrial Appl., Atlanta, GE, USA (2001).
Van Vliet, E., Derksen, J. J., and Van den Akker, H. E. A., A numerical study of a lowdensity
polyethylene tubular reactor using a 3D FDF/LES approach, AIChE 2004 Annual Mtg.,
Austin, TX, USA (2004).
Van Vliet, E., Derksen, J. J., and Van den Akker, H. E. A. AIChE J 51, 725739 (2005).
HARRY E. A. VAN DEN AKKER 228
Van Vliet, E., Derksen, J. J., and Van den Akker, H. E. A., Numerical Study on the Turbulent
Reacting Flow in the Injector Region of an LDPE Tubular Reactor. Proceedings of the 12th
European Conference on Mixing, Bologna, Italy, pp. 719726 (2006).
Van Wageningen, W. F. C., Kandhai, D., Mudde, R. F., and Van den Akker, H. E. A. AIChE J 50,
16841696 (2004).
Venneker, B. C. H., Turbulent ow and gas dispersion in stirred vessels with pseudo plastic uids,
Ph.D. Thesis, Delft University of Technology, Delft, Netherlands (1999).
Venneker, B. C. H., Derksen, J. J., and Van den Akker, H. E. A. AIChE J 48, 673685 (2002).
Voncken, R. M., Holmes, D. B., and Den Hartog, H. W. Chem. Eng. Sci. 19, 209213 (1964).
Von Smoluchowski, M. Phys. Chem. 92, 129156 (1917).
Vuik, C., Fast iterative solvers for the discretized incompressible NavierStokes equations, Delft
University of Technology, TMI TR9398 (1993).
Wang, T., Wang, J., and Jin, Y. AIChE J 52, 125140 (2006).
Wei, H., and Garside, J. Chem. Eng. Res. Des. 75, 219227 (1997).
Westerterp, K. R., Van Dierendonck, L. L., and De Kraa, J. A. Chem. Eng. Sci. 18, 157176 (1963).
Wilcox, D. C., Turbulence Modelling for CFD. DCW Industries Inc., La Canada (CA) (1993).
Wu, H., and Patterson, G. K. Chem. Eng. Sci. 44, 22072221 (1989).
Yeoh, S. L., Papadakis, G., and Yianneskis, M. Chem. Eng. Res. Des. 82(A7), 834848 (2004a).
Yeoh, S. L., Papadakis, G., Lee, K. C., and Yianneskis, M. Chem. Eng. Technol. 27, 257263
(2004b).
Yeoh, S. L., Papadakis, G., and Yianneskis, M. Chem. Eng. Sci. 60, 22932302 (2005).
Yianneskis, M., Popiolek, Z., and Whitelaw, J. H. J. Fluid Mech. 175, 537555 (1987).
Zwietering, Th. N. Chem. Eng. Sci. 8, 244253 (1958).
THE DETAILS OF TURBULENT MIXING PROCESS 229
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL
REACTORS
Rodney O. Fox
Corresponding author. Tel: +1 515 294 9104; Fax: +1 515 294 2689. Email: rofox@iastate.edu
231
Advances in Chemical Engineering, vol. 31
ISSN 00652377
DOI 10.1016/S00652377(06)310046
Copyright r 2006 by Elsevier Inc.
All rights reserved.
controlled experiments to isolate the chemical kinetics in the absence of transport
effects. The second essential element is detailed knowledge of the transport phe
nomena in chemical reacting systems. Indeed, because commercial chemical
reactors are almost always operated in a regime where mass and energy transport
affect or even control the product distribution from the system, understanding
the coupling between transport processes and chemical reactions is an essential
step in the design and scaleup of chemical reacting systems.
From its beginning, the holy grail of CRE has been a computational model
that is capable of predicting reactor performance based on the fundamental
molecularscale parameters describing the chemical kinetics and the transport
coefcients. In principle, the latter can be measured experimentally in small
scale laboratory experiments (or estimated using computational chemistry). The
chemical reaction engineer then incorporates this information into a compu
tational model to predict the behavior of the plantscale reactor. By avoiding
the need for pilotscale experiments, this experimentfree scaleup approach
should result in more rapid process development at much lower cost. Admit
tedly, while CRE has made considerable progress toward this goal, much work
remains to be accomplished. Nevertheless, due to the tremendous expansion in
computing power over the last 30 years, computational models used in CRE can
now account for much more detail than was previously thought possible. This
trend is unlikely to abate, and thus, to remain relevant to industry, chemical
reaction engineers of the future must become adept at employing detailed ow
models for chemical reacting systems.
For example, over the past 15 years computational uid dynamics (CFD) has
become an important tool in CRE for understanding the coupling between
transport processes and chemical reactions. The denition of CFD has in the
process expanded from its original emphasis on uid dynamics (i.e., momentum
transport) to include mass and energy transport, as well as detailed chemical
reactions. One might argue that it would, therefore, be more accurate to refer to
the eld as computational transport phenomena. On the other hand, because
CFD models rarely resolve all of the relevant scales (as described below), one
might also argue that computational chemical reaction engineering is a more
accurate description. However, both of these names are perhaps too broad and
lose the essential focus on the fact that CFD always includes a description of
momentum transport in the uid phase(s). Thus, the original name continues to
be used to designate all computational approaches that solve for the spatial
distribution of the velocity, concentration, and temperature elds. CFD models
have also been developed for multiphase systems, and commercial CFD codes
now offer a wide range of options for modeling chemical reactors.
Despite the many advances, the users of CFD codes must keep in mind that
the underlying transport equations are based on models, which may or may not
be valid for a particular application. This fact often escapes the minds of
newcomers to the eld who are typically overwhelmed by the numerical issues
associated with convergence, gridindependence, and postprocessing. Even
RODNEY O. FOX 232
many CFD experts tend to avoid the issue by working exclusively in an
application area where acceptably accurate models are already available. Un
fortunately (at least for industrial users!), the application of CFD to chemical
reactor analysis introduces new modeling challenges with each new reactor type.
In the simplest case of laminar ow with fully resolved concentration and tem
perature elds, the models have a strong fundamental basis that typically
involves molecularscale transport coefcients. These CFD models are based on
the socalled microscopic balance equations that are taught to undergraduate
students in chemical engineering courses on transport phenomena. The exten
sion of the microscopic balance equations to multiphase ow systems is also
well understood, but brings with it a wide range of new ow phenomena in even
the simplest cases.
In reality, most of the applications in CRE for which CFD is used cannot be
treated using the microscopic balance equations alone. Instead, CFD models are
introduced to describe the effects of unresolved phenomena on the resolved
quantities. These models introduce phenomenological transport coefcients,
much like the ones developed in CRE models for chemical reactors. In fact, in
many cases, spatial transport is dominated by convection and the molecular
scale spatial transport can be neglected. Nevertheless, just as in classical CRE
models for interphase mass/energy transport, the molecularscale transport
coefcients appear in the dimensionless numbers used to formulate the phen
omenological coefcients. Indeed, because the accuracy of CFD predictions are
strongly dependent on the accuracy of these socalled subgridscale (SGS)
models, the modeling skills developed in CRE over its long history are a crucial
component in the development of CFD for chemical reactor design and anal
ysis. In fact, it would not be pretentious to claim that, due to their considerable
abilities to deal with the coupling between chemical reactions and transport
phenomena, chemical reaction engineers are uniquely qualied to develop the
SGS models needed for CFD modeling of chemical reacting systems.
In the remainder of this work we review the current status of CFD models for
chemical reacting systems. In some cases (e.g., singlephase reacting ow) the
current models are fully predictive in the limits of high and low Reynolds
numbers, and quite accurate in the transition region between these limits. In
other cases (e.g., multiphase reacting ow), the predictive abilities of current
CFD models are, in general, limited. Nevertheless, for particular multiphase
reactors (e.g., gassolid reacting ow), powerful models exist and are making
their way into commercial CFD codes. The goal of the presentation will not be
to describe every model in detail, but rather to indicate the current status of
models for treating reacting ows and to point out areas where further research
is needed. The reader interested in a deeper understanding of the particular
aspects a model will be pointed to the appropriate research literature for further
reading. Moreover, consistent with our desire to use CFD for chemical reactor
analysis and design, we will not discuss models whose primary purpose is to
describe nonreacting ows. For singlephase reactors, excellent descriptions of
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 233
such models are available in the literature. Likewise, CFD models for multi
phase ows are described by other authors in this issue.
The organization of the material is as follows. In Section II we provide a
general introduction to CFD models for chemical reacting systems, and to the
critical modeling issues that arise in their development and application. In
Section III we describe the current state of the art in CFD models for single
phase reacting ows. In Section IV we extend the discussion of singlephase
reacting ows to include systems that produce ne particles that follow the
continuousphase ow. In these systems, the principal novelty is the inclusion of
a population balance model for the particulate phase. In Section V, we describe
the current state of the art in CFD models for multiphase reacting ows. Because
this last area is the least developed, but most rapidly advancing, we will limit our
discussion to CFD models that can potentially be used to describe plantscale
reactors (i.e., multiuid models and related meaneld descriptions). Even for
these models, we will not cover the details on how momentum exchange is
treated between phases. Rather, we will focus our attention on factors that affect
directly the chemical reactions. Finally, in Section VI conclusions are drawn
concerning the current status of CFD models for chemical reactor analysis, and
an attempt is made to point out the research directions where progress can be
expected in the near future.
II. Computational Fluid Dynamics for Reacting Systems
In this Section we give an overview of the formulation of CFD models for
reacting systems, with particular emphasis on systems requiring SGS models.
For the reader to understand the procedure followed to create a CFD model for
a chemical reactor, we cover rst the basic formulation. Then, because the SGS
models are often needed due to the ow being turbulent, we next review the
principal length and time scales present in turbulent transport. We then give
examples of SGS phenomena and their corresponding models in turbulent
reacting ows. Finally, we end the section with a brief discussion of the types of
reactor systems that can currently be treated using CFD.
A. BASIC FORMULATION OF CFD MODELS
When applying CFD to model a chemical reactor, we are interested in
knowing how the basic quantities (density, velocity, concentrations, etc.) vary
with the spatial location in the reactor at a given time instant. The starting point
for developing a CFD model is the microscopic balance equation, described in
detail in standard textbooks on transport phenomena (Bird et al., 2002). Letting
F denote a quantity of interest, the general form of its microscopic balance
RODNEY O. FOX 234
equation is
@F
@t
= UF = J
f
_ _
S
f
(1)
where U is the convective velocity, J
f
is a molecularscale model for the diffu
sive ux, and S
f
is a molecularscale source term. Typical examples of quantities
of interest are uid density r, species mass fractions rY
a
, and the uid
momentum rU. Likewise, for multiphase systems similar quantities are of in
terest, but for each individual phase present in the reactor. The generalized
source term S
f
will then include mass/momentum/heattransfer between phases.
For complex uids (e.g., nonNewtonian ows), molecularscale models for J
f
and S
f
can be quite complicated and can lead to numerical difculties, requiring
specially developed CFD solvers.
As mentioned in Section I.A, CFD codes were originally developed to solve
for the uid momentum for Newtonian uids, for which the righthand side of
Eq. (1) is well understood (Bird et al., 2002). However, even for such uids, it is
not possible to accurately solve the microscopic balance equation for Reynolds
numbers commonly observed in chemical reactors. It is thus necessary to dis
tinguish between directnumerical simulations (DNS) and CFD models using
phenomenological descriptions of the turbulence. The two most widely used
CFD approaches for describing turbulent ow are largeeddy simulations (LES)
and Reynoldsaveraged NavierStokes (RANS) models (Pope, 2000). In both
approaches, it is no longer possible to solve Eq. (1) directly for F due to the
enormous computational cost. Instead, Eq. (1) is ltered (LES) or ensemble
averaged (RANS), yielding
@
~
F
@t
=
~
U
~
F =
uf = J
f
_ _
~
S
f
(2)
This transport equation cannot be solved directly because it involves several
unclosed terms. The SGS ux
uf represents the spatial transport of F by the
unresolved velocity uctuations. Models for this term can generally be written
in the form of a generalized transport equation:
@
~
F
@t
=
~
U
~
F =
~
J
f
~
J
Tf
_ _
~
S
f
~
S
Tf
(3)
where the SGS diffusive ux is denoted by
~
J
Tf
and SGS source term by
~
S
Tf
. To
distinguish this expression from Eq. (1), we will refer to Eq. (3) as the CFD
transport equation. Thus, only in the (rare) case of DNS will the CFD transport
equation correspond to the microscopic balance equation.
In chemical reacting systems, the Reynolds number of the ow is not the only
source of computational challenges. Indeed, even for laminar reacting ows the
chemical source term can be extremely stiff and tightly coupled to the diffusive
transport terms. Averaging, as done above to treat turbulent ows, does not
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 235
alleviate this difculty. Thus, turbulent reacting ows offer many difcult
challenges and require specialized models to describe the coupling between
molecular diffusion and chemical reactions (Fox, 2003). We will look at some of
the more widely applicable models in Section III.
Keeping in mind the discussion leading to Eq. (3), the formulation of a CFD
model for a chemical reactor can be broken down into the following broad
steps:
(i) First we must identify the set of state variables needed to completely de
scribe the reactor. Typical examples are
~
F 2 ~ r; ~ r
~
U; ~ r
~
Y
a
; ~ r
~
T; ~ r
~
f
b
_
where, in addition to the quantities introduced earlier,
~
T is the uid
temperature and
f
b
is a set of scalar quantities. The latter are introduced to
dene, for example, the closure for the chemical source term (Fox, 2003)
and the turbulence model (Pope, 2000). Note that the identication of the
state variables is analogous to what is done in classical CRE models.
Thus, chemical reaction engineers are generally well acquainted with the
methodology needed to complete this step. The only new quantity that does
not appear in lumped CRE models is the uid velocity. However, chem
ical engineers are typically introduced to momentum balances in courses on
transport phenomena, and thus understand its signicance.
(ii) The next and arguably the most difcult step is to nd closures for the CFD
transport equation, expressed in terms of the state variables. For example,
in turbulent ows the diffusiveux terms can often be modeled successfully
as gradientdiffusion terms:
~
J
f
~
J
Tf
D
f
D
Tf
_ _
=
~
F (4)
where D
f
is a moleculardiffusion coefcient and D
Tf
is a turbulentdiffu
sion coefcient. In highReynoldsnumber ows, D
f
is negligible compared to
D
Tf
. Note that in general D
Tf
will depend on the scalar quantities
f
b
ap
pearing in the turbulence model. Closure of the source terms in Eq. (3) is
much more difcult, and requires fundamental knowledge about how the
local ow eld interacts with the quantity of interest (e.g., how the local
turbulence level affects the rates of diffusive mixing and chemical reactions at
the subgrid scale). Nevertheless, the nal closures must be expressed as fol
lows in terms of the state variables:
~
S
f
~ r;
~
U;
~
Y
a
;
~
T;
~
f
b
_ _
and
~
S
Tf
~ r;
~
U;
~
Y
a
;
~
T;
~
f
b
_ _
Note that these closures describe SGS phenomena and hence are essentially
local in space (i.e., interior to a computational grid cell). For this reason, it is
RODNEY O. FOX 236
often possible to use DNS of statistically homogeneous systems (i.e., for which
the [ltered] state variables
~
F do not depend on x) to develop closure models
for
~
S
f
and
~
S
Tf
. This procedure has been widely used in singlephase tur
bulence modeling (Pope, 2000; Fox, 2003), and more recently in multiphase
ow systems (e.g., Bunner and Tryggvason, 2003; Nguyen and Ladd, 2005).
For the latter, the generalized source terms include mass/momentum/heat
transfer between phases, and as discussed in Section V the closure models
involve dimensionless parameters such as the particle Reynolds number.
(iii) The coupled system of CFD transport equations now appears as follows in
closed form:
@
~
F
@t
=
~
U
~
F = D
f
D
Tf
_ _
=
~
F
~
S
f
~
S
Tf
(5)
and the remaining task is to nd a suitable numerical algorithm to solve
them. Fortunately, CFD experts have developed powerful and robust al
gorithms for solving equations in the form of Eq. (5), and these are now
available in commercial CFD codes. Thus, from the perspective of the
chemical reaction engineer working in industry, the efcient application of
CFD to chemical reactor analysis and design will inevitably involve the use
of a commercial CFD code. The next step in the CFD model formulation
will thus be to introduce the closure models developed in the previous step
into the CFD code. This is facilitated in most commercial CFD codes by
the availability of socalled userdened scalars. In many cases, the basic
turbulence and multiphase models will already be available in a commercial
code. The chemical reaction engineer will thus only need to add the
specialized closure models (in terms of
~
f
b
) needed to describe the state
variables in a particular application.
(iv) Once the CFD model equations have been implemented in the code, the
next step is to create a computational grid to capture the specic geometry
of the chemical reactor. The qualities of the grid strongly affect the accu
racy and the speed of convergence of the numerical algorithm. Thus, for
complex reactor geometries, it may make sense to hire a specialist in grid
generator to carry out this step.
(v) The remaining steps involve solving the CFD model and postprocessing of
the results. The latter is greatly facilitated by the builtin functions available
in most commercial CFD codes. It is at this point that reactor analysis and
design actually come into full play. By experimenting with variations in the
operating conditions and reactor geometry, the CFD model can be used to
enhance product selectivity and reactor performance.
When applying the steps outlines above, the prudent engineer will start by
modeling an existing reactor for which plantscale data are available for val
idation of the CFD results. If the agreement is poor, usually this will be due to
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 237
inadequate choices for the state variables and/or closure models. Nevertheless,
one should also examine the computational results to see if there are numerical
errors leading, for example, to inconsistencies in the mass, species, or energy
balances. Getting acceptable agreement may take several iterations of changes
in the closures. For many cases, this process can be facilitated by breaking it
down into independent steps (e.g., oweld predictions can be validated before
adding the chemistry). After reasonable agreement between the model and data
is obtained, the CFD model can be safely used to explore alternative design
scenarios.
B. LENGTH AND TIME SCALES IN TURBULENT FLOWS
As mentioned before in Eq. (3), the most common source of SGS phenomena
is turbulence due to the Reynolds number of the ow. It is thus important to
understand what the principal length and time scales in turbulent ow are, and
how they depend on Reynolds number. In a CFD code, a turbulence model will
provide the local values of the turbulent kinetic energy k and the turbulent
dissipation rate e. These quantities, combined with the kinematic viscosity of the
uid n, dene the length and time scales given in Table I. Moreover, they dene
the local turbulent Reynolds number Re
L
also given in the table.
The integral scale of a turbulent ow characterizes the largest and most
energetic ow structures. In a CFD simulation, the local grid size will be pro
portional to the integral length scale L
u
. Likewise, the characteristic lifetime of
the largest eddies is proportional to the integral time scale t
u
. The Kolmogorov
scale characterizes the smallest ow structures and is resolved by neither LES
nor RANS simulations (only in DNS). Note that the ratios of the length and
time scales are as follows:
L
u
Z
Re
3=4
L
and
t
u
t
Z
Re
1=2
L
(6)
Thus, as the local turbulent Reynolds number increases, the separation between
the scales will increase. As a general rule, Re
L
will be proportional to the
TABLE I
THE PRINCIPAL LENGTH AND TIME SCALES, AND REYNOLDS NUMBERS CHARACTERIZING A TURBULENT FLOW
DEFINED IN TERMS OF THE TURBULENT KINETIC ENERGY k, AND TURBULENT DISSIPATION RATE , AND THE
KINEMATIC VISCOSITY m
Quantity Integral scale Kolmogorov scale
Length L
u
k
3/2
/e Z (n
3
/e)
1/4
Time t
u
k/e t
Z
(n/e)
1/2
Reynolds number Re
L
k
2
/en Re
Z
1
RODNEY O. FOX 238
macroscopic Reynolds number for the ow (i.e., Re dened in terms of a char
acteristic ow velocity and length scale.)
In general, for a xedow geometry, the integral length scale will remain
approximately constant (e.g., in a turbulent jet the integral length scale is
proportional to the jet diameter). Likewise, the integral scale velocity, dened
by L
u
/t
u
, will be proportional to the characteristic velocity of the ow (e.g., the
jet velocity). Thus, as the Reynolds number increases (e.g., to enhance turbulent
mixing), Z, t
u
, and t
Z
will all decrease. In the CFD simulation, the grid will
remain approximately the same and the time step must decrease to follow t
u
.
This implies that at high Reynolds numbers less and less of the smallscale ow
structures are captured by the CFD simulation.
To estimate the amount of turbulent kinetic energy lost when ltering at a
given grid size, it is useful to introduce a normalized model energy spectrum
(Pope, 2000) as follows:
E
u
k Ck
5=3
f
L
k f
Z
k (7)
where k is the dimensionless wavenumber (inverse length), and the Kolmogorov
constant is C 1.61 (based on the most recent DNS (Watanabe and Gotoh,
2004)). The nondimensional cutoff functions are dened by
f
L
k
k
k
2
c
L
1=2
_ _
5=3p
0
(8)
and
f
Z
k exp b k
4
=Re
3
L
c
4
Z
1=4
c
Z
_ _ _ _
(9)
wherein p
0
2 and b 5.2. The parameters c
L
and c
Z
are found by applying
two integral constraints as follows:
1
_
1
0
E
u
k dk (10)
and
Re
L
_
1
0
2k
2
E
u
k dk (11)
Note that the nal form of the energy spectrum depends only on the local
turbulent Reynolds number. As an example, spectra found with different Re
L
are shown in Fig. 1.
In the normalized energy spectrum, k 1 corresponds to the inverse of the
local integral length scale and k Re
3=4
L
to the inverse of the local Kolmogorov
length scale. The range of wavenumbers in Fig. 1 over which the slope is 5/3 is
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 239
called the inertial range. Thus, for the ow to be considered turbulent (as op
posed to transitional ow), Re
L
must be larger than approximately 20. In con
trast, highReynoldsnumber turbulence (i.e., with a signicant inertial range)
does not begin until Re
L
is larger than 450. In RANS turbulence models de
signed for lowReynoldsnumber turbulent ows, the model parameters are
functions of Re
L
, and as the local turbulent Reynolds number approaches zero,
the microscopic balance equation (Eq. 1) is recovered. In contrast, in LES
turbulence models the lter size is typically xed at some Reynoldsnumber
independent wavenumber k
c
410. Thus, the fraction of turbulent kinetic energy
captured by LES can be found from
f
c
_
k
c
0
E
u
k dk (12)
and varies from f
c
1 for small Re
L
up to a constant value less than one for
very large Re
L
(Pope, 2000).
In the discussion above, we have considered only the velocity eld in a tur
bulent ow. What about the length and time scales for turbulent mixing of a
scalar eld? The general answer to this question is discussed in detail in Fox
(2003). Here, we will only consider the simplest case where the scalar eld f is
inert and initially nonpremixed with a scalar integral length scale L
f
that is
approximately equal to L
u
. If we denote the molecular diffusivity of the scalar
by G, we can use the kinematic viscosity to dene a dimensionless number in the
following way:
Sc
n
G
(13)
called the Schmidt number. In gases, typical values of the Schmidt number are
near unity, while in liquids values near 1,000 are quite common. The Schmidt
10
7
10
6
10
5
10
4
10
3
10
2
10
1
10
0
10
1
10
0
10
2
10
4
10
6
10
8
10
10
10
12
E
u
Re
L
= 10
0
Re
L
= 10
2
Re
L
= 10
4
Re
L
= 10
6
Re
L
= 10
8
FIG. 1. The normalized model turbulent energy spectrum for a range of Reynolds numbers.
RODNEY O. FOX 240
number and the Kolmogorov length scale can be used to dene the Batchelor
length scale as follows:
l
B
Sc
1=2
Z Sc
1=2
Re
3=4
L
L
u
(14)
which is the length scale where molecular diffusion occurs. In a nonpremixed
turbulent ow seen under magnication (e.g., using planar laserinduced u
orescence), the smallest observable structures over which concentration gradi
ents are seen have characteristic size l
B
. Note that for large Sc, the Batchelor
scale can be very small even at low Reynolds numbers.
The degree of local mixing in a RANS simulation is measured by the scalar
variance hf
02
i, which ranges from zero for complete mixing (i.e., f hfi is
uniform at the SGS) up to f
max
hfihfi f
min
where hfi is the mean
concentration and f
max
and f
min
are the maximum and minimum values, re
spectively. The rate of local mixing is controlled by the scalar dissipation rate e
f
(Fox, 2003). The scalar time scale analogous to the turbulence integral time
scale is (Fox, 2003) as follows:
t
f
2hf
02
i
f
(15)
In a RANS simulation of scalar mixing, a model for e
f
must be supplied to
compute hf
02
i. In fully developed turbulence, t
f
can be related to t
u
by con
sidering the scalar energy spectrum, as rst done by Corrsin (1964).
To determine how the scalar time scale dened in Eq. (15) is related to the
turbulence integral time scale given in Table I, we can introduce a normalized
model scalar energy spectrum (Fox, 2003) as follows:
E
f
k C
OC
Re
3bk5=4
L
k
bk
f
L
kf
B
k (16)
where the scaling exponent is dened by
bk 1
2
3
7 6f
D
k
_
f
Z
k (17)
and ObukhovCorrsin constant is C
OC
0.670.68 (Sreenivasan, 1996; Watanabe,
and Gotoh, 2004; Yeung et al., 2005). In the model spectrum, the diffusivescale
cutoff function is dened by
f
D
k 1 c
D
Sc
dk=2
k=Re
3=4
L
_ _
exp c
D
Sc
dk=2
k=Re
3=4
L
_ _
(18)
with c
D
2.59, and the diffusive exponent is
dk
1
2
1
4
f
Z
k (19)
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 241
The Batchelorscale cutoff function is dened by
f
B
k 1 c
d
Sc
dk
k=Re
3=4
L
_ _
exp c
d
Sc
dk
k=Re
3=4
L
_ _
(20)
wherein the scalardissipation constant c
d
is found by applying an integral con
straint as follows:
Re
L
Sc
_
1
0
2k
2
E
f
kdk (21)
Note that the scalardissipation constant computed from Eq. (21) depends only on
Re
L
and Sc.
In Fig. 2, the normalized model scalar energy spectrum is plotted for a xed
Reynolds number (Re
L
10
4
) and a range of Schmidt numbers. In Fig. 3, it is
shown for Sc 1000 and a range of Reynolds numbers. The reader interested in
the meaning of the different slopes observed in the scalar spectrum can consult
Fox (2003). By denition, the ratio of the time scales is equal to the area under
the normalized scalar energy spectrum as follows:
t
f
2t
u
_
1
0
E
f
kdk (22)
Thus, from Figs. 2 and 3 we can conclude that the timescale ratio will depend
on Re
L
and Sc.
In the literature on turbulent mixing, the mechanicaltoscalar timescale ratio
is dened by
R
2t
u
t
f
(23)
10
6
10
5
10
4
10
3
10
2
10
1
10
0
10
1
10
1
10
3
10
5
10
7
10
9
E
Sc = 0.05
Sc = 0.5
Sc = 5
Sc = 50
Sc = 500
Sc = 5000
FIG. 2. Normalized model scalar energy spectra for a range of Schmidt numbers and Re
L
10
4
.
RODNEY O. FOX 242
and is plotted based on Eq. (16) for a range of Schmidt numbers as a function of
Reynolds number in Fig. 4. Note that for very large Reynolds numbers, R is
independent of Sc and approaches a constant value of R C/C
OC
2.37, i.e.,
the ratio of the Kolmogorov and the ObukhovCorrsin constants. In contrast,
for Re
L
less than 10
6
, R is strongly dependent on both the Reynolds and
Schmidt numbers. The dependence on Reynolds number is especially signicant
for Schmidt numbers far from unity. For example, liquidphase reactors used
for material processing (Mahajan and Kirwan, 1993, 1996; Johnson and
Prudhomme, 2003a,b) have high Schmidt numbers and operate at low to mode
rate Reynolds numbers (Liu and Fox, 2006). In a CFD simulation, t
u
can be
found from the turbulence model and t
f
from the data in Fig. 4. Thus, the
curves in Fig. 4 dene a SGS model for t
f
parameterized in terms of Re
L
and Sc.
10
7
10
6
10
5
10
4
10
3
10
2
10
1
10
0
Sc = 0.05
Sc = 0.5
Sc = 5
Sc = 50
Sc = 500
Sc = 5000
Re
L
4.0
3.5
3.0
2.5
2.0
1.5
1.0
0.5
0.0
R
FIG. 4. Mechanicaltoscalar timescale ratio found from the normalized model scalar energy
spectra.
10
6
10
8
10
4
10
2
10
0
10
0
10
2
10
4
10
6
10
8
10
10
10
14
10
12
Re
L
= 10
0
Re
L
= 10
2
Re
L
= 10
4
Re
L
= 10
6
Re
L
= 10
8
FIG. 3. Normalized model scalar energy spectra for a range of Reynolds numbers and Sc 10
3
.
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 243
We will revisit this topic in Section III when discussing CFD models for mixing
sensitive reactions. Note that while the discussion above applies to RANS tur
bulence models, the method can be extended to LES by integrating over the SGS
wavenumbers (i.e., starting at k
c
).
In summary, we have seen that the principal length and time scales in a single
phase turbulent ow depend on the local turbulent Reynolds number Re
L
. In a
CFD code, standard turbulence models will provide the local values of k and e.
Given the uid viscosity, it will thus be possible to compute the local turbulent
Reynolds number and related integralscale quantities such as L
u
and t
u
. Using
the model energy spectra, we have also shown how the scalar mixing time t
f
depends on the Schmidt number and local turbulent Reynolds number. In
principle, a similar analysis could be carried out for multiphase turbulent ows
to understand the scaling laws for the length and time scales and their depend
ence on the relevant dimensionless numbers. Unfortunately, DNS of multiphase
ow is still in its infancy and experimental measurements of energy spectra are
difcult to obtain. Nevertheless, we can expect signicant progress in our un
derstanding of turbulent multiphase ows using DNS for particular systems
(e.g., gassolid ows) in the coming years.
C. MODELS FOR SUBGRID SCALE PHENOMENA
As noted earlier, in most applications of CFD to chemical reactor design and
analysis the CFD transport equation (Eq. 3) will require SGS closures. (Here we
use subgrid scale to refer to LES and RANS models for terms not fully resolved
by the computational grid.) Thus, one of the principal tasks of a chemical reaction
engineer is to develop the SGS models that accurately describe the chemistry and
physics occuring at the unresolved scales of the ow. As discussed above, for a
singlephase turbulent ow, the unresolved scales are those smaller than the in
tegral length and time scales (L
u
and t
u
, respectively). Obviously, the SGS models
will be highly dependent on the type of ow under consideration (e.g., singlephase
vs. multiphase, nonreacting vs. reacting, etc.), and a complete listing of all such
models would be lengthy and uninformative to the general reader. Thus, instead of
giving general examples, in this section we will demonstrate how the SGS model is
developed for a particular example (singlephase turbulent mixing). In the sub
sequent sections we will extend this model to reacting scalars of various types.
Readers interested in more details on the models can consult Fox (2003).
One of the rst questions that arises when considering a chemical reactor is
Can the reactor be considered perfectly mixed?. In CRE, this question implies
at least two physical situations:
1. The reactor is perfectly macromixed if the mean concentration at every point
in the reactor is equal to the volume average.
2. The reactor is perfectly micromixed if the instantaneous, local concentration
at every point in the reactor is equal to the local mean concentration.
RODNEY O. FOX 244
The classical CRE model for a perfectly macromixed reactor is the contin
uous stirred tank reactor (CSTR). Thus, to x our ideas, let us consider a stirred
tank with two inlet streams and one outlet stream. The CFD model for this
system would compute the ow eld inside of the stirred tank given the inlet
ow velocities and concentrations, the geometry of the reactor (including bafes
and impellers), and the angular velocity of the stirrer. For liquidphase ow with
uniform density, the CFD model for the ow eld can be developed independ
ently from the mixing model. For simplicity, we will consider this case. Nev
ertheless, the SGS models are easily extendable to ows with variable density.
Following the steps for formulation of a CFD model introduced earlier, we
begin by determining the set of state variables needed to describe the ow.
Because the density is constant and we are only interested in the mixing prop
erties of the ow, we can replace the chemical species and temperature by a
single inert scalar eld x(x, t), known as the mixture fraction (Fox, 2003). If we
take x 0 everywhere in the reactor at time t 0 and set x 1 in the rst inlet
stream, then the value of x(x, t) tells us what fraction of the uid located at
point x at time t originated at the rst inlet stream. If we denote the inlet
volumetric ow rates by q
1
and q
2
, respectively, for the two inlets, at steady state
the volumeaverage mixture fraction in the reactor will be
x
q
1
q
1
q
2
(24)
Thus, the reactor will be perfectly mixed if and only if x
x at every spatial
location in the reactor. As noted earlier, unless we conduct a DNS, we will not
compute the instantaneous mixture fraction in the CFD simulation. Instead, if
we use a RANS model, we will compute the ensemble or Reynoldsaverage
mixture fraction, denoted by hxi. Thus, the rst state variable needed to describe
macromixing in this system is x h i. If the system is perfectly macromixed, x h i
x
at every point in the reactor. The second state variable will be used to describe
the degree of local micromixing, and is the mixturefraction variance hx
02
i.
When the variance is zero, the uid is perfectly micromixed so that x x h i. The
maximum value of the variance at any point in the reactor is x h i1 x h i, and
varies from zero in the feed streams to a maximum of 1/4 when x h i 1=2.
At this point, we should clarify an alltocommon misconception in the
chemicalengineering literature concerning the meaning of Reynolds average.
Unfortunately, many textbooks and journal articles still dene it as a time
average or a volume average over an interval that is not too long, but not too
short. This denition confuses methods for estimating the expected value from
experimental data for a single realization (i.e., time and volume averages), which
are statistics, with the underlying expected values with respect to all possible
realizations. In general, a statistic will be different every time it is computed,
while an expected value is constant at a given point in space and time. Thus,
when deriving closures for expected values such as x h i and hx
02
i, we start with a
general transport theory based on the joint probability density function (PDF)
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 245
as described in Pope (2000) and Fox (2003). Space or time averages only come
into the picture when we must validate the predictions of the CFD model
against experimental data. For example, if the ow is statistically stationary,
then the time average of x(x, t) can be used to estimate x h i(x). (Note that by
denition of statistically stationary the expected values will not depend on
time.) Likewise, if the ow is statistically homogeneous, then the volume average
of x(x, t) can be used to estimate x h i(t). In chemical reactors, the ow is almost
never homogeneous (if it were, CFD would not be needed). Nevertheless, one
still nds authors who confuse micromixing, rigorously dened in terms of the
variance hx
02
i, with deviations of the mean x h i from its volumeaverage.
In reality, such uctuations correspond to poor macromixing and are a
mathematical artifact caused by lumping inhomogeneous ow into a homo
geneous model (e.g., by modeling laminar ow in a tubular reactor using a plug
ow model). Finally, we should note that identical statistical concepts can be
used to derive CFD models for scalar mixing in lowReynoldsnumber chaotic
ows encountered in highviscosity mixing. The principal difculty in these
ows is nding general state variables to describe the length and time scales of
the ow.
The remaining state variables in our CFD model for turbulent mixing are
needed to describe the ow eld in the reactor. In a RANS model for turbulent
ow, the mean velocity U h i appears in the Reynoldsaverage momentum bal
ances. The latter is closed by providing a turbulence model for the Reynolds
stresses (Pope, 2000). If a turbulentviscositybased model is used, two state
variables are introduced to describe the local turbulent integral time scale and
length scale (see Table I). Common choices are the turbulent kinetic energy k,
and the turbulent dissipation rate, e. The set of state variables used to described
turbulent mixing in the reactor are thus
~
F 2 U h i; k; ; x h i; hx
02
i
_
Note that when solving the CFD transport equations, the mean velocity and
turbulence state variables can be found independently from the mixturefraction
state variables. Likewise, when validating the CFD model predictions, the ve
locity and turbulence predictions can be measured in separate experiments (e.g.,
using particleimage velocimetry [PIV]) from the scalar eld (e.g., using planar
laserinduced uorescence [PLIF]).
Now that the state variables have been determined, we can go to steps (ii) and
(iii), which involve nding closed CFD transport equations. The derivation of
the RANS equations is described in detail in Fox (2003), and will not be re
peated here. Instead, we will simply give the CFD transport equations and
discuss the closures appearing in the equations. The ve transport equations are
@r U h i
@t
= r U h i U h i = m m
T
_ _
= U h i =p rg (25)
RODNEY O. FOX 246
@rk
@t
= r U h ik = m m
T
=s
k
_ _
=k P
k
r (26)
@r
@t
= r U h i = m m
T
=s
_ _
=
k
C
1
P
k
C
2
r (27)
@r x h i
@t
= r U h i x h i = DD
T
= x h i (28)
and
@rhx
02
i
@t
= r U h ihx
02
i
_ _
= DD
T
=s
x
=hx
02
i
2D
T
= x h i j j
2
r
x
29
Note that although the density is constant, we have included it in the transport
equations to be consistent with the formulation used in commercial CFD codes.
The lefthand sides of Eqs. (25)(29) have the same form as Eq. (5) and
represent accumulation and convection. The terms on the righthand side can be
divided into spatial transport due to diffusion and source terms. The diffusion
terms have a molecular component (i.e., m and D), and turbulent components.
We should note here that the turbulence models used in Eqs. (26) and (27) do not
contain corrections for low Reynolds numbers and, hence, the moleculardiffu
sion components will be negligible when the model is applied to highReynolds
number ows. The turbulent viscosity is dened using a closure such as
m
T
rC
m
k
2
= (30)
The turbulent diffusivity is dened by introducing a socalled turbulent Schmidt
number Sc
T
:
D
T
m
T
=Sc
T
(31)
which should not be confused with the molecular Schmidt number Sc. Like the
other diffusionmodel constants (i.e., s
k
, s
e
, and s
x
), Sc
T
has been determined
using canonical turbulent mixing experiments (see Pope (2000) and Fox (2003)
for details). We should note, however, that these constants must sometimes
be adjusted for noncanonical ows.
The source terms on the righthand sides of Eqs. (25)(29) are dened as
follows. In the momentum balance, g represents gravity and p is the modied
pressure. The latter is found by forcing the mean velocity eld to be solenoidal
= U h i 0. In the turbulentkineticenergy equation (Eq. 26), P
k
is the source
term due to mean shear and the nal term is dissipation. In the dissipation
equation (Eq. 27), the source terms are closures developed on the basis of the
form of the turbulent energy spectrum (Pope, 2000). Finally, the source terms
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 247
for the mixturefraction variance (Eq. 29) are due to production by mean mix
turefraction gradients and dissipation by micromixing. As written, Eq. (29) is
not yet closed: we need to add a model for the mixturefraction dissipation rate
e
x
. Using Eq. (23), the latter can be modeled by
x
R
k
hx
02
i (32)
where R depends on Re
L
and Sc as shown in Fig. 4. The CFD model for
turbulent mixing is now complete, and can be solved to investigate the degree of
macro and micromixing in a chemical reactor.
The next step in the CFD model formulation involves adding Eqs. (25)(29)
to a CFD code. For this particular example, this step is facilitated in some
commercial CFD codes that have the model already included in the standard
release of the code. The nal step is to solve the model and to compare with
experimental data when available. In this step, it may be useful to dene new
variables during postprocessing to quantify the degree of mixing, mixing zones
or the characteristic times for macro and micromixing. See Liu and Fox (2006)
for examples of how this can be done using output from a CFD mixing model.
The CFD model developed above is an example of a moment closure.
Unfortunately, when applied to reacting scalars such as those considered in
Section III, moment closures for the chemical source term are not usually
accurate (Fox, 2003). An alternative approach that yields the same moments
can be formulated in terms of a presumed PDF method (Fox, 1998). Here we
will consider only the simplest version of a multienvironment micromixing
model. Readers interested in further details on other versions of the model can
consult Wang and Fox (2004).
The basis idea behind multienvironment models is that the mixture fraction
at any location in the reactor can be approximated by a distribution function in
the form of a sum of delta functions as follows:
f
x
z; x; t
N
n1
p
n
x; t d z x
n
x; t (33)
where p
n
is the mass fraction of environment n, and x
n
the mixture fraction in
environment n. Using the denition of mixturefraction moments, we have
x h i
N
n1
p
n
x
n
(34)
and
hx
02
i
N
n1
p
n
x
2
n
x h i
2
(35)
RODNEY O. FOX 248
In other words, if we know p
n
and x
n
at every point in the reactor, then we can
compute up to 2N1 independent mixturefraction moments.
The simplest model of this type is the twoenvironment model (N 2) for
which the independent state variables in the CFD model are
~
F 2 U h i; k; ; p
1
; x
1
; x
2
_
In theory, this model can be used to x up to three moments of the mixture
fraction e:g:; hxi; hx
2
i; and hx
3
i. In practice, we want to choose the CFD
transport equations such that the moments computed from Eqs. (34) and (35)
are exactly the same as those found by solving Eqs. (28) and (29). An elegant
mathematical procedure for forcing the moments to agree is the direct quad
rature method of moments (DQMOM), and is described in detail in Fox (2003).
For the twoenvironment model, the transport equations are
@rp
1
@t
= r U h ip
1
_ _
= DD
T
=p
1
(36)
@rp
1
x
1
@t
= r U h ip
1
x
1
_ _
= DD
T
=p
1
x
1
rgp
1
p
2
x
2
x
1
D
T
x
1
x
2
p
1
=x
1
2
p
2
=x
2
2
_ _
37
and
@rp
2
x
2
@t
= r U h ip
2
x
2
_ _
= DD
T
=p
2
x
2
rgp
1
p
2
x
1
x
2
D
T
x
2
x
1
p
1
=x
1
2
p
2
=x
2
2
_ _
38
where p
2
1p
1
. Summing together Eqs. (37) and (38) and using Eq. (34), the
reader can easily show that Eq. (28) is recovered. With a little algebra (Fox,
2003), one can also show using all three equations and Eq. (34) that Eq. (29) will
be recovered if we let the micromixing parameter be
g
R
2k
(39)
Although we will not do so here, with a little more work one can use Eqs.
(36)(38) to nd the transport equation for hx
3
i. The twoenvironment model
thus provides an extra piece of information that can be compared to exper
imental data.
The next step would be to implement the CFD transport equation for the state
variables in a CFD code. This is a little more difcult for the twoenvironment
model (due to the gradient terms on the righthand sides of Eqs. 37 and 38)
than for the moment closure. Nevertheless, if done correctly both models will
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 249
predict exactly the same values for the mean and variance of the mixture fraction.
(See Wang and Fox (2004) and Liu and Fox (2006) for specic examples.) The
real advantage of the twoenvironment model comes when dealing with reacting
scalars. Unlike the moment method, multienvironment models can easily be
extended to multiple reacting scalars with virtually no changes in the model
formulation and, more importantly and surprisingly, are often nearly as accurate
as more sophisticated closures (Wang and Fox, 2004). We will look at examples
of how this is done in Sections III and IV.
In summary, we have presented two different SGS models for singlephase
turbulent mixing of an inert scalar. The goal of this presentation was not to
show the reader the specic details of how models are derived and tested, but
rather to show how a rather complicated physical process can be modeled by
adding additional scalars to a CFD model in the form of Eq. (5). Once the
equations are in this form, they can be solved in a commercial CFD code for
arbitrarily complex reactor geometries. The primary task faced when developing
a CFD model for a new reacting system is to develop closures in terms of an
appropriate set of state variables. For chemical reaction engineers, the usual
starting point will be an existing CFD model for the uid phase(s), which has
been developed and (hopefully!) validated by experts in uid dynamics. Given
such a model for momentum transport, the chemical reaction engineer can focus
on the signicant task of describing local mass/heattransfer and chemical re
actions. Thus, the availability of accurate models for momentum transport is
the baseline requirement when faced with a new reactor system, and essentially
determines which systems are amenable to CFD.
D. REACTOR SYSTEMS AMENABLE TO CFD
It would be difcult to construct an exhaustive list of reactor systems that can
be treated to some degree using CFD. However, we can give a partial list with a
few examples to illustrate the technical issues. First, the simplest systems to treat
with CFD are laminarow reactors for which the microscopic transport equation
can be solved directly (i.e., no SGS modeling is required.) For such reactors, it is
possible to use detailed chemical kinetics in complex ow geometries involving
heat and masstransfer and catalytic surfaces. Nevertheless, even for laminar
systems computational difculties can arise, for example, when the working with
liquid systems wherein the Schmidt (Prandtl) number is much larger than unity.
For such cases the scalar eld will require a much ner grid than the velocity eld
to fully resolve all of the chemistry and physics (i.e., reactiondiffusion layers).
One might therefore consider using a micromixing model for the scalar elds to
describe the molecular mixing below the grid resolution of the velocity eld. In
principle, such a model would have the same form as those used for turbulent
reacting ows (see Section III), but with a micromixing rate (or local scalar
dissipation rate) found from the local strainrate tensor of the velocity eld.
RODNEY O. FOX 250
Finally, we can also mention that laminarow systems with nonNewtonian
uids often require special numerical algorithms that are usually not available in
CFD codes designed mainly for turbulent ows.
Turbulent singlephase ow reactors can also be treated quite accurately with
current CFD technology. The key issues in this case are the SGS models and the
modeling of heat/mass transfer and reactions at ow boundaries. These issues
arise in the CFD transport equation due to the inability to resolve the smallest
scales of the ow or boundary layers. For turbulent reacting ows, it is now
possible to handle relatively complex chemistry. Nevertheless, due to the com
putational cost, the total number of chemical species that can be transported by a
CFD code for a large computational grid is on the order of 10100. Furthermore,
due to numerical stiffness of many kinetic schemes, simply adding a large number
of scalar equations coupled through the chemical source terms leads to unreal
istically long computing times. It is thus still very much of interest to nd smart
methods for reducing the number of transported scalars needed to describe
complex chemistry. Several useful methods have been proposed to do this (e.g.,
using the quasi steady state for free radicals (Kolhapure et al., 2005), but methods
based on tabulation in terms of a set of progress variables (e.g., Fiorina et al.,
2005) appear to be promising for complex gasphase reaction systems. Difcult
complications arise, however, if the chemical reactor has multiple inlets or re
cycled product streams. Such reactors cannot be classied as either nonpremixed
or premixed (which are the types that can be most easily handled using tabu
lation), and the number of degrees of freedom in scalar phase space is large
enough that it is very difcult to determine a priori an appropriate set of progress
variables to describe the ow. It, thus, may be necessary to carry a large number
of scalar elds to describe such reactors, but one can still use tabulation schemes
(Raman et al., 2004) to handle the numerical stiffness of the chemical source term.
Multiphase reactor systems offer many challenges to CFD modelers. In terms
of complexity, uidsolid systems are more amenable to CFD modeling than
gasliquid systems. Nevertheless, progress has been made in both elds. For
uidsolid systems, we can distinguish (see Section IV) between reactor systems
with ne particles that follow exactly the uid and systems with solidparticle
velocities different than the uids velocity. In the rst case, the solids can be
treated as a scalar eld advected by the (singlephase) uid. In the second case, the
solid phase must have its own momentum equation that is coupled to the mo
mentum equation for the uid. The momentum equation for the solid phase
requires many modeling assumptions to describe all possible ow regimes. From
the point of view of the uid dynamics, in general, dilute uidsolid systems (e.g.,
circulating uidized beds), dominated by uidphase turbulence, are easier to deal
with than dense systems (e.g., bubbling uidized beds). However, the addition of
chemical species and reactions is challenging in both cases. In theory, a general
CFD code for uidsolid ows must account for homogeneous reactions occur
ing in the uid (or solid) phase and heterogeneous reactions occuring at the
interface between phases. Given that the solid phase is very often a complex
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 251
porous material with microchannels that cannot be resolved by the CFD code, it
will be necessary to develop subgridscale models to describe mass/heat transfer
to the particle surface and through the pores. Indeed, in many cases, the solid
phase and surface reactions will be masstransfer limited and the overall con
version predicted by the CFD code will be determined by the model used to
describe mass/heat transfer between the uid and solid phases. Additional com
plications arise for uidsolid systems when the solid particles change in size due,
for example, to surface growth, breakage, or agglomeration. It is then necessary
to include a description of the particle size distribution (Fan et al., 2004). Finally,
we should note that although turbulence models have made considerable progress
for dilute gassolid systems (Minier and Peirano, 2001), the same cannot be said
for dense systems. As a result, CFD simulations of bubbling uidized beds are
usually done without turbulence models (see Section V) and require relatively ne
computational grids to capture integralscale properties of the ow (e.g., total
pressure drop). The high computational cost of such models makes them intrac
table for analyzing plantscale uidizedbed reactors.
From the perspective of CFD, the most difcult reactor systems are gasliquid
and liquidliquid ows. Using the denser phase as the reference phase, such
systems range from dilute (e.g., liquid sprays) to dense (e.g., bubbly ow). From
the point of view of the uid dynamics, these systems are challenging because the
interface between the phases is deformed by (and deforms) the ow. A completely
general CFD model would need to keep track of the uid velocity in each phase
and the location and velocity of the interface. Although it is possible to use this
approach for specic model problems, it is intractable for actual reactor systems
where a lessdetailed approach must be applied. For example, continuum model
can be used that describes gas and liquid as interpenetrating uids. However, it is
then necessary to introduce models for momentum, mass, and energy exchange
between the phases that describe the unresolved processes occuring at the phase
interphase. Unlike in uidsolid ows where the interfaces are rigid, in gasliquid
ows the interface can change due to the uid dynamics and chemical/physical
processes occuring at the interface. Moreover, under industrial conditions where
the volume fraction of the gas phase is often very high, turbulence and bubble
coalescence and breakage must be accounted for in the CFD model (Sanyal et al.,
2005). Unfortunately, it is very difcult to validate (and thus to improve) mul
tiphase turbulence models using modern laserbased measurement techniques.
CFD models for gasliquid chemical reactors remain, therefore, the least devel
oped and should be applied with caution for reactor design and analysis.
III. MixingDependent, SinglePhase Reactions
CFD models for singlephase chemical reactors are by far the most advanced
and widely used in industry. The number of different chemical reactors that can be
RODNEY O. FOX 252
modeled by CFD is very broad and ranges for laminar ow reactors with detailed
gasphase chemistry coupled with catalyticsurface chemistry to complex turbulent
hydrocarbon ames. For many of the more complicated ow congurations, spe
cialized CFD codes have been developed to take advantage of the particular
characteristics of the ow. Thus, we will not attempt to describe the entire range of
ow phenomena that can be predicted using CFD in any detail. For turbulent
reacting ows, a recent monograph (Fox, 2003) covering specic SGS models with
abundant references to the literature is recommended as a starting point for an
yone wanting to know more about the subject. Here, we will conne our attention
to a few specic examples of singlephase turbulent reacting ows to give the
reader a general idea of the key modeling issues in the context of CFD.
The topics in this section are arranged in the order of the computational dif
culty faced when treating the chemical source term. For completeness, we should
note that the simplest case, which simply neglects SGS uctuations of the scalar
elds, will not be discussed for two reasons. First, its implementation in a CFD
code is trivial (at least for cases where it is accurate!) and it is the usual default
model in a commercial CFD code. Second, since it is accurate in the limit where the
reaction rates are slow compared to the ow time scales, CFD is often not required
to understand how to scale up chemical reactors with slow chemistry. The dis
cussion here will thus proceed in the opposite direction: starting with very fast
chemistry and progressing toward socalled niterate chemistry. As discussed in
Fox (2003), the speed of the reactions is taken with respect to the resolved scales of
the uid ow. Thus, in a DNS, the smallest characteristic time scale is the
Kolmogorov time scale t
Z
(see Table I), and a fast reaction occurs on time scales
shorter than t
Z
. In contrast, in RANS simulations the ow time scale is given by the
local integral time scale t
u
. In comparison, for a classical CRE model the ow
time scale is the residence time, which is typically much larger than t
u
. Therefore,
we can conclude that CFD models will start to have utility for reactor analysis
whenever the reaction time scales are smaller than the residence time of the reactor.
Before looking at specic SGS models, we should highlight highly exothermic
chemical reactions (e.g., combustion). The CFD models for these systems are
complicated by the fact that the reaction rates change dramatically across the
ow domain depending on the local temperature. Thus, these systems can be
have as not only nonreacting ows under ambient conditions but also innitely
fast reactions once ignited. For this reason, combustion models for premixed
and nonpremixed systems are usually formulated very differently (Peters, 2000;
Poinsot and Veynante, 2001; Veynante and Vervisch, 2002). In contrast, if we
consider fast, nearly isothermal reactions in the liquid phase the range of
behaviors is more limited in terms of the observed reaction rates. For example,
it does not make sense to discuss CFD models for a premixed acidbase re
action, because once mixed the reaction occurs instantaneously. For this reason,
liquidphase reactions that are sensitive to mixing are almost always operated
under nonpremixed conditions. We will thus limit our attention to these cases in
the following discussion.
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 253
A. ACID BASE AND EQUILIBRIUM CHEMISTRY
Acidbase reactions are the archetypical instantaneous reactions. If we let A
denote the acid concentration and B the base concentration, the chemical source
term for both the acid and base can be expressed as
SA; B kAB (40)
where the rate constant k is extremely large. In essence, acid and base cannot
coexist at the same spatial location so that either A 0 when B40, or B 0
when A40. These zones with excess acid or base are separated by stoichiometric
surfaces whereon A B 0. In a CFD simulation of an acidbase reaction it
makes no sense to try to solve the problem directly using the chemical source
term. Indeed, even if k were only moderately large, including the source term
will lead (at best) to very slow convergence.
To overcome this difculty, we can introduce a new variable x dened in
terms of a linear combination of A and B such that the chemical source term for
x is null. Consider an acidbase reaction of the form
ArB !P (41)
The microscopic transport equations for A and B are, respectively,
@A
@t
U =A = G
A
=AkAB (42)
and
@B
@t
U =B = G
B
=B rkAB (43)
where G denotes the moleculardiffusion coefcient. Note that if we multiply
Eq. (42) by r and then subtract Eq. (43), we can eliminate the chemical source
term as follows:
@ rAB
@t
U =rAB r
2
G
A
rAG
B
B (44)
However, the diffusion term is now more complicated and cannot be closed
unless A and B are known separately.
As discussed for the general case in Fox (2003), to proceed further we must
assume that G
A
EG
B
so that Eq. (44) can be written as
@ rAB
@t
U = rAB Gr
2
rAB (45)
RODNEY O. FOX 254
The applicability of this approximation depends on the relative importance of
the convection and the diffusion terms, and it becomes more accurate for cases
dominated by convection (i.e., at large Reynolds numbers).
As discussed earlier, acidbase reactions are always nonpremixed. For exam
ple, a semibatch reactor could initially be lled with base at concentration B
0
and
acid is added with concentration A
0
. Likewise, a continuous reactor could be run
with two feed streams: one for acid and one for base. For both of these examples,
the degree of mixing between the acid stream and the rest of the reactor contents
can be quantied by introducing the mixture fraction x, which obeys
@x
@t
U =x Gr
2
x (46)
with boundary conditions x 1 in the acid inlet stream and x 0 in the base inlet
stream (and inside the reactor at t 0). For reactors with more than two inlet
streams, it is possible to dene a mixturefraction vector n (Fox, 2003), which
obeys Eq. (46) and has N components with the properties 0 x
n
1 and
S
N
n1
x
n
1. The modeling approaches discussed below for a single mixture frac
tion component x can thus be extended to n to treat more complex ow con
gurations (Fox, 2003).
The mixture fraction as dened above describes turbulent mixing in the re
actor and does not depend on the chemistry. However, by comparing Eqs. (45)
and (46), we can note that they have exactly the same form. Thus, for the
acidbase reaction, the mixture fraction is related to rAB by
x
rAB B
0
rA
0
B
0
(47)
and the stoichiometric mixture fraction is given by
x
st
B
0
rA
0
B
0
(48)
Using the fact that A and B cannot coexist at the same spatial location, we then
nd
A
x x
st
A
0
B
0
=r
_ _
if x4x
st
0 otherwise
_
(49)
and
B
0 if x4x
st
x
st
x rA
0
B
0
otherwise
_
(50)
Thus, the CFD simulation need to only treat the turbulent mixing problem for
the mixture fraction. Once x (or its statistics) are known, the acid and base
concentrations can be found from Eqs. (49) and (50), respectively.
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 255
The acidbase reaction is a simple example of using the mixture fraction to
express the reactant concentrations in the limit where the chemistry is much
faster than the mixing time scales. This idea can be easily generalized to the case
of multiple fast reactions, which is known as the equilibriumchemistry limit. If
we denote the vector of reactant concentrations by / and assume that it obeys a
transport equation of the form
@/
@t
U =/ Gr
2
/ S / (51)
then in the equilibrium limit we need to only consider the solution to a simpler
equation that includes only the chemical source term as follows:
d/
dt
S / (52)
In fact, we are only interested in the value of / for t N subject to initial
conditions that depend on the mixture fraction as follows:
/
0
x x/
1
1 x /
2
(53)
where /
1
is the reactant concentration vector in the rst inlet stream (dened
by x 1) and /
2
is the reactant concentration vector in the second inlet stream.
When formulating the equilibriumchemistry approximation we implicitly
assume that the solution to Eq. (52) for large t depends only on the mixture
fraction, and not on the mixing history of the uid element. For some mixing
sensitive reactions (see Section III.B below), this assumption does not hold and
the equilibriumchemistry approximation is not applicable. These reactions are
typically irreversible and the nal product distribution depends on the mixing
path in concentration phase space traversed by the uid particle. In general, the
equilibriumchemistry approximation should only be used for systems of fast
reversible reactions. For this case, /
N
(x) found from solving Eq. (52) will de
pend only on x. Note that adding a reverse reaction to the acidbase reaction
(Eq. 41) discussed above will not change the basic conclusion that A and B can
be determined from x. Only the nal formulae (Eqs. 49 and 50) will change, and
these can be found using the methods described in Fox (2003).
In a turbulent ow, the local value (i.e., at a point in space) of the mixture
fraction will behave as a random variable. If we denote the probability density
function (PDF) of x by f
x
(z) where 0rzr1, the integer moments of the mixture
fraction can be found by integration:
hx
n
i
_
1
0
z
n
f
x
z dz (54)
In most applications, the moments of principal interest are the mean hxi and
variance hx
02
i hx
2
i hxi
2
. The most widely used approach for approximating
RODNEY O. FOX 256
f
x
is the presumed PDF method wherein the PDF depends only on a small set of
moments. For example, the beta PDF can be used and has the functional form
f
x
z
a b 1 !
a 1 ! b 1 !
z
a1
1 z
b1
(55)
where a and b depend on the mean and variance as follows:
a x h i
x h i 1 x h i
hx
02
i
1
_ _
and b
1 x h i
x h i
a (56)
The spatial dependencies of x h i and hx
02
i are found by solving Eqs. (28) and (29),
respectively.
A typical CFD model for acidbase and equilibrium chemistry solves Eqs.
(25)(29), and then uses Eq. (55) to approximate f
x
. Once f
x
is known, the
expected values of the reactant concentrations are computed by numerical
quadrature from the formula
h/i
_
1
0
/
1
z f
x
z dz (57)
For example, A h i and B h i can be computed using Eqs. (49) and (50), respectively.
Note that instead of Eq. (55), we could use the simpler expression for f
x
given by
Eq. (33), which avoids the need for numerical quadrature. In both cases, the mean
and variance of the mixture fraction are identical (and thus both models account
for niterate mixing effects.) In practical applications, the differences in the
predicted values of h/i can often be small (Wang and Fox, 2004).
B. CONSECUTIVECOMPETITIVE AND PARALLEL REACTIONS
To go beyond the equilibriumchemistry limit to consider cases where some of
the reaction rates are nite compared to the ow time scales, we need an efcient
method to solve for chemical species with chemicalsource terms. The straight
forward approach for doing this is to simply solve a transport equation for each
chemical species with its corresponding chemicalsource term. However, it is often
the case that one or more of the reaction steps is very fast compared to the ow
time scales, leading to numerical difculties or poor convergence. An elegant
method for avoiding this problem is to rewrite the transport equations in terms of
the mixture fraction and a set of reactionprogress variables (Fox, 2003).
Some typical examples of the reactions that can be treated in this manner are
consecutivecompetitive reactions:
AB!
k
1
R
AR!
k
2
S 58
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 257
and parallel reactions as follows:
AB!
k
1
P
1
AC!
k
2
P
2
59
In most applications, the rst reaction in each set is an acidbase reaction so
that k
1
is very large. For Eq. (59), B and C are premixed and added to A under
conditions such that B is in stoichiometric excess to A. Likewise, for Eq. (58), B
is reacted in stoichiometric excess with A to produce the desired product R.
Under these conditions, the rst reaction in each set is favored. However, if
mixing occurs with the same time scale as the second reaction, the undesired by
product (S in Eq. (58) and P
2
in Eq. (59)) will be produced. Thus, the amount of
byproduct produced is a sensitive measure of the quality of mixing in the
chemical reactor.
The description of Eqs. (58) and (59) in terms of the mixture fraction and
reactionprogress variables is described in detail by Fox (2003). Here we will
consider a variation of Eq. (59) wherein the acid acts as a catalyst in the second
reaction (Baldyga et al., 1998):
AB!
k
1
P
1
AC!
k
2
AP
2
60
This parallel reaction set was used, for example, by Johnson and Prudhomme
(2003a) to investigate the quality of mixing in a conned impingingjets reactor.
Following the steps outline in Fox (2003), the reactant concentrations in Eq.
(60) can be written in terms of the mixture fraction x and two reactionprogress
variables Y
1
and Y
2
as
c
A
A
0
1 x 1 x
s1
Y
1
(61)
c
B
B
0
x x
s1
Y
1
(62)
and
c
C
C
0
x x
s2
Y
2
(63)
where the two stoichiometric mixture fractions are
x
s1
A
0
A
0
B
0
and x
s2
A
0
A
0
C
0
(64)
and A
0
, B
0
, and C
0
are the inlet concentrations of reactants A, B, and C,
respectively. Note that in the absence of chemical reactions, the reaction
progress variables are dened such that Y
1
Y
2
0.
RODNEY O. FOX 258
The microscopic transport equations for the reactionprogress variables can
be found from the chemical species transport equations by generalizing the
procedure used above for the acidbase reactions (Fox, 2003). If we assume that
G
A
EG
B
EG
C
, then the transport equations are given by
@Y
a
@t
U =Y
a
Gr
2
Y
a
S
a
x; Y
1
; Y
2
for a 1; 2 (65)
where the chemicalsource terms are
S
1
x; Y
1
; Y
2
k
1
B
0
x
s1
c
A
c
B
B
0
x
s1
k
1
1 x
1 x
s1
Y
1
_ _
x
x
s1
Y
1
_ _
(66)
and
S
2
x; Y
1
; Y
2
k
2
C
0
x
s2
c
A
c
C
B
0
x
s1
k
2
1 x
1 x
s1
Y
1
_ _
x
x
s2
Y
2
_ _
(67)
Note that since the reaction rates must always be nonnegative, the chemically
accessible values of the reactionprogress variables will depend on the value of
the mixture fraction. We will discuss this point further by looking next at the
limiting case where the rate constant k
1
is very large and k
2
is nite.
In many applications, due to the large value of k
1
, the rst reaction is es
sentially instantaneous compared to the characteristic ow time scales. Thus, if
the transport equation is used to solve for Y
1
, the chemicalsource term S
1
will
make the CFD code converge slowly. To avoid this problem, Y
1
can be written
in terms of x by setting the corresponding reactionrate expression (S
1
) equal to
zero as follows:
Y
11
min
x
x
s1
;
1 x
1 x
s1
_ _
(68)
It is then no longer necessary to solve a transport equation for Y
1
and the
numerical difculties associated with treating the rst reaction with a niterate
chemistry solver are thereby avoided.
As with the acidbase reaction, Eq. (68) implies that A and B cannot coexist
at any point in the ow. Using this inniterate approximation, we need only
solve transport equations for x and Y
2
, where the source term for Y
2
is now
S
21
x; Y
2
B
0
x
s1
k
2
1 x
1 x
s1
Y
11
_ _
x
x
s2
Y
2
_ _
(69)
Note that S
2N
must be nonnegative, and thus the expression above only holds
for x and Y
2
values that satisfy this condition. For all other values, S
2N
is null.
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 259
Applying Eq. (68), we nd that when S
2N
is nonzero, it equals
S
21
x; Y
2
A
0
k
2
1
x
x
s1
_ _
x
x
s2
Y
2
_ _
if 0 x x
s1
and 0 Y
2
x=x
s2
70
The region in xY
2
composition space where this chemical source term is non
zero is shown in Fig. 5. Note that the maximum conversion of C occurs when
x x
s1
and corresponds to Y
2max
x
sl
/x
s2
or (using Eq. 63) to c
C
0 (i.e.,
complete conversion).
As mentioned earlier, in applications of Eq. (60) the reactor is operated with
excess B so that x
s1
o1/2. If the mixing in the reactor is good, the mixture
fraction in all uid particles at the exit of the reactor will be equal to the mean
x x h i41=2. Thus, if mixing were much faster than the characteristic reaction
time of Eq. (70) ((A
0
k
2
)
1
), then the chemicalsource term in Eq. (70) would be
zero and no reaction would occur so that Y
2
0 at the exit. In contrast, any
local deviations from perfect mixing can lead to zones in the reactor where
xrx
s1
and hence to the production of the byproduct Y
2
. In the opposite limit
where k
2
is large compared to the mixing rates, the maximum attainable value
for Y
2
when x
s1
rxr1 is the mixing line (Fox, 2003), dened by
Y
2mix
x Y
2 max
1 x
1 x
s1
_ _
for x
s1
x 1 (71)
0 1
0
nonzero
source term
'mixing' line
s1
Inaccessible region
Y
2max
Y
2
FIG. 5. Region in xY
2
phase space with nonzero chemical source term and the mixing line.
RODNEY O. FOX 260
and shown as a dashed line in Fig. 5. Using this expression, we nd that the
maximum attainable conversion is
X
max
x
s1
1
x
x1 x
s1
for x
s1
x 1 (72)
The accessible region in xY
2
phase space for the reaction given in Eq. (60) is
represented by the triangular region in Fig. 5 found by connecting the feed
points and the maximum conversion point. Phasespace trajectories begin at the
two feed streams [stream 1: (0; 0) and stream 2: (1; 0)]. If x
s1
is less than the
outlet value of the mixture fraction, then the amount of byproduct formed is
determined by the amount of time spent in the region with nonzero source term
(t
mix
) and the characteristic time of the second reaction (t
r
). If t
r
is large com
pared to t
mix
, then the byproduct concentration will be near zero. If the inverse
is true, then the byproduct concentration will be near X
max
.
The CFD model for the reaction given in Eq. (60) in the limit where the rst
reaction is very fast must account for uctuations in x and Y
2
due to turbulent
mixing. In general, this is done by solving for their joint PDF (Fox, 2003),
denoted here by f (z, y). There are several ways this can be accomplished:
1. Solve the joint PDF transport equation.
2. Assume a functional form for the joint PDF.
3. Assume a functional form for the conditional PDF of Y
2
given x and use a
presumed PDF for f
x
.
Method 1 will be discussed in Section III.C. Method 3 can be implemented in
several different forms (Baldyga, 1994; Klimenko and Bilger, 1999; Fox and
Raman, 2004), but the lowest order approximation requires a model for the
conditional expected value of Y
2
given that x z (denoted by hY
2
jzi) where
hY
2
j0i hY
2
j1i 0. By denition, hY
2
jzi will be a singlevalued function of
mixture fraction and will lie in the triangular region in Fig. 5. The simplest such
model is the one proposed by Baldyga (1994), which uses a linearinterpolation
procedure to nd the conditional moment from the unconditional moment hY
2
i
(Fox, 2003). Here we will look at a multienvironment model that is based on
method 2.
The multienvironment model for the joint PDF generalizes Eq. (33) by
writing
f z; y; x; t
N
n1
p
n
x; t dz x
n
x; t dy Y
2n
x; t (73)
where Y
2n
is the value of Y
2
corresponding to environment n. Here we will
consider only the twoenvironment model (N 2) where the CFD models for
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 261
p
1
, x
1
and x
2
are given by Eqs. (36), (37), and (38), respectively. Similarly, the
CFD models for the reactionprogress variable in the two environments are
@rp
1
Y
21
@t
= r U h ip
1
Y
21
_ _
= DD
T
=p
1
Y
21
rp
1
S
21
x
1
; Y
21
rgp
1
p
2
Y
22
Y
21
D
T
Y
21
Y
22
p
1
j=Y
21
j
2
p
2
j=Y
22
j
2
_ _
74
and
@rp
2
Y
22
@t
= r U h ip
2
Y
22
_ _
= DD
T
=p
2
Y
22
rp
2
S
21
x
2
; Y
22
rgp
1
p
2
Y
21
Y
22
D
T
Y
22
Y
21
p
1
j=Y
21
j
2
p
2
j=Y
22
j
2
_ _
75
Except for the chemical source term, these equations have the same form as
those used for the mixture fraction. Note that the chemical source term (S
2N
) is
evaluated using the mixture fraction and reactionprogress variable in the par
ticular environment. The average chemical source term hS
21
x; Y
2
i will thus
not be equal to S
21
hxi; hY
2
i unless micromixing occurs much faster than the
second reaction.
The CFD model for the reaction given in Eq. (60) with k
1
N has state
variables
~
F 2 U h i; k; ; p
1
; x
1
; x
2
; Y
21
; Y
22
_
By denition of the reactionprogress variables, Y
21
and Y
22
are zero for the
inlet streams, and nonnegative inside the reactor due to the chemical source
term. Once the CFD model has been solved, the reactant concentrations in each
environment n are found from
c
An
A
0
1 x
n
1 x
s1
Y
1n
(76)
c
Bn
B
0
x
n
x
s1
Y
1n
(77)
and
c
Cn
C
0
x
n
x
s2
Y
2n
(78)
where
Y
1n
min
x
n
x
s1
;
1 x
n
1 x
s1
_ _
(79)
The mean reactant concentrations are then dened by (p
2
1p
1
)
c
A
h i p
1
c
A1
p
2
c
A2
(80)
c
B
h i p
1
c
B1
p
2
c
B2
(81)
RODNEY O. FOX 262
and
c
C
h i p
1
c
C1
p
2
c
C2
(82)
The overall conversion of C, denoted by X, is computed using
X 1
c
C
h i
C
0
x h i
(83)
The value of X at the reactor outlet is a sensitive measure of the degree of mixing
in the reactor. If X51, then mixing in the reactor is rapid compared to the
second reaction in Eq. (60). In contrast, if XE1, then mixing is slow.
The CFD model described above has been used by Liu and Fox (2006) to
simulate the experiments of Johnson and Prudhomme (2003a) in a conned
impingingjets reactor. In these experiments, two coaxial impinging jets with
equal ow rates are used to introduce the two reactantstreams. The jet Re
ynolds number Re
j
determines the uid dynamics in the reactor. Typical CFD
results are shown in Fig. 69 for a jet Reynolds number of Re
j
400 and a
reaction time of t
r
4.8 msec. The latter is controlled by xing the inlet con
centrations of the reactants. Further, details on the reactor geometry and the
CFD model can be found in Liu and Fox (2006).
X(mm)
Z
(
m
m
)
2 1 0 1 2
2
1
0
1
2
3
4
5
1.0
0.9
0.8
0.7
0.6
0.5
0.5
0.4
0.3
0.2
0.1
0.0
p
1
X(mm)
Z
(
m
m
)
2 1 0 1 2
2
1
0
1
2
3
4
5
1.0
0.9
0.8
0.7
0.6
0.5
0.5
0.4
0.3
0.2
0.1
0.0
p
2
FIG. 6. Volume fractions p
1
and p
2
in the crosssection of the conned impingingjets reactor.
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 263
X(mm)
Z
(
m
m
)
2 1 0 1 2
2
1
0
1
2
3
4
5
1.00
0.91
0.82
0.73
0.64
0.55
0.45
0.36
0.27
0.18
0.09
0.00
1
X(mm)
Z
(
m
m
)
2 1 0 1 2
2
1
0
1
2
3
4
5
1.00
0.91
0.82
0.73
0.64
0.55
0.45
0.36
0.27
0.18
0.09
0.00
2
FIG. 7. Mixture fractions x
1
and x
2
in the crosssection of the conned impingingjets reactor.
X(mm)
Z
(
m
m
)
2 1 0 1 2
2
1
0
1
2
3
4
5
0.67
0.61
0.55
0.49
0.43
0.37
0.30
0.24
0.18
0.12
0.06
0.00
Y
21
X(mm)
Z
(
m
m
)
2 1 0 1 2
2
1
0
1
2
3
4
5
0.67
0.61
0.55
0.49
0.43
0.37
0.30
0.24
0.18
0.12
0.06
0.00
Y
22
FIG. 8. Reactionprogress variables Y
21
and Y
22
in the crosssection of the conned impingingjets
reactor.
RODNEY O. FOX 264
In Fig. 6, the volume fractions for each environment p
1
and p
2
are shown in a
crosssection of the reactor, which includes the inlet jets and the outlet tube.
Note that because p
2
1p
1
and the inlet ow rates are equal, the contour plots
are symmetric about the vertical axis. For the same reason, p
1
p
2
1/2 on the
vertical axis. In the lefthand inlet stream p
1
1, corresponding to reactant A.
In the righthand inlet stream p
2
1, corresponding to reactants B and C. On
the outlet crosssection mixing is nearly complete so that p
1
Ep
2
. Finally, note
that since Eq. (36) does not contain a term for micromixing, the distribution of
p
1
and its deviation from 1/2 measures the degree of macromixing at any point
in the reactor. We can therefore conclude the reactor is fairly well macromixed
everywhere except in the region near the inlet jets.
In Fig. 7, the mixture fractions in each environment x
1
and x
2
are shown. By
denition of the inlet conditions, in the inlet tubes x
1
0 and x
2
1. The var
iations away from the inlet values represent the effect of micromixing. For ex
ample, if we set g 0 in Eqs. (36) and (37) to eliminate micromixing, then x
1
and
x
2
would remain at their inlet values at all points in the reactor. Note that the
spatial distributions of x
1
and x
2
are antisymmetric with respect to the vertical
axis (as would be expected from the initial conditions.) In the outlet tube, x
1
and
x
2
are very near the perfectly micromixed value of 1/2. Finally, by comparing
Fig. 6 and Fig. 7, we can observe that macromixing occurs slightly faster than
micromixing in this reactor (i.e., p
n
are closer to their outlet values than are x
n
.)
X(mm)
Z
(
m
m
)
2 1 0 1 2
2
1
0
1
2
3
4
5
619
563
506
450
394
338
281
225
169
113
56
0
c
C1
X(mm)
Z
(
m
m
)
2 1 0 1 2
2
1
0
1
2
3
4
5
619
563
506
450
394
338
281
225
169
113
56
0
c
C2
FIG. 9. Reactant concentrations c
C1
and c
C2
in the crosssection of the conned impingingjets
reactor.
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 265
The results shown in Figs. 6 and 7 can be combined to compute the mean
mixture fraction hxi and its variance hx
02
i from Eqs. (34) and (35), respectively.
Example plots are shown in Liu and Fox (2006) and, as expected, they agree
with the solution found by solving the moment transport equations directly
(Eqs. 28 and 29).
In Fig. 8, the reactionprogress variables in each environment Y
21
and Y
22
are
shown. By denition of the inlet conditions, in the inlet tubes Y
21
Y
22
0.
Recall that Y
2
is produced by the second (niterate) reaction in Eq. (60). Thus,
as observed in the plots, it is largest in zones in the reactor where A is in excess.
The largest values are found near the left wall of the reactor where the con
vective velocity is low and A is in slight stoichiometric excess. The residence time
for uid particles in this region is relatively long compared to the reaction time.
In general, Y
21
is larger than Y
22
, which is easily explained by the fact that A
enters the reactor in environment one. Finally, note that at the reactor outlet Y
2
is not uniformly mixed across the tube. Thus, despite the high energy dissipation
in the reactor (as measured by the pressure drop), the macromixing at the outlet
is not complete (Liu and Fox, 2006).
In Fig. 9, the distribution of reactant C is shown in each environment. As c
C
is a linear combination of x and Y
2
(Eq. 78), we can distinguish features of both
Fig. 7 and Fig. 8 in the plots in Fig. 9. In particular, because C is injected in the
righthand inlet stream, c
C2
and x
2
appear to be quite similar. Finally, as shown
in Liu and Fox (2006), the CFD predictions for the outlet conversion X are in
excellent agreement with the experimental data of Johnson and Prudhomme
(2003a). For this reactor, the local turbulent Reynolds number Re
L
is relatively
small. The good agreement with experiment is thus only possible if the effects of
the Reynolds and Schmidt numbers are accounted for using the correlation for
R shown in Fig. 4. Further details on the simulations and analysis of the CFD
results can be found in Liu and Fox (2006).
The example reactions considered in this section all have the property that
the number of reactions is less than or equal to the number of chemical species.
Thus, they are examples of socalled simple chemistry (Fox, 2003) for which
it is always possible to rewrite the transport equations in terms of the mix
ture fraction and a set of reactionprogress variables where each reaction
progress variablereactionprogress variable 4 depends on only one reaction.
For chemical mechanisms where the number of reactions is larger than the
number of species, it is still possible to decompose the concentration vector
into three subspaces: (i) conservedconstant scalars (whose values are null
everywhere), (ii) a mixturefraction vector, and (iii) a reactionprogress vector.
Nevertheless, most commercial CFD codes do not use such decomposi
tions and, instead, solve directly for the mass fractions of the chemical spe
cies. We will thus look next at methods for treating detailed chemistry
expressed in terms of a set of elementary reaction steps, a thermodynamic
database for the species, and chemical rate expressions for each reaction step
(Fox, 2003).
RODNEY O. FOX 266
C. DETAILED CHEMISTRY
In a CFD model with detailed chemistry, the user must provide a chemical
mechanism involving K chemical species A
b
of the form (Fox, 2003)
K
b1
u
f
bi
A
b
k
f
i
k
r
i
K
b1
u
r
bi
A
b
for i 2 1; . . . ; I reactions (84)
The rate constants (k
f
i
and k
r
i
) and the stoichiometric coefcients (u
f
bi
and u
r
bi
) are
all assumed to be known. Likewise, the reaction rate functions R
i
for each
reaction step, the equation of state for the density r, the specic enthalpies for
the chemical species H
k
, as well as the expression for the specic heat of the uid
c
p
must be provided. In most commercial CFD codes, user interfaces are avail
able to simplify the input of these data. For example, for a combusting system
with gasphase chemistry, chemical databases such as ChemkinII greatly sim
plify the process of supplying the detailed chemistry to a CFD code.
The reaction rates R
i
will be functions of the state variables dening the
chemical system. While several choices are available, the most common choice
of state variables is the set of species mass fractions Y
b
and the temperature T.
In the literature on reacting ows, the set of state variables is referred to as the
composition vector /:
/ Y
1
; . . . Y
K
; T
T
where the mass fractions sum to unity: Y
1
+?+Y
K
1. The microscopic bal
ance equation for the composition vector has the form of Eq. (1) (Bird et al.,
2002). For a turbulent reacting ow, the CFD transport equation will thus have
the form of Eq. (3) after averaging.
With detailed chemistry, the most difcult term to close in the CFD transport
equation will be the averaged chemical source term
S/: As described in detail
in Fox (2003), the chemical source term depends on the reaction rates R
i
, which
can be highly nonlinear in the composition vector /. For this reason, simple
closures that neglect correlations between components of the composition vec
tor are usually inaccurate. Nevertheless, the default closure for detailed chem
istry in most commercial CFD is the socalled laminarchemistry
approximation:
S/ S
~
/. In words, this closure approximates the average
chemical source term by its value evaluated at the average composition vector.
In general, the laminarchemistry approximation overpredicts the reaction rate
of the principal reactants, which in reality will be reduced by niterate tur
bulent mixing (Fox, 2003). The simplest example is the reaction rate for the
acidbase reaction
S k
N
p
: Thus, when N is small the statistical error is quite signicant.
Ideally, we would like to have an acceptably accurate method that works for N
as small as one, and whose accuracy increases for larger N in a welldened
deterministic manner. Our method for accomplishing this task is called the direct
quadrature method of moments (DQMOM) (Fox, 2003; Wang and Fox, 2004;
Marchisio and Fox, 2005). We briey describe the resulting CFD model below.
Readers interested in more details should consult the references given above.
The application of DQMOM to the closed composition PDF transport equa
tion is described in detail by Fox (2003). If the IEM model is used to describe
micromixing and a gradientdiffusivity model is used to describe the turbulent
uxes, the CFD model will have the form
@rp
n
@t
= r U h ip
n
_ _
= DD
T
=p
n
for n 1; . . . ; N 85
RODNEY O. FOX 268
and
@rp
n
f
an
@t
= r U h ip
n
f
an
_ _
= DD
T
=p
n
f
an
rp
n
S
a
/
n
_ _
rp
n
g hf
a
i f
an
_ _
rb
an
86
where the subscript n denotes the environment and the subscript a denotes
the component of the composition vector. Thus, p
n
is the mass fraction of
environment n, and /
n
is the composition vector in environment n. The reader
will recognize these equations as an Nenvironment generalization of the
twoenvironment model introduced earlier. Note that, as in transported
PDF simulations, the chemical source term S
a
(/
n
) appears in closed form in
Eq. (86).
By denition, the sum of the mass fractions is unity: p
1
+?+p
N
1. Thus,
one of the equations for p
n
in Eq. (85) is redundant. In Eq. (86), f
an
is one of the
K+1 components of the composition vector in environment n. The mean com
position hf
a
i appearing in the micromixing model is dened by
hf
a
i
N
n1
p
n
f
an
(87)
Note that if the mass fractions are used to dene the composition vector, then
by denition
K
a1
f
an
1 (88)
where the sum is over all the K chemical species. This implies that the sum of
Eq. (86) over all species must yield Eq. (85), and thus that one of the chemical
species equations is redundant. Redundancy can be avoided in the CFD model
by solving Eq. (85) only for n 1,y, N1, and Eq. (86) for n 1,y, N but
with only K1 mass fractions. To avoid numerical errors, the mass fraction of
the species not solved for should be relatively large and, preferably, correspond
to a chemically inert species. The composition vector will then have K (including
temperature) components, and a total of N(K+1)1 transport equations must
be solved to represent the model in the CFD code. Alternatively, since the sum
of the chemical source term in Eq. (86) over all chemical species is null, the CFD
model can solve only Eq. (86) for a 1,y, K+1 and not use Eq. (85). This
leads to N(K+1) equations (i.e., one more than required
1
), but has the benet
that all of the transport equations have the same form.
The nal term in Eq. (86) is the correction term b
an
; which comes from
applying DQMOM to the transport equation for the composition PDF (Fox,
1
The extra equation is the mass continuity equation for r.
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 269
2003). This term is computed by solving a linear system of equations dened for
each a 1,y, K+1 by
N
n1
f
m1
an
b
an
N
n1
m1p
n
f
m2
an
D
T
j=f
an
j
2
for m 1; . . . ; N 89
This expression results from forcing the transport equations for the rst N
moments of f
a
, denoted by f
m
a
_
; to agree with the composition PDF transport
equation (Fox, 2003). For example, with N 2 the linear system in Eq. (89) can
be written in matrix form as
1 1
f
a1
f
a2
_ _
b
a1
b
a2
_ _
0
p
1
D
T
=f
a1
2
p
2
D
T
=f
a2
2
_ _
(90)
Solving for b
an
yields
b
a1
D
T
f
a2
f
a1
p
1
=f
a1
2
p
2
=f
a2
2
_ _
(91)
and
b
a2
D
T
f
a2
f
a1
p
1
=f
a1
2
p
2
=f
a2
2
_ _
(92)
The reader will recognize these terms as having of the same form as the cor
rection terms in the twoenvironment model discussed earlier. With N 1,
b
a1
0 and the model reduces to the laminarchemistry approximation. With
N 2, additional information is obtained concerning the secondorder mo
ments of the composition vector. Likewise, by using a larger N, the Nthorder
moments are controlled by the DQMOM correction terms found from Eq. (89).
As noted earlier, the sum of the mass fractions is unity and thus Eq. (86) will
be consistent with Eq. (85) only if the sum of the correction term b
an
over all
chemical species a 1,y, K is null. In general, this will not be the case if Eq.
(89) is used. Another difculty that can arise is that the mass fractions in two
environments may be equal, e.g., f
a1
f
a2
, and thus the coefcient matrix in
Eq. (89) will be singular. This can occur, for example, in the equilibriumchem
istry limit where the compositions depend only on the mixture fraction, i.e.,
/ /
N
(x). For chemical species that are not present in the feed streams, the
equilibrium values for x 0 and x 1 are zero, but for intermediate values of
the mixture fraction, the equilibrium values are positive. This implies that the
equilibrium values will be the same for at least two values of the mixture frac
tion in the range 0oxo1. Thus, in the equilibrium limit it is inevitable that two
environments will have equal mass fractions for certain species at some point in
the ow eld. Since singularity implies an underlying correlation between
RODNEY O. FOX 270
components of the composition vector, in most cases these difculties can be
overcome by computing the correction terms using a set of moments that is
different than hf
m
a
i (i.e., one can include cross moments hf
mp
a
f
p
b
i. For example,
one possible choice of moments that also ensures that the sum of the correction
terms is null is to use the cumulative mass fractions. We will describe this choice
next, but the reader should keep in mind that other choices may be required if
the coefcient matrix in Eq. (89) becomes singular.
Let Y
a
denote the mass fractions of the K chemical species describing the
reacting ow. By denition,
K
a1
K Y
a
1: Assuming that the chemical species
are numbered such that the major species (e.g., reactants) appear rst,
2
followed
by the minor species (e.g., products), we can dene a linear transformation by
X
b
b
a1
Y
a
for b 1; . . . ; K (93)
Note that by denition X
K
1 and thus X
b
is the cumulative mass fraction of
the rst b species. The inverse transformation corresponding to Eq. (93) is
Y
1
X
1
and Y
a
X
a
X
a1
for a ! 2 (94)
The correction terms can be computed using either Y
a
or X
a
. However, it is clear
that the correction terms will depend on which choice is used since the moments
controlled by DQMOM are different (i.e., hY
m
a
i vs. hX
m
a
i). Thus, for example,
with xed N one set of moments may lead to singular correction terms, but not
the other. In general, we can continue to solve for the mass fractions in the CFD
model, but with the correction terms computed using the cumulative mass
fractions as follows.
1. Given the mass fractions in each environment Y
an
, use Eq. (93) to compute
the cumulative mass fractions X
an
.
2. Use the integer moments of X
a
to compute the correction terms as follows:
N
n1
X
m1
an
b
an
N
n1
m1p
n
X
m2
an
D
T
rX
an
j j
2
for m 1; . . . ; N 95
3. Compute the correction terms for the mass fractions b
an
using Eq. (94) as
follows:
b
1n
b
1n
and b
an
b
an
b
a1n
for a ! 2 (96)
2
In the computer code, a sorting algorithm can be used to put the mean mass fractions hY
a
i in
descending order before dening X
b
. By keeping track of the order of the indices, one can easily
dene the inverse transformation needed to compute Y
a
from X
b
.
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 271
Note that since X
Kn
1, Eq. (95) leads to the degenerate case where
b
Kn
0, and thus the sum of the correction terms over all species is null.
Also, if one of the mass fractions is null
3
(say Y
g
0) then X
g
X
g1
and
thus the correction term for such mass fractions will be null due to Eq. (96).
As discussed in detail in Wang and Fox (2004), on the one hand, for non
reacting systems the DQMOM approach with N environments will exactly re
produce the moment equations for each chemical species up to order N.
However, N cannot be chosen to be too large because the coefcient matrix in
Eq. (89) can become poorly conditioned. In contrast, the moments estimated
from the transported PDF simulations will have statistical uctuations that can
only be reduced by time/ensemble averaging. In this sense, DQMOM is pref
erable. On the other hand, for reacting systems the moments of chemical species
obey transport equations that contain unclosed averages involving the nonlinear
chemical source term. In both approaches, these averages are approximated by
f
m1
a
S
a
/
_
_
c
m1
a
S
a
w f
/
dw %
N
n1
p
n
f
m1
an
S
a
/
n
_ _
(97)
where in a transported PDF code p
n
is the particle weight and /
n
is the particle
composition vector (Fox, 2003). The integral in Eq. (97) is just the denition of
the expected value. It can be seen that the integral is replaced by a nite sum
over a set of N quadrature points. Thus, the accuracy of the moments will
depend on the degree of nonlinearity of S(/). For a weakly nonlinear source
term, a loworder quadrature method may be adequate and DQMOM would be
preferred. Such cases have been successfully investigated by Wang and Fox
(2004). In contrast, for strong nonlinearities (e.g., typical of combusting sys
tems), a large value of N may be needed to approximate the integral in Eq. (97)
and transported PDF simulations would be preferred.
In practice, it may be difcult to determine in advance which method is best
to use for a particular application. For example, the CFD results may be more
sensitive to largescale inhomogeneities in the ow eld than to the chemical
source term closure. A rational approach to determine whether a more detailed
SGS model is needed might be to start with N 1 (laminarchemistry approx
imation) and compare the predicted mean chemical species elds to the two
environment model (N 2). If the differences are small, then the simpler model
is adequate. However, if the differences are large, then the CFD simulation can
be repeated with N 3 and the results compared to N 2. Naturally, once this
procedure has converged, it will still be necessary to validate the CFD results
with experimental data whenever possible. Indeed, it may be necessary to
3
If the mass fractions are sorted in descending order, then all of the null mass fractions will be
grouped together at the end of the array. The procedure can thus be simplied by using only the non
zero mass fractions to dene X
b
. In practice, one can dene a lower limit for Y
a
and set the
correction term to zero for mass fractions below this limit.
RODNEY O. FOX 272
improve the turbulent transport and micromixing models before satisfactory
agreement is obtained even with the stateoftheart transported PDF method.
IV. Production of Fine Particles
The CFD models discussed in the previous section considered mass balances
for a nite number of chemical species. In this section, we will extend these
models to include systems wherein a second phase is produced. Such systems
include aerosols (Friedlander, 2000; Wright et al., 2001), reactive precipita
tion (Pohorecki and Baldyga, 1983; Garside and Tavare, 1985; Pohorecki
and Baldyga, 1988; Villermaux and David, 1988; Marcant and David, 1991;
Mahajan and Kirwan, 1993; David and Marcant, 1994; Seckler et al., 1995;
Mahajan and Kirwan, 1996; Aoun et al., 1999; Johnson and Prudhomme,
2003b,c), colloids (Oles, 1992; Sandku hler et al., 2003, 2005; Waldner et al.,
2005), and ame synthesis of nanoparticles (Kodas et al., 1987; Akhtar et al.,
1991; Xiong and Pratsinis, 1991; Kruis et al., 1993; Pratsinis et al., 1996; Zhu
and Pratsinis, 1997; Briesen et al., 1998; Pratsinis, 1998; Kammler and Pratsinis,
1999, 2000; Kammler et al., 2001; Mueller et al., 2004a,b; Tani et al., 2004a,b)
(to name just a few examples). We will limit the discussion to systems wherein
the particles are not too large, in other words, to particles that on average
follow the local uid velocity (Davies, 1949). Systems with larger particles will
be considered in Section V. In terms of the CFD model, the primary difference
between ne and large particles is that the former can be treated as pseudo
species by extending the mass balances and chemical kinetic expressions that we
have already considered (Piton et al., 2000; Johannessen et al., 2000, 2001). In
contrast, for larger particles, a separate momentum balance is required for each
phase (Fan et al., 2004), which signicantly complicates the solution procedure
used to solve the CFD model.
The denition of a ne particle can be made more quantitative by intro
ducing the article Stokes number St, which measures the particletouid re
sponsetime ratio to changes in the velocity (Fuchs, 1964):
St
gr
s
d
2
s
12r
f
n
f
(98)
In this denition, r
s
and r
f
are the solid and uid densities, respectively. The
characteristic diameter of the particles is d
s
(which is used in calculating the
projected crosssectional area of particle in the direction of the ow in the drag
law). The kinematic viscosity of the uid is n
f
and g is a characteristic strain rate
for the ow. In a turbulent ow, g can be approximated by 1/t
Z
when d
s
is
smaller than the Kolmogorov length scale Z. (Unless the turbulence is extremely
intense, this will usually be the case for ne particles.) Based on the Stokes
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 273
number, we can neglect the momentum equation for the solid phase when
Sto0.14 (Dring, 1982). Note that for systems with growing or aggregating
particles, the characteristic diameter of the largest particles should be used to
compute the Stokes number. Thus, as a general rule of thumb, in a liquidphase
turbulent ow solid particles (with r
s
Er
f
) will follow the uid if d
s
rZ. Since
typical minimum values for the Kolmogorov length scale in practical systems
are in the range 10100 mm, nucleation and growth of nanoparticles and crys
tals, and colloidal aggregation can all be treated as systems involving ne par
ticles.
A fundamental modeling challenge that arises when dealing with particulate
systems is the need to describe the particle size distribution (PSD). Depending
on the application (Randolph and Larson, 1988; Ramkrishna, 2000), this is
done by dening a population balance equation (PBE) governing a number
density function (NDF) n(l; x, t). The latter is dened by an internal coor
dinate l, which may correspond to mass, volume, or (as done here) length.
4
In
words, n(l; x; t, is the number concentration of particles with lengths in the
range [l, l+dl]) and thus it has units of (number)/(volume length).
The NDF is very similar to the PDFs introduced in the previous section to
describe turbulent reacting ows. However, the reader should not confuse them
and must keep in mind that they are introduced for very different reasons. The
NDF is in fact an extension of the nitedimensional composition vector / to
the case of an innite number of scalars (parameterized by 0ploN). Thus,
even in the case of laminar ow where the PDFs are not needed, the NDF still
introduces an extra dimension (l) to the problem description. The choice of the
state variables in the CFD model used to solve the PBE will depend on how the
internal coordinate is discretized. Roughly speaking (see Ramkrishna (2000)
for a more complete discussion), there are two approaches that can be em
ployed:
1. Sectional methods that represent the NDF by a histogram (Kumar and
Ramkrischna, 1996).
2. Quadrature methods that approximate integral constraints (e.g., moments) of
the NDF (McGraw, 1997).
For cases with only one internal coordinate, either approach can be imple
mented in a CFD code (but the computational cost for the same accuracy can be
very different). However, for cases with more than one internal coordinate, only
the quadrature methods are computationally tractable on current computers.
Thus, in the examples below, we will describe only CFD models based on the
4
The preferred choice of internal coordinates is discipline dependent. Nevertheless, the conser
vation of solid mass will imply constraints on particular moments of the PSD. In general, given the
relationship between the various choices of coordinates, it is possible (although not always practical)
to rewrite the PBE in terms of any choice of internal coordinate.
RODNEY O. FOX 274
quadrature approach. Details on particular applications of this approach can be
found in our recent publications (Marchisio et al., 2001a,b, 2003a,b; Wang and
Fox, 2003; Wang et al., 2005a).
A. MIXINGDEPENDENT NUCLEATION AND GROWTH
As a rst example of a CFD model for neparticle production, we will
consider a turbulent reacting ow that can be described by a species concen
tration vector c. The microscopic transport equation for the concentrations is
assumed to have the standard form as follows:
@c
@t
= Uc = G=c Sc (99)
All of the chemical species, except one, will be assumed to be completely soluble.
The one partially insoluble species will nucleate and grow a solid phase. A
typical example is A+BP where P is a sparingly soluble compound. The rates
of nucleation J and molecular surface growth G can be functions of the local
concentration vector c, the particle size l, and the local turbulence properties.
Neglecting aggregation and breakage processes, a microscopic PBE for this
system can be written as follows:
@n
@t
= Un = G
n
=n Jl; c
@
@l
Gl; cn (100)
Note that we have used the uid velocity U to describe convection of particles,
which is valid for small Stokes number. In most practical applications, J is a
highly nonlinear function of c. Thus, in a turbulent ow the average nucleation
rate will depend strongly on the local micromixing conditions. In contrast, the
growth rate G is often weakly nonlinear and therefore less inuenced by tur
bulent mixing.
The quadrature approach for treating Eq. (100) introduces a moment trans
formation dened by
m
k
_
1
0
l
k
nldl for k 0; 1; . . . ; 1 (101)
Applying this transformation to Eq. (100) yields
@m
k
@t
= Um
k
= G=m
k
J
k
c kG
k1
c (102)
where the moments of the nucleation function are dened by
J
k
c
_
1
0
l
k
Jl; c dl (103)
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 275
and the growth rates for the moments are dened by
G
k
c
_
1
0
l
k
Gl; cnl dl (104)
Note that in the special case of sizeindependent growth, this term can be ex
pressed as a closed function of the moments, i.e., G
k
c Gcm
k
. Note also
that when deriving Eq. (102) we have neglected the sizedependence of G
n
. This
is justied in turbulent ows and, in any case, to do otherwise would require a
micromixing model that accounts for differential diffusion (Fox, 2003).
In principle, any functional form could be used for the nucleation rate.
However, to simplify the discussion, we will assume that only particles of zero
size are formed by nucleation so that Eq. (103) becomes
J
k
c d
0;k
Jc (105)
where d
j
,
k
is the Kronecker delta and J(c) contains the dependence of the nu
cleation rate on the local composition vector. Note that under this assumption
nucleation only appears in the equation for the moment of order k 0.
Due to the presence of the unknown NDF inside the integral in Eq. (104), the
growth term is not closed (i.e., it cannot be computed exactly in terms of the
moments unless G is independent of l). In the quadrature method of moments
(QMOM), the integral is approximated by a sum over a set of M weights (w
m
)
and M abscissas (l
m
):
G
k
c
M
m1
w
m
l
k
m
Gl
m
; c (106)
The weights and abscissas are determined in QMOM by forcing them to agree
with the quadrature approximation of the rst 2 M moments:
m
k
M
m1
w
m
l
k
m
for k 0; 1; . . . ; 2M 1 (107)
This system of 2 M nonlinear equations is illconditioned for large M, but can be
efciently solved using the productdifference (PD) algorithm introduced by
McGraw (1997). Thus, given the set of 2 M moments on the lefthand side of
Eq. (107), the PD algorithm returns w
m
and l
m
for m 1,y, M. The closed
microscopic transport equation for the moments can then be written for
k 0,y, 2 M1 as
@m
k
@t
= Um
k
= G=m
k
d
0;k
Jc k
M
m1
w
m
l
k1
m
Gl
m
; c (108)
RODNEY O. FOX 276
To proceed further, this expression and Eq. (99) must be averaged to nd the
CFD transport equations for the species concentrations and the moments.
By dening the composition vector to include the species concentrations and
the moments as follows:
/ c
1
; . . . ; c
K
; m
0
; . . . ; m
2M1
T
we can observe that the microscopic transport equations have the same form as
those used for detailed chemistry in Section III.C. Thus, any turbulent reacting
ow model that can be used for detailed chemistry can also be used as a CFD
model for neparticle production. For example, using the DQMOM approach
to treat the composition PDF leads to Eqs. (85) and (86) with source terms for
the moments given by the last two terms in Eq. (108) as follows:
S
k
/ d
0;k
Jc k
M
m1
w
m
l
k1
m
Gl
m
; c (109)
Note that each environment in the micromixing model will have its own set of
concentrations c
an
and moments m
kn
, reecting the fact that the PSD is coupled
to the chemistry and will thus be different at every SGS point in the ow. The
PD algorithm is applied separately in each environment to compute the weights
(w
mn
) and abscissa (l
mn
) from the quadrature formula as follows:
m
kn
M
m1
w
mn
l
k
mn
for k 0; 1; . . . ; 2M 1 (110)
Thus, the source terms for each environment S(c) and S
k
(/) will be closed. Of
particular interest are the local nucleation rates J(c
n
). As discussed in Wang and
Fox (2004), due to poor micromixing the local nucleation rates can be much
larger than those predicted by the average concentrations J(hci). This results in a
rapid increase in the local particle number density m
0n
due to the creation of a
very large number of nuclei. As discussed below, this will have signicant con
sequences on the local rate of aggregation.
The CFD model for nucleation and growth can now be solved to determine
the average species concentrations hci and the average moments of the NDF
hm
k
i. However, to properly interpret the computational results, care must be
taken in dening the averages for terms involving the moments. Starting from
the denition of the moments from the microscopic NDF (Eq. 101), the average
moments are dened by
m
k
h i
_
1
0
l
k
hnli dl (111)
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 277
where hnli is the Reynoldsaverage NDF. For the multienvironment model,
the Reynoldsaverage quantities are dened by
m
k
h i
N
n1
p
n
m
kn
and hnli
N
n1
p
n
n
n
l (112)
where n
n
(l) is the NDF in environment n. Using the denition of the weights
and abscissas (Eq. 110), the average moments can be expressed as follows:
m
k
h i
N
n1
p
n
M
m1
w
mn
l
k
mn
M
m1
hw
m
l
k
m
ia
M
m1
w
m
h i l
m
h i
k
(113)
Thus the weights and abscissas for the average NDF cannot be found to be
averaging those for the NDF in each environment. Due to the nonlinear re
lationship between the moments and weights and abscissas, this result is not
surprising.
5
However, it does illustrate that hw
m
i and hl
m
i are not the relevant
quantities needed to reconstruct hm
k
i.
B. BROWNIAN AND SHEARINDUCED AGGREGATION AND BREAKAGE
As mentioned above, when local nucleation rates are high, the local number
density of particles will be large and aggregation will be favored. For very small
particles (e.g., submicron), the dominant aggregation mechanism is Brownian
motion (Einstein, 1905). Physically, the particles move by random walks driven
by momentum exchange with solvent molecules. When two particles collide,
they will stick together with a probability p
B
to form a doublet. By this mech
anism, larger clusters are eventually formed with a fractal dimension d
f
(Meakin, 1988; Sorensen, 2001; Lattuada et al., 2003a,b). If the volume fraction
of particles is high enough, the clusters can reach innite size in a nite time
5
If we interpret the weights and abscissas as a deltafunction representation of the NDF:
n
n
l
M
m1
w
mn
dl l
mn
N
n1
M
m1
p
n
w
mn
dl l
mn
By using the PD algorithm, this set of NM delta functions is reduced to a set of only M delta
functions, but with the same values for the rst 2 M moments hm
k
i as the original set. It should be
obvious that this cannot be accomplished by simply averaging the weights and abscissas.
RODNEY O. FOX 278
span through a process called gelation (Sandku hler et al., 2005). However, if the
clusters are growing in a turbulent ow eld, once their characteristic size is
greater than approximately one micron other sheardriven processes become
dominant (Oles, 1992). These are typically classied as shearinduced growth,
breakage and restructuring. As in the case of chemical kinetics, aggregation and
breakage models are required to describe these phenomena. From the perspec
tive of developing CFD models, we can assume that such kinetic expressions
are available. Thus, our focus here will be on how to implement the PBE in a
CFD simulation to predict the moments of the PSD.
The simplest aggregation and breakage models can be formulated in terms of
the NDF n(u), which uses volume as the independent variable.
6
The microscopic
transport equation for the NDF has the form (Wang et al., 2005a,b)
@n
@t
= Un = G
n
=n Ju; c
@
@u
Gu; cn
Au Bu 114
where A(u) and B(u) are the aggregation and breakage terms, respectively.
Although we do not do so here, these terms can be assumed to be dependent on
the species concentrations c without changing the form of the CFD model. For
binary aggregation and breakage, the aggregation term can be expressed as
follows (Ramkrishna, 2000):
Au
1
2
_
u
0
bu s; snu sns ds
nu
_
1
0
bu; snsds 115
where b is the aggregation kernel. A typical breakage term has the form
Bu
_
1
u
bujsasnsds aunu (116)
where b is the daughtersize distribution and a is the breakage kernel.
For Brownian aggregation, the aggregation kernel can be written as follows
(Elimelech et al., 1995; Sandku hler et al., 2003):
bu; s
2p
B
k
B
T
3m
u
1=d
f
s
1=d
f
_ _
u
1=d
f
s
1=d
f
_ _
(117)
6
It is also possible to use length as the independent variable as described in Wang et al. (2005b).
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 279
while for shearinduced aggregation it has the form (Oles, 1992; Elimelech et al.,
1995) as follows:
bu; s gau; s u
1=d
f
s
1=d
f
_ _
3
(118)
A general expression can be found by combining these two cases (Melis et al.,
1999). In these expressions, k
B
is the Boltzmann constant, T is the uid tem
perature (Kelvin), m is the uid viscosity, g is the local shear rate, and a is an
efciency factor. For shearinduced breakage, the kernel is usually t to exper
imental data (Wang et al., 2005a,b). A typical form is (Pandya and Spielman,
1983) as follows:
au c
1
g
c
2
u
c
3
=d
f
(119)
where c
1
c
3
are empirical tting parameters. A typical daughtersize distribu
tion is
bujs du fs du 1 f s (120)
where f 1/2 corresponds to equalsize daughters and f51 corresponds to
erosion (i.e., a very small and a very large daughter).
At this point, we should step back and make a few comments concerning
Eq. (114). First, it should be obvious to the reader that many modeling assump
tions have already been invoked to arrive at forms for the aggregation and
breakage terms and the rate functions needed to dene them. Therefore, we should
keep in mind that, just as when working with chemical kinetics, the CFD pre
dictions can be no better than the basic physical/chemical models used to describe
the source terms. By their nature, aggregation and breakage terms are much more
difcult to formulate accurately than gasphase chemical mechanisms. Thus, we
should expect a greater degree of mismatch with experimental data for CFD
solutions for aggregationbreakage systems than we are accustomed to with, for
example, combustion systems (Raman et al., 2004). Due to the uncertainties in the
kinetics for aggregation and (especially) breakage, the use of a highly sophis
ticated SGS model for turbulent mixing (e.g., transported PDF or CMC models)
is likely unwarranted for most systems of interest. In fact, at present the greatest
need in this domain is carefully designed experiments to accurately measure the
rate constants appearing in the aggregation and breakage expressions.
A second general observation can be made by comparing the aggregation
terms (Eq. 115) to the breakage terms (Eq. 116): The former is second order in
n(u) and the latter is rst order. This implies that when n(u) is very large (e.g.,
due to high local nucleation), aggregation will be favored. In fact, in shear
dominated systems gelation is prevented only by virtue of the fact that the
aggregation efciency a drops off rapidly for large clusters. Thus, in systems
RODNEY O. FOX 280
with nucleation, growth, aggregation, and breakage (e.g., aggregating nano
particles), the PSD can be nonzero over a very wide range of cluster volumes
(i.e., 34 orders of magnitude is not uncommon). If sectional methods are used
to approximate n(u), such systems typically require a relatively large number of
sections (e.g., 25100) for reasonable accuracy (Marchisio et al., 2003b). For
this reason, when used in a CFD model, sectional methods require an unfavor
able tradeoff between reasonable computational cost and accuracy. Quad
rature methods, which have lower cost for equivalent accuracy (Marchisio et al.,
2003b), are thus a better choice for combining with CFD to describe these
complex systems.
A nal observation concerns the shear rate g appearing in both the aggre
gation and breakage kernels. Since from the outset we have assumed that the
clusters are small compared to the Kolmogorov length scale, the local shear rate
seen by a cluster is the instantaneous shear (e/u)
1/2
, where e is the uctuating
dissipation rate (Fox, 2003). The uctuating dissipation is very different than the
average dissipation e computed by the turbulence model (Fox and Yeung, 2003).
On average hi , but e(t) uctuates strongly on a characteristic time scale
proportional to the Kolmogorov time scale t
Z
. Whether these uctuations must
be accounted for in the CFD model depends on the characteristic aggregation
time. Marchisio et al. (2006) estimate that when the local solid volume fraction
exceeds 10
3
, the uctuations must be included in the CFD model. Otherwise, g
can be set equal to (e/u)
1/2
. In any case, we see from this discussion that g scales
with the local turbulent Reynolds number like g$Re
L
1/2
. Thus, shearinduced
aggregation and breakage will be important phenomena in turbulent ows and
their importance will increase with increasing Reynolds number.
We now turn to the question of developing a CFD model for neparticle
production that includes nucleation, growth, aggregation, and breakage.
Applying QMOM to Eq. (114) leads to a closed set of moment equations as
follows:
@m
k
@t
= Um
k
= G=m
k
S
k
/ (121)
where the moment source term is given by (Marchisio et al., 2003a)
S
k
/ d
0;k
Jc k
M
m1
w
m
u
k1
m
Gu
m
; c
M
m1
w
m
au
m
b
k
m
u
k
m
_ _
1
2
M
m1
M
p1
w
m
w
p
u
m
u
p
k
u
k
m
u
k
p
_ _
bu
m
; u
p
122
and
b
k
m
_
1
0
u
k
buju
m
du (123)
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 281
Thus, just as we saw with Eq. (109), the moment source term has the form found
with detailed chemistry (i.e., the righthand side of Eq. (122) depends only on
/).
7
The CFD transport equation can therefore be developed along the same
lines that we discussed earlier for nucleation and growth. In other words, the
DQMOM model can be used to describe micromixing of the moments of the
PSD at the subgrid scale, along with turbulent transport models to describe
macromixing. One new factor that can arise when dealing with aggregation is
that the moment source terms can be stiff. To handle this problem in CFD
simulations, Wang and Fox (2003) successfully implemented a tabulation
method originally designed for combustion chemistry.
C. MULTIVARIATE POPULATION BALANCES
The CFD model described above is adequate for particle clusters with a
constant fractal dimension. In most systems with uid ow, clusters exposed to
shear will restructure without changing their mass (or volume). Typically re
structuring will reduce the surface area of the cluster and the fractal dimension
will grow toward d
f
3, corresponding to a sphere. To describe restructuring,
the NDF must be extended to (at least) two internal coordinates (Selomulya
et al., 2003; Zucca et al., 2006). For example, the joint surface, volume NDF can
be denoted by n(s, u; x, t) and obeys a bivariate PBE.
With two (or more) internal coordinates, numerical approaches for the PBE
using sectional methods become intractable in the context of CFD. A practical
alternative is to use a nite number of samples to approximate the NDF in
terms of delta functions as follows:
ns; u
M
m1
w
m
ds s
m
du u
m
(124)
The weights w
m
and abscissas s
m
and u
m
are related to the bivariate moments by
m
k
1
k
2
__
s
k
1
u
k
2
ns; udsdu
M
m1
w
m
s
k
1
m
u
k
2
m
(125)
Thus, it would be natural to attempt to extend the QMOM approach to handle a
bivariate NDF. Unfortunately, the PD algorithm needed to solve the weights and
abscissas given the moments cannot be extended to more than one variable. Other
methods for inverting Eq. (125) such as nonlinear equation solvers can be used
(Wright et al., 2001; Rosner and Pykkonen, 2002), but in practice are compu
tationally expensive and can suffer from problems due to illconditioning.
7
There are, however, important differences. For example, in detailed chemistry the source terms
do not depend on the ow quantities. In contrast, all of the rate functions for particulate systems can
potentially depend on the local ow quantities such as the instantaneous shear rate.
RODNEY O. FOX 282
To overcome the difculty of inverting the moment equations, Marchisio and
Fox (2005) introduced the direct quadrature method of moments (DQMOM).
With this approach, transport equations are derived for the weights and ab
scissas directly, thereby avoiding the need to invert the moment equations dur
ing the course of the CFD simulation. As shown in Marchisio and Fox (2005),
the NDF for one variable with moment equations given by Eq. (121) yields two
microscopic transport equations of the form
@w
m
@t
= Uw
m
= G=w
m
a
m
a
m
(126)
and
@w
m
u
m
@t
= Uw
m
u
m
= G=w
m
u
m
b
m
b
m
(127)
where the source terms a
m
and b
m
are found from the moment source terms S
k
,
and a
m
and b
m
are correction terms that depend on G9ru
m
9
2
. The role of the
correction terms is to ensure that Eqs. (126) and (127) are mathematically
equivalent to Eq. (121), and they result from the diffusion terms during the
nonlinear change of variables (Eq. 107). Thus, they will be null in the absence of
molecular diffusion and their forms do not depend on the moment source terms.
The extension of DQMOM to bivariate systems is straightforward and, for
the surface, volume NDF, simply adds another microscopic transport equation
as follows:
@w
m
s
m
@t
= Uw
m
s
m
= G=w
m
s
m
c
m
c
m
(128)
Example calculations for a bivariate system can be found in Marchisio and Fox
(2006) and Zucca et al. (2006). We should note that for multivariate systems the
choice of the moments used to compute the source terms is more problematic
than in the univariate case. For example, in the bivariate case a total of 3 M
moments must be chosen to determine a
m
, b
m
and c
m
. In most applications,
acceptable accuracy can be obtained with 3rMr5. Thus, the maximum
number of moments that will be required is 15, and one might decide to use
bivariate moments up to order four as shown below:
m
0,0
,
m
1,0
, m
0,1
,
m
2,0
, m
1,1
, m
0,2
,
m
3,0
, m
2,1
, m
1,2
, m
0,3
,
m
4,0
, m
3,1
, m
2,2
, m
1,3
, m
0,4
.
However, there is no guarantee that this set of moments will be linearly inde
pendent. Even if they do work, we can note that compared to the univariate case
where the ten moments m
0
m
9
would be employed, the accuracy (as measured
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 283
by the highestorder moment used) decreases as the number of internal coor
dinates increases. Thus, the only way to obtain equivalent accuracy would be to
increase M.
Leaving aside the question of determining which moments to use, we also
need a consistent method for deriving a CFD transport equation for turbulent
reacting ows from the microscopic transport equations for the weights and
abscissas. In particular, since the moments have a nonlinear dependence on the
weights and abscissas, the denition of the micromixing model in terms of
the weights and abscissas must be consistent with that used for the moments.
8
The development of a consistent model is most easily done by proceeding in
three independent steps as follows:
1. Find the multienvironment model for the case where the moment source terms
are null and there is no micromixing: a
m
a
m
b
m
b
m
c
m
c
m
0:
2. Extend the model to include micromixing, but no moment source terms:
a
m
b
m
c
m
0.
3. Extend the model to include the moment source terms due to nucleation,
growth, etc.
We will now briey illustrate how these steps are carried out.
The multienvironment model for turbulent transport of the bivariate mo
ments in the absence of moment source terms has the form
@p
n
@t
= Up
n
= G
T
=p
n
(129)
and, for arbitrary values of k
1
and k
2
@p
n
m
k
1
k
2
n
@t
= Up
n
m
k
1
k
2
n
= G
T
=p
n
m
k
1
k
2
n
b
k
1
k
2
n
(130)
where the bivariate moments in environment n are given in terms of the
weights and abscissas by
m
k
1
k
2
n
M
m1
w
mn
v
k
1
mn
s
k
2
mn
(131)
As discussed earlier with other transported scalars, the correction term appearing
on the righthand side of Eqs. (130) is found by solving a linear system dened by
N
n1
m
k1
k
1
k
2
n
b
n
k
1
k
2
n
N
n1
k 1p
n
m
k2
k
1
k
2
n
G
T
=m
k
1
k
2
n
2
for k 1; . . . ; N (132)
8
We develop the CFD equations using the DQMOM model for micromixing. Nevertheless, care
must also be taken when using other micromixing models, including transported PDF methods.
RODNEY O. FOX 284
Note that in this expression k is unrelated to k
1
and k
2
, which are determined by
the choice of moment m
k
1
k
2
n
. Equation (132) denes the correction term b
n
k
1
k
2
n
for a
xed set of indices (k
1
, k
2
, n). In other words, given the weights and abscissas we
can use Eq. (131) to compute the bivariate moments, which can then be used in
Eq. (132) to solve for b
n
k
1
k
2
n
.
The multienvironment model for the weights and abscissas has the form
@p
n
w
mn
@t
= Up
n
w
mn
= G
T
=p
n
w
mn
A
mn
A
mn
(133)
@p
n
w
mn
u
mn
@t
= Up
n
w
mn
u
mn
= G
T
=p
n
w
mn
u
mn
B
mn
B
mn
134
and
@p
n
w
mn
s
mn
@t
= Up
n
w
mn
s
mn
= G
T
=p
n
w
mn
s
mn
C
mn
C
mn
135
where the nal two terms on the righthand sides are null in the absence of
micromixing and moment source terms. The correction terms
A
n
mn
; B
n
mn
and C
n
mn
are determined such that they are equivalent to the cor
rection term b
n
k
1
k
2
n
in Eq. (130). In other words, if we use Eq. (131) to replace
m
k
1
k
2
n
in Eq. (130), then Eqs. (133)(135) will be equivalent to Eq. (130) when
the correction terms satisfy the following linear system:
1 k
1
k
2
M
m1
u
k
1
mn
s
k
2
mn
A
n
mn
k
1
M
m1
u
k
1
1
mn
s
k
2
mn
B
n
mn
k
2
M
m1
u
k
1
mn
s
k
2
1
mn
C
n
mn
b
n
k
1
k
2
n
k
1
k
1
1p
n
M
m1
w
mn
v
k
1
2
mn
s
k
2
mn
G
T
=v
mn
j j
2
k
1
k
2
p
n
M
m1
w
mn
v
k
1
1
mn
s
k
2
1
mn
G
T
=v
mn
=s
mn
2
k
2
k
2
1p
n
M
m1
w
mn
v
k
1
mn
s
k
2
2
mn
G
T
=s
mn
j j
2
136
Note that the correction terms are proportional to G
T
and result from
turbulent velocity uctuations (represented by a gradientdiffusion model). For
the multienvironment model the composition vector is dened by
/ w
1
; . . . ; w
M
; w
1
v
1
; . . . ; w
M
v
M
; w
1
s
1
; . . . ; w
M
s
M
T
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 285
Thus, a total of 3 M variables are needed to describe the evolution of the
bivariate NDF in a turbulent ow. The rst step in the construction of the
CFD model is now complete.
In the second step we must add the micromixing terms from the DQMOM
model to Eqs. (133)(135). However, as we discussed earlier, we need to keep in
mind that micromixing conserves the moments of the NDF, and not the weights
and abscissas (see Eq. 113). The micromixing model in environment n for the bi
variate moments has the form
M
k
1
k
2
n
g hm
k
1
k
2
i m
k
1
k
2
n
_ _
(137)
where the moments in environment n are given by Eq. (131) and the average
moments are found by averaging over all N environments:
hm
k
1
k
2
i
N
n1
p
n
m
k
1
k
2
n
(138)
Thus, given the weights and abscissas, the micromixing term for the moments is
closed. Applying DQMOM, the micromixing source terms (which are added to
the righthand sides of Eqs. (133)(135)) can be shown to obey for each
n 1,y, N the linear system dened by
1 k
1
k
2
M
m1
u
k
1
mn
s
k
2
mn
A
mn
k
1
M
m1
u
k
1
1
mn
s
k
2
mn
B
mn
k
2
M
m1
u
k
1
mn
s
k
2
1
mn
C
mn
p
n
M
k
1
k
2
n
139
Note that the righthand side of this expression contains the closed micromixing
term for the moments (Eq. 137). To nd the 3 M micromixing source terms
(A
mn
, B
mn
, C
mn
) from this expression, we must choose a set of 3 M bivariate
moments. Note that because the moment equations are closed when only
micromixing is considered, the chosen moments will be reproduced exactly. A
convenient choice is to use the uncoupled moments m
k0
and m
0k
. (Note that this
same choice should then be used in Eq. (136).) This yields the linear system
1 k
M
m1
u
k
mn
A
mn
k
M
m1
u
k1
mn
B
mn
p
n
M
k0n
for k 0; . . . ; 2M 1 140
which can be solved to nd A
mn
and B
mn
for each n 1,y, N, and the linear
system
k
M
m1
s
k1
mn
C
mn
p
n
M
0kn
k 1
M
m1
s
k
mn
A
mn
for k 1; . . . ; M 141
RODNEY O. FOX 286
which can be solved to nd C
mn
. The second step in the construction of the CFD
model is now complete.
The third step is to add the moment source terms due to nucleation, growth,
aggregation, and restructuring. The exact form of these terms will depend on the
models used to describe these processes (see for example Eq. 122). However, if
we denote the source term for the bivariate moments as S
k
1
k
2
/, where / is the
composition vector including all variables needed to describe the kinetics, then
the source terms (including micromixing) in the CFD model (A
mn
, B
mn
, C
mn
) can
be found by simply adding p
n
S
k
1
k
2
/
n
to the righthand side of Eq. (139). The
extension of DQMOM to multivariate systems with more than two internal
coordinates is discussed in Marchisio and Fox (2005). As far as the CFD model
equations are concerned, no additional complications arise for higher
dimensional systems. Nevertheless, from a practical point of view the reader
should keep in mind that going to higher dimensions will necessarily diminish
the accuracy of the method as compared to the univariate case.
We end here with our discussion of CFD models for neparticle production.
The reader hopefully has a good feel for the issues involved and at least a
cursory understanding of the available models. The current status of the eld is
such that the CFD tools available for the analysis of neparticle systems are
adequate for most applications. Currently, the weakest link in the modeling
process is description of physical processes such as aggregation and breakage,
and their coupling to the ow eld. As mentioned earlier, there is a need for
carefully designed experiments with local measurements of the PSD and
turbulence elds that can be used for CFD validation. As in the eld of
turbulent mixing, the ideal experiment would be spatially homogeneous, or a
most onedimensional, to focus on the source terms for the chemical reactions
and the moments. It would also be useful to conduct a few idealized
experiments using direct numerical simulations (DNS) for homogeneous ne
particle systems undergoing nucleation, growth, aggregation, and breakage.
Although such experiments cannot be directly compared to real systems, they
would still be useful for calibrating CFD models used for micromixing and
shearinduced aggregation and breakage.
V. Multiphase Reacting Systems
The CFD models considered up to this point are, as far as the momentum
equation is concerned, designed for singlephase ows. In practice, many of the
chemical reactors used in industry are truly multiphase, and must be described
in the context of CFD by multiple momentum equations. There are, in fact,
several levels of description that might be attempted. At the most detailed level,
direct numerical simulation of the transport equations for all phases with fully
resolved interfaces between phases is possible for only the simplest systems. For
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 287
example, for uidparticle systems the NavierStokes equation must be solved
for the uid phase between the particles with enough detail to capture
momentum transport at the particle surfaces (Nguyen and Ladd, 2005). At the
same time, Newtonian equations for the particle positions and momenta must be
solved simultaneously to account for uid surface forces and particleparticle
collisions. Obviously, such a detailed model could not be used to describe a large
chemical reactor such as a uidized bed. Less costly methods have thus
been developed and are described in other chapters of this issue. In general, the
methods available in commercial CFD codes are based on the socalled
multiuid model (Drew and Passman, 1999) that makes no attempt to capture
the details at the interfaces between phases. Instead, the uid at the subgrid scale
is described by the volume or mass fractions for each phase much in the same
way that environments are used to describe micromixing in singlephase ows).
We will thus look briey at the structure of multiuid models and describe some
of the modeling assumptions that are required for multiphase reacting ows.
A. MULTIFLUID CFD MODELS
In this section, we will look briey at multiuid CFD models. Our primary
objective is to understand the modeling issues that arise and how they are dealt
with in CFD codes. To x ideas, we will look at a gassolid system (e.g., a
uidized bed) wherein the solid particles undergo growth, aggregation, and
breakage. Unlike in Section IV, we will assume here that the particle Stokes
number (St) can be larger than 0.14. Thus, it will be necessary to account for
particle momentum separately from the gas phase (or at least to account for the
effect of the second phase in the momentum balance). Nevertheless, the particle
size distribution (PSD) can be accounted for using DQMOM (Fan et al., 2004).
For simplicity, we will assume that the particle density r
s
is constant and
independent of particle size, and that the gas density r
g
is constant. At the
subgrid scale, the two phases are described by the volume fractions a
g
and a
s
for
the gas and solid phases, respectively. By denition, a
g
+a
s
1. In addition, the
DQMOM representation of the solid phase will introduce a volume fraction for
each abscissa a
sm
for m 1,y, M. By denition, a
s1
a
sM
a
s
. (See Fan
et al. (2004) for details.)
The multiuid CFD model at its most basic level consists of mass and mo
mentum balances for each phase. For the present example, the mass balance
for the gas phase can be written as follows:
@r
g
a
g
@t
= r
g
a
g
U
g
M
m1
M
gm
(142)
where M
gm
is the masstransfer rate from the gas to the solid phases. As we will
discuss later (Section V.B), a model must be provided to close the masstransfer
RODNEY O. FOX 288
term. In words, r
g
a
g
is the mass of gas per unit volume of the multiphase
mixture (i.e., gas and solid phases), whereas r
g
is the mass of gas per unit
volume of gas. The gas velocity U
g
appears in the convection term for the gas
phase, and is normally not equal to the solids velocities.
The mass balance for the solid phases (m+1,y, M) can be written as follows:
@r
s
a
sm
@t
= r
s
a
sm
U
sm
M
gm
3k
v
r
s
l
2
m
b
m
2k
v
r
s
a
m
(143)
The solid velocities U
sm
(one for each abscissa l
m
) appear in the convection term.
The source terms a
m
and b
m
are found from the DQMOM representation of
aggregation and breakage of the solid particles (Fan et al., 2004). Each particle
phase is represented by its volume fraction (instead of its weight w
m
) and its
characteristic length l
m
. Note that growth of solid particles (with constant den
sity r
s
) requires mass transfer from the gas phase, represented by M
gm
. In the
absence of growth the total solids volume fraction a
s
does not change. Thus, the
aggregation and breakage terms will cancel as follows:
M
m1
3l
2
m
b
m
2a
m
_ _
0 (144)
This property will result from applying DQMOM to the aggregation and
breakage terms in the PBE (Fan et al., 2004).
While the mass balances given above are relatively straightforward (assuming
that a suitable closure can be derived for the masstransfer terms), the mo
mentum balances are signicantly more complicated. In their simplest forms,
they can be written as follows:
@r
g
a
g
U
g
@t
= r
g
a
g
U
g
U
g
a
g
=p = r
g
M
m1
f
gm
r
g
a
g
g 145
and
@r
s
a
sm
U
sm
@t
= r
s
a
sm
U
sm
U
sm
a
sm
=p = r
sm
f
gm
M
n1
f
smn
r
s
a
sm
g 146
for the gas and solid phases, respectively. The two terms of the lefthand sides of
the momentum balances correspond to accumulation and convection. Note that
the conserved quantities (for example r
g
a
g
U
g
) are the momentum of a given
phase per unit volume of the mixture (hence the appearance of a
g
, etc.), and that
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 289
each phase has its own characteristic convection velocity (e.g., U
g
). The terms
on the lefthand side of Eqs. (145) and (146) account for changes in momentum
of each phase. Obviously, since momentum is exchanged between phases at the
interfaces and we are not resolving the interfaces, such phaseinteraction terms
will require models. In contrast, body forces (i.e., gravity g in this example)
appear in closed form.
The rst term on the righthand side of Eqs. (145) and (146) is a pressure term
shared by both phases. The purpose of this term (when r
s
and r
g
are constant) is
to ensure that the volumeaverage velocity, dened by
U
vol
a
g
U
g
M
m1
a
sm
U
sm
(147)
is solenoidal in the absence of mass transfer. Indeed, dividing the mass balances
(Eqs. 142 and 143) by the densities and neglecting the masstransfer terms lead to
= U
vol
0 (148)
Thus, just as for incompressible singlephase ow, the pressure p constrains
the velocity elds to ensure (in the case of multiphase ows) that the sum of the
phase volume fractions equals unity. In the presence of mass transfer, the right
hand side of Eq. (148) is nonzero; nevertheless, the role of the pressure is still the
same. Finally, we should note that in gassolid ows the maximum volume fraction
of the solid phase is less than unity due to physical constraints (i.e., when particles
are close packed there is still room for the gas phase so that 0oa
g
). To accom
modate this constraint, it is common to introduce a solidpressure term p
s
that
becomes extremely large when a
g
approaches its minimum value (e.g., a
g
0.4).
The second term on the righthand side of Eqs. (145) and (146) contains the
viscousstress models s
g
and s
sm
. Even for laminar ow, suitable forms for these
models are difcult to determine a priori. Typical models used in CFD introduce an
effective viscosity m
eff
for each phase, and describe the viscous stresses as follows.
r
x
m
eff;x
rU
x
rU
x
T
_
x g; sm (149)
Leaving aside the difcult question of whether this model holds for multiphase
ows, we still have the problem of determining m
eff,x
in terms of the computed
properties of the ow. The reader should appreciate that choosing an effective
viscosity for a multiphase ow is much more complicated than just adding a
turbulence model as done in singlephase turbulent ows. Indeed, even for a case
involving two uids (e.g., two immiscible liquids) for which the molecular viscosi
ties are constant, the choice of the effective viscosities is not obvious. For example,
even if the massaverage velocity dened by
U
mass
r
g
a
g
U
g
M
m1
r
s
a
sm
U
sm
r
g
a
g
r
s
a
s
(150)
RODNEY O. FOX 290
were laminar, the ow around individual particles could be turbulent (as measured
by the particle Reynolds number dened below) and the effective viscosity should
reect this fact. The simplest models account for particlescale turbulence using
an expression of the form
m
eff;g
m
g
1 C
s
a
s
Re
s
(151)
where m
g
is the molecular viscosity of the gas phase and Re
s
is a particle Reynolds
number dened by
Re
s
r
g
d
s
jU
s
U
g
j
m
g
(152)
and d
s
is the characteristic diameter of the particles. The model constant C
s
is order
unity, but must be t to experimental data. The effective viscosity in Eq. (151) has
the desired behavior in the limits where a
s
or Re
s
are very small; however, it is
unlikely to be accurate when the product of these terms is large.
The situation for cases where U
mass
is also turbulent is even more compli
cated. First, such largescale turbulence can be due to a variety of physical
phenomena, and thus have different characteristics. For example, largescale
turbulence can (as in singlephase ows) be introduced through the boundary
conditions (e.g., turbulent jets) or by using mixing devices (e.g., stirred tanks).
For such cases, it may be possible to make suitable modications to singlephase
turbulence models to arrive at useful expressions for the effective viscosity. In
contrast, largescale turbulence that arises due to internal properties of a mul
tiphase ow (e.g., density differences between phases) is more difcult to de
scribe by simple modications of standard turbulence models.
Second, due to the difculty of accessing multiphase ows with laserbased ow
diagnostics, there is very little experimental data available for validating multi
phase turbulence models to the same degree as done in singlephase turbulent
ows. For example, thanks to detailed experimental measurements of turbulence
statistics, there are many cases for which the singlephase ke model is known to
yield poor predictions. Nevertheless, in many CFD codes a multiphase ke model
is used to supply multiphase turbulence statistics that cannot be measured exper
imentally. Thus, even if a particular multiphase turbulent ow could be adequately
described using an effective viscosity, in most cases it is impossible to know
whether the multiphase turbulence model predicts reasonable values for m
eff
.
Third, many of the multiphase ows of interest to chemical engineers are in
regimes where both particlescale and largescale turbulence are signicant. For
example, in gasliquid bubble columns the particle Reynolds number (based on
the bubble rise velocity) is typically large. Thus, even for low gasow rates,
particlescale turbulence will be signicant. However, at low gasow rates and
with uniform sparging, a bubble column will have no largescale turbulence (i.e.,
the ow regime will be homogeneous) (Garnier et al., 2002; Harteveld et al.,
2003), and thus only the effective viscosity of individual particles should be
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 291
included in the model. As the gas ow rate is increased to a critical value of the
gas holdup (which can be as high as a
g
0.55 (Mudde, 2005)), the ow will
become unstable and largescale turbulence will be generated. Although it has
been attempted in the literature (Thorat and Joshi, 2004), it is unlikely that a two
uid ke turbulence model has the necessary mathematical structure to correctly
predict ow transitions or even the turbulence levels observed in homogeneous
bubbly ows. In contrast, a twouid model using only a particlescale effective
viscosity and appropriate force models can predict ow transitions (Monahan
et al., 2005) in reasonable agreement with experiments (Harteveld et al., 2003).
Nevertheless, it is likely that different CFD models will be required for different
ow regimes (e.g., homogeneous vs. churn turbulent) and the user must be careful
not to extend a particular model beyond its range of applicability.
Returning to the momentum equations, the third term on the righthand side
of Eqs. (145) and (146) contains the gassolid momentumexchange models
f
gm
. Likewise, the fourth term on the righthand side of Eqs. (146) contains the
solidsolid momentumexchange model f
smn
. Note that because solidsolid in
teractions conserve momentum, the latter must be dened such that
M
m1
M
n1
f
smn
0 (153)
Determination of accurate models for f
gm
and f
smn
is nontrivial (Drew and
Passman, 1999), and no consensus exists on the exact forms needed to describe
particular ows. Nevertheless, it is generally acknowledged that the momentum
exchange model must include drag terms with forms similar to
f
gm
a
g
a
sm
C
D
Re
s
3m
g
Re
s
4d
2
s
U
sm
U
g
(154)
where C
D
(Re
s
) is a drag coefcient, and
f
smn
a
sm
a
sn
C
mn
U
sn
U
sm
(155)
where C
mn
depends on the properties of the solid phases (Gao et al., 2006). Note
that the drag models depend on the velocity difference between two phases, and
thus can be nonzero even for cases where the velocity elds are uniform in time
and space. Other forces that can be included depend on gradients (temporal or
spatial) of the velocities or volume fractions (Drew and Passman, 1999), and
thus are only signicant for inhomogeneous ows. However, as can be shown
using linear stability analysis (Batchelor, 1988; Lammers and Biesheuvel, 1996;
Minev et al., 1999; Jackson, 2000; Johri and Glasser, 2002; Sankaranarayanan
and Sundaresan, 2002), spatially uniform solutions to the multiuid model are
usually unstable, implying that the stationary, homogeneous solution to the
multiuid model is not representative of the ow. Thus, even when simulating
homogeneous ows, it is important to include all relevant forces when
RODNEY O. FOX 292
comparing numerical simulations with experiments (Monahan et al., 2005).
Moreover, the computational requirements in terms of gird size needed to attain
gridindependent solutions are relatively high for the laminar twouid model
(Monahan et al., 2005). In the context of modeling chemical reactors, it will be
necessary to develop CFD models for the unresolved scales when applying
multiuid models to real reactors. Although procedures for developing such
models are still being actively investigated and no clear consensus has yet to
emerge (Sundaresan, 2000), here we will limit ourselves to a brief discussion of
the relevant issues.
To simplify the presentation, let us consider the transport equation for U
mass
found by summing together Eqs. (145) and (146):
@^ rU
mass
@t
= ^ rU
mass
U
mass
= ^ r u
d
u
d
=p = ^ r ^ rg (156)
where the phaseaverage density, dened by ^ r r
g
a
g
r
s
a
s
; obeys
@^ r
@t
= ^ rU
mass
0 (157)
Note that these expressions (Eqs. 156 and 157) appear deceptively simple (i.e., as
if the problem can be reduced to modeling a variabledensity, singlephase ow)
because we have hidden the difcult terms in the denition of some new
symbols! First, the phaseaverage stress r is dened by
^ r r
g
M
m1
r
sm
(158)
and (based on the model in Eq. (149)) it is not a simple function of U
mass
. Second,
a new multiphase stress term u
d
u
d
has be introduced and is dened by
^ r u
d
u
d
r
g
a
g
u
g
u
g
r
sm
a
sm
M
m1
u
sm
u
sm
(159)
where u
g
U
g
U
mass
and u
sm
U
sm
U
mass
are the differences between
the phase velocities and the massaverage velocity. We can note that for a
constantdensity, twophase system with r
g
6 r
s
, U
g
and U
s
are related to U
mass
and U
vol
by
U
g
r
s
a
g
r
s
r
g
U
vol
^ r
a
g
r
s
r
g
U
mass
U
s
r
g
a
s
r
s
r
g
U
vol
^ r
a
s
r
s
r
g
U
mass
160
Thus the twouid model can be formulated in terms of any two velocities chosen
from the set U
g
, U
s
, U
mass
and U
vol
, which might be useful, for example, to
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 293
examine the limiting case where r
g
(r
s
. The closures for shear stresses ^ r and
convection due to differences in the phase velocities u
d
u
d
in Eq. (156) are
necessarily ow dependent. Nevertheless, the simplest closures might use Eq.
(149) with U
mass
and an effective viscosity depending on ^ r and assume that the
slip velocity between phases u
d
is a known constant (e.g., depending on density
difference and bubble size). We should also note that it is possible to write the
model in terms of the volumeaverage velocity U
vol
(Eq. 147). However,
the resulting expressions are more complicated than Eqs. (156) and (157).
Up to this point we have not introduced any modeling concepts to deal with
largescale turbulence. However, if the Reynolds number corresponding to
Eq. (156) is large enough, the velocity eld U
mass
will become turbulent. In this
case, the computational resources needed to resolve all of the relevant ow
scales will increase drastically, and the multiuid CFD model will no longer be
tractable for analyzing chemical reactors. To deal with this difculty, we can
introduce a multiphase turbulence model based on Reynolds averaging Eq.
(156). Because ^ r is not constant, we will in fact use the Favre average. For
example, if the Reynoldsaverage velocity is denoted by hU
mass
i, then the Favre
average velocity is dened by
U
mass
h ^ rU
mass
i
h ^ ri
(161)
Note that we have also introduced the Reynoldsaverage, phaseaverage density
h^ ri. Applying the Favre average to Eq. (157) yields a closed expression for the
mass balance as follows:
@h^ ri
@t
= h ^ ri
U
mass
0 (162)
Applying the same process to Eq. (156) yields
@h ^ ri
U
mass
@t
= h ^ ri
U
mass
U
mass
= h ^ ri u
mass
u
mass
= h ^ ri
u
d
u
d
=hpi = ^ r h i h ^ rig 163
where the velocity uctuations due to largescale turbulence are denoted by
u
mass
U
mass
~
U
mass
. The most important unclosed terms in Eq. (163) are the
turbulence stresses u
mass
u
mass
and the Favreaverage multiphase stresses
u
d
u
d
.
The rst of these is usually closed by introducing a multiphase turbulence model
with appropriate modications to include the effect of interfacial momentum
exchange on production (dissipation) of largescale turbulence. The second term
u
d
u
d
is not directly related to turbulent velocity uctuations. Instead, it will
depend on correlations between the phaseaverage density ^ r and the velocity
difference u
d
. For example, if the bubble rise velocity U
r
were constant and
independent of the gas volume fraction, then
u
d
u
d
% U
2
r
e
v
e
v
where e
v
g/gis
RODNEY O. FOX 294
the unit vector in the vertical direction. More generally, the turbulent twouid
model for gasliquid ow should have properties similar to turbulence gener
ated by buoyancy in singlephase ows (Riley and DeBruynKops, 2003). Like
wise, in gassolid ows u
d
will depend on the Favreaverage drag term, and thus
on the particle Stokes number through correlations between gas and solid
phase velocity uctuations (Fan and Zhu, 1998). Although computationally
expensive using presentday computers (Agrawal et al., 2001; Zhang and
VanderHeyden, 2002), it might be instructive to use direct simulations of the
laminar twouid model (i.e., before Favre averaging) to parameterize multi
phase turbulence models as has been done for stably stratied ows (Shih et al.,
2005). This possibility is especially attractive because it offers access to ow
statistics that cannot be measured experimentally. At the very least, it might
allow us to distinguish between the adequacy of the various multiphase tur
bulence models available in the literature (Mudde, 2005).
The goal of the discussion above was obviously not to describe multiphase
turbulence models, but rather to point out the difculties encountered when
trying to derive a consistent set of transport equations. Although it is usually
not done, it is worthwhile to think of the averaging process used to arrive at
Eqs. (162) and (163) in two distinct steps: (1) ensemble averages over different
phase congurations to derive the laminar multiuid model (i.e., Eq. 146)
that can be used to describe multiphase ows without largescale turbulence,
and (2) Reynolds or Favre averages (or even LES) to describe turbulent
multiphase ows. Ideally, direct numerical simulations (DNS) of twophase
ows with resolved interfaces could be used to develop twouid models for
laminar multiphase ows. For example, the recent work of Nguyen and Ladd
(2005) uses DNS to understand the sedimentation of mono and polydisperse
hardsphere suspensions when the largescale ow is laminar, and the work of
Bunner and Tryggvason (2003) uses DNS to investigate bubbly ows. If a
twouid model (which does not resolve the interfaces) could be derived that
adequately reproduces these DNS data, then it could be used to investigate
the effects of largescale turbulence that arises, for example, when the system is
subjected to shear (Lakehal et al., 2002). The results from direct simulations of
the twouid model in the turbulent regime could then be used to develop
and validate multiphase turbulence models along the lines suggested by
Sundaresan (2000). Fortunately, with the continuing advances in computer
power, steady advances in DNS of twophase systems can be expected. There is
thus reason to be optimistic that more powerful multiphase turbulence models
will eventually be available for modeling practical systems such as chemical
reactors.
In the current state of the art, almost all multiphase CFD models available
in commercial codes use some type of turbulence model based on extending
models originally developed for singlephase ows. Such CFD models are
thus meant to describe fully turbulent ows (as opposed to laminar or tran
sitional ows). Nevertheless, many of these models have not been validated
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 295
experimentally in the context of dense multiphase ows for the reasons dis
cussed earlier, and thus should be used with caution even for turbulent mul
tiphase ows. In any case, there is still considerable room for improvement of
multiphase CFD models through comparison with carefully designed experi
ments for canonical ows. Even more so than for singlephase turbulence, it can
be expected that particular models will have limited ranges of applicability and
will have to be tuned for multiphase ows with different physics. For ex
ample, a twouid model for a solidliquid slurry in an agitated reactor will
require different physical models than a churnturbulent bubble column. In the
rst case, largescale turbulence is generated by the agitation system, while in
the second case, it is generated by buoyancy and interfacial dynamics. However,
for specic ows of interest to the chemical industry, it should be possible to
develop reliable CFD models for multiphase ow dynamics that can be used to
investigate scalar transport and chemical reactions needed to model chemical
reactors. In the next section, we will thus look briey at the additional models
needed to describe the transport and production of chemical species and
thermal energy.
B. INTERPHASE MASS/HEATTRANSFER MODELS
Our discussion of multiphase CFD models has thus far focused on describing
the mass and momentum balances for each phase. In applications to chemical
reactors, we will frequently need to include chemical species and enthalpy
balances. As mentioned previously, the multiuid models do not resolve the
interfaces between phases and models based on correlations will be needed to
close the interphase mass and heattransfer terms. To keep the notation simple,
we will consider only a twophase gassolid system with a
g
+a
s
1. If we
denote the mass fractions of N
sp
chemical species in each phase by Y
ga
and Y
sa
,
respectively, we can write the species balance equations as
@r
g
a
g
Y
ga
@t
= r
g
a
g
Y
ga
U
g
= J
ga
M
a
R
ga
, (164)
and
@r
s
a
s
Y
sa
@t
= r
s
a
s
Y
sa
U
s
= J
sa
M
a
R
sa
. (165)
The terms J
ga
and J
sa
are the diffusive uxes of species a in the gas and solid
phases, respectively. Note that in addition to molecularscale diffusion, these
terms include dispersion due to particlescale turbulence. The latter is usually
modeled by introducing a gradientdiffusion model with an effective diffusivity
along the lines of Eqs. (149) and (151). Thus, for large particle Reynolds num
bers the molecularscale contribution will be negligible. The term M
a
is the
RODNEY O. FOX 296
masstransfer rate from the gas to the solid phase for species a. By denition, the
mass fractions sum to unity so that S
N
sp
a1
J
ga
S
N
sp
a1
J
sa
0 and M
g
S
N
sp
a1
M
a
(see Eq. 142). The terms R
ga
and R
sa
are the reaction rates for species a in each
phase. By denition, mass is conserved so that S
N
sp
a1
R
ga
0 and S
N
sp
a1
R
sa
0.
Although we do not write them explicitly here, the reader can appreciate that
the enthalpy balances for each phase will have a form similar to Eqs. (164) and
(165), and can be used to determine the temperatures T
g
and T
s
of each phase.
Likewise, mass transfer will lead to corresponding terms in the momentum
balances (Eqs. 145 and 146) (Bird et al., 2002).
CFD models for turbulent multiphase reacting ows do not solve the
laminar twouid balances (Eqs. 164 and 165) directly. First, Reynolds
averaging is applied to eliminate the largescale turbulent uctuations. Using
Eq. (164) as an example, we can apply Reynolds averaging to nd (with r
g
constant)
@r
g
ha
g
i
Y
ga
@t
= r
g
ha
g
i
Y
ga
U
g
_ _
= r
g
ha
g
i
Y
00
ga
u
00
g
_ _
= hJ
ga
i hM
a
i hR
ga
i 166
The Reynoldsaverage gas volume fraction ha
g
i is found from
@r
g
ha
g
i
@t
= r
g
ha
g
i
U
g
_ _
hM
g
i (167)
wherein the Reynoldsaverage masstransfer term hM
g
i is unclosed. The Favre
average gas velocity
U
g
is dened by
U
g
ha
g
U
g
i
ha
g
i
(168)
and its transport equation is found by Reynolds averaging Eq. (145). Although
we do not write it out explicitly here, the reader should appreciate that the
Reynoldsaverage gasphase momentum equation has a number of unclosed
terms that require models.
Returning to Eq. (166), the third term on the lefthand side involves the
turbulent scalar uxes, dened by
Y
00
ga
u
00
g
ha
g
Y
00
ga
u
00
g
i
ha
g
i
(169)
where the scalar and velocity uctuations are dened by Y
00
ga
Y
ga
Y
ga
and
u
00
g
U
ga
U
g
; and respectively. The usual model for the scalar uxes is gra
dient diffusion as follows:
Y
00
ga
u
00
g
G
T
=
Y
ga
(170)
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 297
where G
T
is a turbulent diffusivity that is computed from the multiphase tur
bulence model. The CFD model can then be written as
@r
g
ha
g
i
Y
ga
@t
= r
g
ha
g
i
Y
ga
U
g
_ _
= G
eff
=
~
Y
ga
_ _
hM
a
i hR
ga
i 171
where we have combined the (usually negligible) particlescale diffusive ux and
the turbulent uxes into an effectivediffusion coefcient (G
eff
).
As mentioned earlier, since the interfaces between phases are not resolved in
the CFD model, the Reynoldsaverage masstransfer terms (hM
a
i), and the
Reynoldsaverage reaction rates (hR
ga
i) in Eq. (171) must be modeled in terms
of known quantities. This situation is very much like classical reaction engi
neering models for multiphase reactors with the important difference that all
quantities are known locally. Such quantities include
a
g
_
;
Y
ga
;
Y
sa
;
T
g
;
T
s
;
U
g
;
U
s
and local multiphase turbulence statistics. Note that these variables, although
local, tell us nothing about the internal structure of the phases (i.e., subgrid
scale information). For example,
T
s
represents the Favreaverage temperature
of a solid particle consistent with the Favreaverage enthalpy of a single
particle.
9
If, as is often the case, the temperature varies strongly between the
center and outer surface of a particle, a SGS model will be required to account
for this effect on, for example, the chemical reactions. The principal advantage
of using the CFD model over a classical CRE model is thus the ability to
account for the effect of local uid dynamics (e.g., j
U
g
U
s
j) on the mass/heat
transfer rate between phases. In CFD codes, this is typically done by using
correlations for hM
a
i written in terms of the local Sherwood (or Nusselt)
number and particle Reynolds number (modied perhaps by a function of ha
g
i
to account for particleparticle interactions). In the case where largescale
mixing is innitely rapid, these correlations will reduce to the classical CRE
models for homogeneous multiphase reactors. However, such cases are rare
(and need not be modeled using CFD), and it is more likely that largescale
mixing will be rate limiting at certain locations within the reactor. Indeed, it is
exactly in such cases that CFD modeling will be of most benet for reactor
design and analysis.
9
This discussion also applies to the original variable T
s
, which represents the ensembleaverage
temperature of particles located at a particular point at a given time. Basically, we know the total
enthalpy of each particle, but we do not know how it is distributed inside any given particle. Since
the reaction rate can be very sensitive to the local temperature, we will need a SGS model to describe
the coupling between intraparticle transport processes and chemical reactions.
RODNEY O. FOX 298
C. COUPLING WITH CHEMISTRY
The species balances given above (Eqs. 164 and 165) include the reaction
source terms on the righthand sides. However, for these expressions to be
useful in CFD modeling, the user must supply the reaction rate functions and
the kinetic parameters. In addition, just as in singlephase turbulent reacting
ows (Fox, 2003), it may be necessary to account for micromixing effects on the
chemical kinetics by using SGS models in the Reynoldsaverage transport
equations (e.g., for hR
ga
i in Eq. (171)). For example, consider the parallel
reactions
AB !R
AC !S
with A in the gas phase and excess B and C in the liquid phase. If the rst
reaction is very rapid, then in the absence of micromixing effects in the liquid
phase only R would be produced. However, if A cannot be rapidly mixed into
the liquid phase (after mass transfer from the gas phase), then some S will be
produced as an unwanted byproduct. In many ways, this situation is analogous
to the reaction systems discussed in Section III.B. If we dene a mixture fraction
x that is unity in the gas phase (i.e., for pure A) and initially zero in the liquid
phase, then the degree of conversion of the rst reaction (Y
1N
) can be para
meterized by the value of x in the liquid phase (see Eq. 68). However, the rate of
the second reaction will depend on the local mixture fraction in a nontrivial
manner (see Eq. 70). Thus, if we simply ignore SGS uctuations in the liquid
mixture fraction we will likely severely underestimate the extent of the second
reaction. In theory, it is possible to write a presumed PDF transport model for
the mixture fraction in the liquid phase. However, unlike in singlephase tur
bulence, the source term for the mixture fraction in the liquid phase is the mass
transfer term and the sink term for mixturefraction uctuations will depend on
the rate of molecular mixing in different regions around the interphase (e.g.,
boundary layer, wake, far eld). A similar situation is encountered in spray
combustion (Reve illon et al., 2004) where evaporating liquid droplets act as
source terms for reactants in the gas phase. It thus may be useful to adapt SGS
models for spray combustion (at least the parts modeling the mixture fraction
mean and variance) to describe SGS mixing in more general settings such as
gasliquid ows. One can also use direct simulations of bubbly reacting ows
(Khinast, 2001; Khinast et al., 2003; Koynov and Khinast, 2004; Raffensberger
et al., 2005) to explore the validity of SGS models developed for twophase
reacting ows.
From the discussion above, we should keep in mind that even if no SGS
micromixing model is used to describe the multiphase ow, it may often be the
case that chemical reactions (and indeed micromixing) will be limited by mass/
heat transfer between the phases. Because the multiuid model (see Eqs. 164 and
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 299
165) includes masstransfer terms, the reaction rates will usually be mass trans
fer limited in cases where the chemical kinetics are fast. Thus, since we are
already relying on correlations to calculate the mass/heattransfer rates, it may
not be fruitful to try to include a detailed description of micromixing inside each
phase. Indeed, it will more likely be the case that improving the correlations
(e.g., to include the effect of chemical reactions on local mass transfer) will have
a greater impact on the accuracy of CFD model predictions.
Finally, to conclude our discussion on coupling with chemistry, we should
note that in principle fairly complex reaction schemes can be used to dene the
reaction source terms. However, as in singlephase ows, adding many fast
chemical reactions can lead to slow convergence in CFD simulations, and the
user is advised to attempt to eliminate instantaneous reaction steps whenever
possible. The question of determining the rate constants (and their dependence
on temperature) is also an important consideration. Ideally, this should be done
under laboratory conditions for which the mass/heattransfer rates are all faster
than those likely to occur in the productionscale reactor. Note that it is not
necessary to completely eliminate mass/heattransfer limitations to determine
usable rate parameters. Indeed, as long as the rate parameters found in the lab
are reliable under wellmixed (vs. perfectmixed) conditions, the actual mass/
heattransfer rates in the reactor will be lower, leading to accurate predictions of
chemical species under mass/heattransferlimited conditions.
VI. Conclusions and Future Perspectives
From the brief overview of CFD models presented in this work, the reader will
hopefully have gained an initial appreciation of the utility and power of CFD tools
for chemical reactor analysis and design. In Section II we have discussed the basic
formulation and specic steps needed to set up a CFD model. We have also
introduced the key concept of subgridscale (SGS) modeling and its importance in
describing unresolved phenomena in reactorscale CFD models. However, it is
worth repeating here that the development of SGS models for chemical reactions
and molecular transport is a natural extension of traditional chemical reaction
engineering modeling activities, and thus one of the key areas where chemical
engineers can have a large impact on the eld. In Section III we gave an example
of an SGS model developed for mixingsensitive chemical reactions. This simple
model for masstransferlimited chemical reactions can be easily extended to other
applications such as highSchmidtnumber laminar ows. In Section IV we dis
cussed efcient methods for adding a population balance equation to a CFD code
to model the production of ne particles. More generally, the ability to represent
the evolution of a population of entities (e.g., particles, bubbles, drops) in the
context of CFD results in a very powerful tool for describing complex reacting
ows. At present, methods based on the direct quadrature method of moments
RODNEY O. FOX 300
(DQMOM) are still in their initial stages of development. Nevertheless, they have
already been applied with great success by a growing number of researchers to a
wide variety of problems. The versatility of DQMOM was demonstrated in Sec
tion IV by applying it twice to the one modeling problem: rst to model micro
mixing between uid elements containing a bivariate number density function
(NDF), and then to represent the bivariate NDF itself. Finally, in Section V we
discussed the challenges associated with CFD models for multiphase reacting
ows. Although there are still a number of open problems to be solved in mul
tiphase ow CFD, when used with caution existing CFD models can be used for at
least qualitative analysis of chemical reactors.
In my opinion, the perspective for future developments in the eld of CFD
modeling of chemical reactors is quite strong. On the one hand, the continued
growth of computational power both through faster computers and better
algorithms will make it possible to solve more and more complex problems. On
the other hand, we are fortunate in this eld that the basic microscopic balance
equations are known, even though they lead to complex multiscale, multiphysics
phenomena. The accessibility of largescale computing facilities will allow us to
explore this complexity using direct simulations for specic academic pro
blems that can be used to test the SGS models needed for CFD simulations of
industrial reactors. In general, advances in the development of SGS models will
require collaboration between computational physicist/chemist working on di
rect simulation of academic problems and chemical reaction engineers develo
ping multiphysics models. Indeed, the reader should appreciate that it is almost
never the case that a reliable SGS model can be developed by simply analyzing
the results from a largescale direct simulation. Inversely, SGS model developed
without validation against detailed experimental or direct simulation data are
usually of limited value. Instead, the more fruitful approach is to rst develop a
tentative SGS model based on a preliminary understanding of the physics/
chemistry of the problem, and then to design a largescale direct simulation to
test key assumptions/predictions of the model. This dialogue between model
development and model validation is continued until a suitably reliable SGS
model is found. SGS models developed in this manner have a strong funda
mental underpinning and have a much greater chance of being applicable to a
wide range of operating conditions.
In summary, my recommendation for future progress in the eld is not to
follow the deceptively simple path of rushing toward the application of large
scale CFD simulations to complex industrial reactor systems if the basic SGS
models have not rst been shown to be reliable on academic problems.
10
Rather, I would recommend that we proceed more cautiously with adequate
attention given to the development of the fundamental physical understanding
required to develop reliable CFD models. While this path will obviously require
10
Despite this word of caution, one should not lose site of the fact that there are many industrial
reactor systems that can be accurately simulated with existing SGS models!
CFD MODELS FOR ANALYSIS AND DESIGN OF CHEMICAL REACTORS 301
patience and perseverance, in the long run it will undoubtedly be the surest way
to attain the chemical reaction engineers longsought goal of experimentfree
scaleup of chemical reactors.
ACKNOWLEDGMENTS
The authors work in the area of CFD analysis of chemical reactors has been
supported nearly continuously for the last 15 years by the U.S. National Science
Foundation. The work on gassolid multiphase ows and population balances
was funded by the U.S. Department of Energy. The author would also like to
acknowledge support from several companies, including Air Products and
Chemicals, BASF, BASELL, BP Chemicals, Dow Chemical, DuPont Engi
neering, and Univation Technologies. Last, but not least, the author wishes to
acknowledge his many collaborators over the years who are many in number to
name them individually.
REFERENCES
Agrawal, K., Loezos, P. N., Syamlal, M., and Sundaresan, S. J. Fluid Mech. 445, 151185 (2001).
Akhtar, M. K., Xiong, Y., and Pratsinis, S. E. AIChE J. 37, 15611570 (1991).
Aoun, M., Plasari, E., David, R., and Villermaux, J. Chem. Eng. Sci. 54, 11611180 (1999).
Baldyga, J. Chem. Eng. Sci. 49, 19852003 (1994).
Baldyga, J., Bourne, J. R., and Walker, B. Can. J. Chem. Eng. 76, 641649 (1998).
Batchelor, G. K. J. Fluid Mech. 193, 75110 (1988).
Bird, R. B., Stewart, W. E., and Lightfoot, E. N., Transport Phenomena. 2nd edn. John Wiley &
Sons, New York, USA (2002).
Briesen, H., Fuhrmann, A., and Pratsinis, S. E. Chem. Eng. Sci. 53, 41054112 (1998).
Bunner, B., and Tryggvason, G. J. Fluid Mech. 495, 77118 (2003).
Corrsin, S AIChE J. 10, 870877 (1964).
David, R., and Marcant, B. AIChE J. 40, 424432 (1994).
Davies, C. N. The sedimentation and diffusion of small particles. Proceedings of the Royal
Society of London. Series A, Mathematical and Physical Sciences. 200, 110113 (1949).
Drew, D. A., and Passman, S. L., Theory of Multicomponent Fluids. SpringerVerlag, Inc., New
York, USA (1999).
Dring, R. P. ASME J. Fluid Eng 104, 15 (1982).
Einstein, A. Annalen der Physik 17, 549560 (1905).
Elimelech, M., Gregory, J., Jia, X., and Williams, R. A., Particle Deposition and Aggregation,
Measurement, Modelling and Simulation. ButterworthHeinemann, Woburn (1995).
Fan, L. S., and Zhu, C., Principles of GasSolid Flows. Cambridge University Press, New York
(1998).
Fan, R., Marchisio, D. L., and Fox, R. O. Powder Technol. 139, 720 (2004).
Fiorina, B., Gicquel, O., Vervisch, L., Carpentier, S., and Darabiha, N. Combust. Flame 140,
147160 (2005).
Fox, R. O. Chem. Eng. Process. 37, 521535 (1998).
RODNEY O. FOX 302
Fox, R. O., Computational Models for Turbulent Reacting Flows. Cambridge University Press,
Cambridge, UK (2003).
Fox, R. O., and Raman, V. Phys. Fluids 16, 45514564 (2004).
Fox, R. O., and Yeung, P. K. Phys. Fluids 15, 961985 (2003).
Friedlander, S. K., Smoke, Dust, and Haze. 2nd edn Oxford University Press, New York, USA
(2000).
Fuchs, N. A., The Mechanics of Aerosols. Pergamon Press, New York, USA (1964).
Gao, D., Fan, R., Subramaniam, S., Fox, R. O., and Hoffman, D. K., J, Fluid Eng. 128, 6268
(2005).
Garnier, C., Lance, M., and Marie , J. L. Exp. Therm. Fluid Sci. 26, 811815 (2002).
Garside, J., and Tavare, N. S. Chem. Eng. Sci. 40, 14851493 (1985).
Harteveld, W. K., Mudde, R. F., and Van Den Akker, H. E. A. Can. J. Chem. Eng. 81, 389394
(2003).
Jackson, R., The Dynamics of Fluidized Particles. Cambridge Univer