Sei sulla pagina 1di 15

Chemical Engineering Science 58 (2003) 33373351

www.elsevier.com/locate/ces
Implementation of the quadrature method of moments in CFD
codes for aggregationbreakage problems
Daniele L. Marchisio

, R. Dennis Vigil, Rodney O. Fox


Department of Chemical Engineering, Iowa State University, 2114 Sweeney Hall, Ames, IA 50011-2230, USA
Received 10 January 2003; received in revised form 7 April 2003; accepted 11 April 2003
Abstract
In this work the quadrature method of moments (QMOM) is implemented in a commercial computational uid dynamics (CFD)
code (FLUENT) for modeling simultaneous aggregation and breakage. Turbulent and Brownian aggregation kernels are considered in
combination with dierent breakage kernels (power law and exponential) and various daughter distribution functions (symmetric, erosion,
uniform). CFD predictions are compared with experimental data taken from other work in the literature and conclusions about CPU time
required for the simulations and the advantages of this approach are drawn.
? 2003 Elsevier Ltd. All rights reserved.
Keywords: Population balance; Computational uid dynamics; Quadrature method of moments; Aggregation; Breakage; Particulate systems
1. Introduction
Aggregation and breakage play an important role in a
number of important chemical processes such as precipita-
tion, crystallization, separation processes, and reaction in
multiphase systems. Modeling and simulation of these pro-
cesses is complicated due to the diculties inherent in de-
scribing the evolution of a distribution of particle sizes and
because of the incomplete understanding of the mechanisms
by which aggregation and breakage occur, including the role
of hydrodynamics. This latter problem is often neglected,
despite considerable evidence that aggregation is strongly
inuenced by mixing. For example, Brown and Glatz
(1987) investigated the eect of the operating conditions on
particle size established during breakage of protein particles
prepared under isoelectric precipitation in an agitated ves-
sel. They found that the aggregation rate increases with both
particle concentration and shear rate. In another study, the
inuence of the type of ow on the aggregation rate was in-
vestigated using a TaylorCouette reactor, a pipe-ow reac-
tor, and a at-bottomed tank reactor (Krutzer, van Diemen,
& Stein, 1995). These experiments showed that at equal

Corresponding author. Tel.: 515-294-3186; fax: 515-294-2689.


E-mail address: marchis@iastate.edu (D. L. Marchisio).
energy dissipation rates, the aggregation rate is higher for
isotropic turbulent ows than for non-isotropic ows.
Raphael and Rohani (1996) investigated the eect of ag-
gregation on the particle size distribution (PSD) during sun-
ower protein precipitation and found that the maximumsize
of the aggregates is determined by the hydrodynamics of the
reactor and the mean residence time of the particles in the re-
actor. Serra, Colomer, and Casamitjana (1997) investigated
aggregation and breakup of particles in a TaylorCouette
reactor. In their experimental work they analyzed the eect
of particle concentration, shear rate, and particle initial di-
ameter. Their results showed the existence of three regions
determined by particle concentration and type of ow es-
tablished (laminar or turbulent). Moreover, they found that
the nal aggregate diameter in the turbulent regime is inde-
pendent of monomer size and is instead controlled by the
Kolmogorov micro-scale, whereas in laminar ow the nal
mean particle size decreases as the diameter of the primary
particles is increased.
One means for characterizing the morphology of clus-
ters of particles formed by aggregationbreakage processes
is the fractal dimension, which provides an indication of
the compactness of aggregates. The relationship between
fractal dimension and physical properties has been studied
from both the experimental (Serra & Casamitjana, 1998a),
and theoretical viewpoints (Filippov, Zurita, & Rosner,
2000; Jlang & Logan, 1991). Hansen and co-workers used
0009-2509/03/$ - see front matter ? 2003 Elsevier Ltd. All rights reserved.
doi:10.1016/S0009-2509(03)00211-2
3338 D. L. Marchisio et al. / Chemical Engineering Science 58 (2003) 33373351
a TaylorCouette reactor to study orthokinetic aggregation
for monodisperse and bidisperse colloidal systems (Hansen,
Malmsten, Bergenstahl, & Bergstrom, 1999). Particle ag-
gregation was investigated by direct observation using a
CCD camera, and the observed aggregates were character-
ized by a high fractal dimension, suggesting that clusters
are rearranged and densied by the shear.
Aggregation and breakage are often the last steps of a
complex sequence of phenomena, such as nucleation, fast
reactions, combustion, and molecular growth (Marchisio,
Barresi, & Garbero, 2002; Rosner & Pyykonen, 2002;
Baldyga & Orciuch, 2001). Such systems inevitably lead
to non-negligible spatial heterogeneities, and therefore a
method of modeling these processes that accounts for hy-
drodynamics is crucial for predicting reactor performance.
One approach to account for non-ideal mixing is to use
CFD methods. In such an approach the reactor is represented
by a computational grid and the continuity and Navier
Stokes equations are solved over the computational domain.
When dealing with turbulent ows, the set of equations is
unclosed and turbulence models are used to solve the clo-
sure problem. In addition to these equations, the population
balance equation for the solid phase has to be solved.
The population balance is a continuity statement writ-
ten in terms of the PSD and has consistently received
attention since Smoluchowski (1917) introduced the math-
ematical formalism nearly a century ago. A compre-
hensive overview of the mathematical issues involved,
the numerical methods available, and possible develop-
ments for the future have been given by Ramkrishna
(1985, 2000) and Ramkrishna and Mahoney (2002).
A general form of the mean-eld population balance for a
spatially extended systemcan be written as follows (repeated
indices implies summation):
cn(; x, t)
ct
+u
i

cn(; x, t)
cx
i

c
cx
i
_
(I
t
+ I)
cn(; x, t)
cx
i
_
=
c
c
)
[n(; t)
)
] + h(; t), (1)
where (
1
, . . . ,
n
) is the property vector that species
the state of the particle, n(; x, t) is the number density func-
tion, u
i
is the Reynolds-average velocity in the ith direc-
tion, x
i
is the spatial coordinate in the ith direction, I is the
molecular diusivity and I
t
is the turbulent diusivity. For
turbulent ows I
t
I and thus it is commonly assumed
that I
t
+ I I
t
. The ux in -space is denoted by

)

d
)
dt
, ) 1, . . . , N, (2)
and h(; t) represents the net rate of introduction of new
particles into the system (Hulburt & Katz, 1964).
The main problem in solving the above equation is the
presence of the extra variables
i
, which dene the particle
size, shape, etc. In most CFD codes it is possible to intro-
duce user-dened scalars by using user-dened subroutines,
but these scalars must only be functions of time and space.
Hence, in order to reduce the dimension of the problem,
several methods have been developed.
The discretized population balance (DPB) approach is
based on the discretization of the internal coordinates (i.e.,
the components of the property vector). A detailed compar-
ison of the performance of the most popular DPB methods
has been carried out by Vanni (2000a). The principle ad-
vantage of the DPB method is that the PSD is calculated di-
rectly. However, in order to maintain reasonable accuracy,
a large number of scalars (i.e., particle classes) are required.
As a consequence, the DPB approach is computationally in-
tractable for spatially heterogeneous systems and therefore
not suitable for CFD applications.
An alternative to PBE approaches is to implement a
stochastic analog via a Monte Carlo algorithm (Smith &
Matsoukas, 1998; Lee & Matsoukas, 2000; Rosner & Yu,
2001). These methods have the advantage of satisfying
mass conservation as well as correctly accounting for uc-
tuations that arise as the system mass accumulates in a small
number of large aggregates. However, incorporation of
these methods into CFD codes is also not computationally
tractable because of the large number of scalars required.
In contrast to the DPB or stochastic approaches, the mo-
ment method (MM) is suitable for use with CFD codes be-
cause the internal coordinates are integrated out such that
solution only requires a small number of scalars (i.e., 4
6 moments of the PSD) at each grid point. Of course the
vast reduction in the number of scalars, which makes im-
plementation in CFD codes feasible, comes at the cost of a
less-detailed description of the PSD. However, provided that
the PSD function is suciently simple (e.g., monomodal or
bimodal), a low-order moment description may be sucient.
The method was rst proposed many years ago by Hulburt
and Katz (1964), but it has not found wide applicability due
to the diculty of expressing the transport equations for the
moments of the PSD in terms of the moments themselves.
More recently, several approaches for contending with this
closure problem have been developed, and a discussion
of these can be found in Diemer and Olson (2002).
When the population balance is written in terms of one
internal coordinate (e.g., particle length or particle volume)
the closure problem has been successfully solved with the
use of a quadrature approximation (McGraw, 1997), where
weights and abscissas of the quadrature approximation can
be found by using the product-dierence (PD) algorithm de-
scribed by Gordon (1968). The so-called quadrature method
of moments (QMOM) has been validated for several prob-
lems (e.g., molecular growth, aggregation, breakage) and
by using dierent internal coordinates (Marchisio, Pikturna,
Fox, Vigil, & Barresi, 2003a; Barret & Webb, 1998). More-
over, the method has been extended to the study of aerosol
D. L. Marchisio et al. / Chemical Engineering Science 58 (2003) 33373351 3339
dynamics by using population balances with two internal
coordinates: particle volume and surface area (Wright, Mc-
Graw, & Rosner, 2002).
The main purpose of this work is to implement the
QMOM approach to solving the population balance equa-
tion in a commercial CFD code and to verify the feasibility
of these calculations for practical applications. The spe-
cic problem considered is aggregation and breakage in
turbulent TaylorCouette ow.
The CFD-based predictions for the particle size distribu-
tions are generated by using several dierent combinations
of conventional expressions for the aggregation and breakup
kernels and are compared with experimental data from Serra
et al. (1997) and Serra and Casamitjana (1998a, b). The
paper is organized as follows: governing equations are dis-
cussed, aggregation and breakage models are presented, and
the results of the case study considered, particularly with re-
spect to the number of scalars involved in the calculations,
the quality of the predictions, and the CPU time in compar-
ison with alternative approaches.
2. Governing equations
In this section the QMOM is presented and explained.
Particular attention will be devoted to its implementation in
FLUENT.
2.1. Quadrature method of moments
In this work, we consider a number density function de-
ned in terms of particle length (
1
L). The resulting
population balance is (Randolph & Larson, 1988)
cn(L; x, t)
ct
+u
i

cn(L; x, t)
cx
i

c
cx
i
_
I
t
cn(L; x, t)
cx
i
_
=
c
cL
[G(L)n(L; x, t)] + B(L; x, t) D(L; x, t), (3)
where G(L) is the growth rate, and B(L; x, t) and D(L; x, t)
are, respectively, the birth and death rates due to aggrega-
tion and breakage. The moments of the PSD are dened as
follows:
m
k
(x, t) =
_

0
n(L; x, t)L
k
dL, (4)
and thus the transport equation for the kth moment is
cm
k
(x, t)
ct
+u
i

cm
k
(x, t)
cx
i

c
cx
i
_
I
t
cm
k
(x, t)
cx
i
_
=(0)
k
J(x, t) +
_

0
kL
k1
G(L)n(L; x, t) dL
+B
a
k
(x, t) D
a
k
(x, t) + B
b
k
(x, t) D
b
k
(x, t), (5)
where J(x, t) is the nucleation rate and where
B
a
k
(x, t) =
1
2
_
+
0
n(z; x, t)
_
+
0
[(u, z)
(u
3
+ z
3
)
k}3
n(u; x, t) du dz, (6)
D
a
k
(x, t) =
_
+
0
L
k
n(L; x, t)
_
+
0
[(L, z)
n(z; x, t) dz dL, (7)
B
b
k
(x, t) =
_
+
0
L
k
_
+
0
a(z)b(L|z)n(z; x, t) dz dL, (8)
D
b
k
(x, t) =
_
+
0
L
k
a(L)n(L; x, t) dL (9)
are the moments of the birth and death rates, respectively,
due to aggregation and breakage. The aggregation kernel
[(L, z) represents the rate coecient for aggregation of two
particles with lengths L and z whereas the breakage kernel
a(L) denes the rate coecient for breakage of a particle of
length L. The fragment distribution function for the break-
age of a particle of size L is given by b(L|z). A detailed
derivation of these equations can be found elsewhere (see
Marchisio, Vigil, & Fox, 2003b); discussion of the choice
of aggregation and breakage kernels can be found in
the following sections.
The QMOM is based on the following quadrature approx-
imation
m
k
=
_
+
0
n(L)L
k
dL
N
d

i=1
w
i
L
k
i
, (10)
where weights (w
i
) and abscissas (L
i
) are determined
through the product-dierence (PD) algorithm from the
low-order moments (Gordon, 1968); a detailed explanation
of the method can be found in Appendix A. By using the
PD algorithm, a quadrature approximation with N
d
weights
and N
d
abscissas can be constructed using the rst 2N
d
moments of the PSD. For example, if N
d
= 3 only the rst
six moments (m
0
, . . . , m
5
) are tracked, and the quadrature
approximation is given by
m
k
=
N
d
=3

i=1
w
i
L
k
i
= w
1
L
k
1
+ w
2
L
k
2
+ w
3
L
k
3
. (11)
Knowledge of w
i
and L
i
suces to close the transport equa-
tions for the moments, and in the specic case of aggrega-
tion and breakage without nucleation and molecular growth,
the moment equations become
cm
k
(x, t)
ct
+u
i

cm
k
(x, t)
cx
i

c
cx
i
_
I
t
cm
k
(x, t)
cx
i
_
=
1
2

i
w
i

)
w
)
(L
3
i
+ L
3
)
)
k}3
[
i)
3340 D. L. Marchisio et al. / Chemical Engineering Science 58 (2003) 33373351

i
L
k
i
w
i

)
[
i)
w
)
+

i
a
i

b
(k)
i
w
i

i
L
k
i
a
i
w
i
, (12)
where [
i)
= [(L
i
, L
)
), a
i
= a(L
i
), and

b
(k)
i
=
_
+
0
L
k
b(L|L
i
) dL. (13)
It has been shown (Marchisio et al., 2003a, b; McGraw,
1997; Barret & Webb, 1998) that by using this quadrature
approximation it is possible to track the moments of the PSD
with very high accuracy.
2.2. Aggregation kernel
Aggregation of solids suspended in a uid is a complex
phenomena that in general depends upon particleparticle
and particleuid interactions, particle morphology, as well
as uid mixing. However, regardless of the specic mech-
anism by which aggregation occurs, the process can be di-
vided up into two steps: (1) particleparticle collision and
(2) sticking or fusion of colliding particles. The aggre-
gation kernel can then be expressed as the product of the
sticking coecient and a collision frequency function.
In certain simplied cases, the aggregation kernel can
be computed exactly, provided that certain assumptions are
satised. For example, if long-range particleparticle forces
are absent, if particles are Euclidean (as opposed to fractal),
and if the particles are suciently small so that they neither
inuence the uid phase nor are disturbed by uid shear, then
the collision frequency may be described by the Brownian
kernel:
[(L, z) =
2k
B
1
3j
(L + z)
2
Lz
, (14)
where k
B
is Boltzmanns constant, 1 is the absolute temper-
ature, j is the uid viscosity, and L and z are particle sizes,
as rst proposed by Smoluchowski (1917).
When particles are suciently large compared to shear
gradients, particleparticle collision frequencies are inu-
enced by uid motion. In the case of turbulent ow, the ra-
tio between particle size and the Kolmogorov micro-scale is
of primary importance. When particles are smaller than the
Kolmogorov micro-scale the aggregation kernel can be com-
puted as follows (Adachi, Cohen Stuart, & Fokkink, 1994):
[(L, z) =
4
3
_
3
10
_
1}2 _
c
v
_
1}2
(L + z)
3
, (15)
where c is the turbulent dissipation rate and v is the kinematic
viscosity. In fact, by using Taylors statistical analysis of
isotropic turbulence and assuming a normal distribution of
the velocity gradient, it can be shown that the shear rate
produced by uctuating turbulent velocities is proportional
to
_
c}v. Saman and Turner (1956) used this result to ex-
plain the formation of drops in clouds, and Tontrup, Gruy,
and Cournil (2000) used it for modeling turbulent aggrega-
tion of titania particles in water.
For chemically inert systems, aggregation eciency
is primarily determined by the balance between attrac-
tive (e.g. Van der Waals) and repulsive (e.g. electrical
double layer) forces. This balance is strongly inuenced by
the ionic strength of the solution. When the ionic strength
increases, the double layer repulsion is reduced, and at very
high values it is practically negligible. In these conditions
the aggregation eciency is usually considered to be equal
to one. Aggregation eciency can also be related to the
probability of fragmentation of newly formed aggregates.
In fact, the fragmentation kernels presented in the next
section are suitable for breakage of particles that have had
sucient time to restructure after collision (Gruy, 2001). If
a new aggregate does not have sucient time for restructur-
ing and is fragmented into its original two components, it
can be referred to as null aggregation eciency (Brakalov,
1987). Another important factor in determining the aggre-
gation eciency in precipitating particle system is the rate
of formation of bridges in zones of high supersaturation
(Baldyga, Jasinska, & Orciuch, 2002; Hounslow, Mumtaz,
Collier, Barrick, & Bramley, 2001). In the present work,
we consider only aggregation and breakage of particles in
chemically inert systems.
2.3. Breakage
Particle breakage functions can be factored into two parts.
The breakage kernel, a(L), is the rate coecient for breakage
of a particle of size L, and b(z|L) denes the probability
that a fragment of size z is formed from the breakage of an
L-sized particle. Several expressions for the breakage kernel
relevant to hydrodynamic breakage mechanisms have been
derived by dierent authors and they are summarized in
Table 1.
Breakage kernel (1) in Table 1 was derived for disruption
of disperse liquid droplets in turbulent ow under the hy-
pothesis of local isotropy and assuming that the droplet size
is within the inertial subrange. Abreakage event occurs if the
kinetic energy transmitted by an eddy to a droplet is greater
than the droplet surface energy. The breakup frequency is
derived by considering the fraction of eddies present in the
system with sucient kinetic energy to cause breakage. This
can be done by estimating the fraction of drops breaking and
by estimating the characteristic time required for breakage
(t
b
) through the second-order structure function of the uid
velocity D
LL
(r). This function represents the covariance of
the dierence in velocity dierences between two points at
a given distance r and according to turbulent theory (Frisch,
1995) it can be written as
D
LL
(r) = c
2
(cr)
2}3
, (16)
D. L. Marchisio et al. / Chemical Engineering Science 58 (2003) 33373351 3341
Table 1
Breakage kernels proposed in literature
System a(L) Reference
(1) Liquidliquid a(L) = k
1
L
2}3
c
1}3
exp
_

k
2
o
j
d
c
2}3
L
5}3
_
Coulaloglou and Tavlarides (1977)
(2) Liquidliquid a(L) = [
z
erfc
_
3.5(
L
L
s
)
5}6
_
Narsimhan, Gupta, and Ramkrishna (1979)
(3) Liquidliquid a(L) = c
4
_
c
L
2
_
1}3 _
1

min
(1+)
2

11}3
exp
_

12c
[
o
[j
d
c
2}3
L
5}3

11}3
_
d Luo and Svendsen (1996)
(4) Solidliquid a(L) =
1

15
(
c
v
)
1}2
exp
_

t
[
j(c}v)
1}2
_
Ayazi Shamlou, Stravrinides, Titchener-Hooker, and Hoare (1994)
(5) Solidliquid a(L) = c
1
v
:
c
[
L

Luo and Svendsen (1996)


Kramer and Clark (1999)
Wojcik and Jones (1998)
where c
2
is a numerical constant and c is the turbulent dis-
sipation rate. Assuming that the motion of the daughter
droplets is similar to the relative motion of two lumps of
uid in a turbulent ow and that the distance r is the droplet
diameter, we can write that
_
L
t
b
_
2
D
LL
(L) = c
2
(cL)
2}3
, (17)
which can be solved for t
b
. The nal expression for a(L) is
a function of the surface tension o and the droplet density
j
d
and is reported in Table 1 [see breakage kernel (1)].
Breakage kernel (2) in Table 1 was derived by consid-
ering the frequency that eddies arrive at the surface of a
disperse phase droplet, whose maximum stable diameter is
indicated by L
s
. Under the same phenomenological simpli-
cations of the above mentioned models, and considering the
dierent contributions of the population of eddies in the in-
ertial subrange, it is possible to derive an expression for the
breakage rate without unknown constants [see kernel (3) in
Table 1], where = p}L, p is the eddy length-scale, c
4
is a
constant of order unity (c
4
=0.92), [=3}2, c
[
is the increase
coecient in the surface area [c
[
= [
2}3
+ (1 [)
2}3
1],
and [ is the volume fraction of one of the two fragments
(the function is symmetric about [ = 0.5).
As with breakage of liquid droplets, breakage of solid ag-
gregates depends strongly on the ratio between the particle
size and the smallest turbulent eddy. When the size of a par-
ticle is greater than the size of the smallest eddies, break-
age is likely to occur by instantaneous normal stresses due
to pressure uctuations acting on the surface of the parti-
cle. For particle sizes smaller than the turbulent micro-scale,
breakage is likely to be caused by shear stresses originat-
ing from the turbulent dynamic velocity dierences acting
on the opposite sides of the particle. Under these conditions
the breakage kernel can be written as a function of the ag-
gregate strength (t
[
), which can be estimated as follows:
t
[
=
9
8
k
c
[F
1
L
2
o
, (18)
where L
o
is the diameter of a primary particle, k
c
is the co-
ordination number, and F is the inter-particle force between
two primary particles (Ayazi Shamlou et al., 1994). The re-
sulting breakage kernel is reported in Table 1 [see kernel
(4)]. The inter-particle force F can be computed as follows:
F =
AL
o
12H
2
o
, (19)
where A is the Hamaker constant for the liquid-particle sys-
tem and H
o
is the distance between two primary particles.
The coordination number is based on experimental observa-
tion and can be calculated as
k
c
15[
1.2
, (20)
where [is the volume fraction of solid within the aggregates.
This quantity can be determined once the fractal dimension
d
[
of the aggregates in known
[(L) = C
_
L
L
o
_
d
[
3
, (21)
where C = 0.414d
[
0.211 (Vanni, 2000b).
Starting from the same basic considerations, but using a
simplied approach, Kramer and Clark (1999) developed
another hydrodynamic breakage model for solid aggregates.
They classied failure modes by dening manifestation, in-
duction and location of the failure, and stated that aggregates
break when the maximum eigenvalue of the stress tensor is
greater than the aggregate strength (t
[
). Using the assump-
tion that the number of aggregate bonds with a strength at or
below the failure strength t
[
is not a linear function of the
strain rate and that size-dependency should also be consid-
ered in a breakup model, the resulting form of the breakage
kernel is the one given as expression 5 in Table 1.
By invoking semi-theoretical considerations and tting to
experimental data, several values of the exponent have
been proposed = 0, 1}3, 2}3, 1 (Luo & Svendsen, 1996).
Pandya and Spielman (1982) found that =1, whereas Peng
and Williams (1994), by tting the model to experimental
3342 D. L. Marchisio et al. / Chemical Engineering Science 58 (2003) 33373351
Table 2
Fragment distribution functions: L
o
is the size of the primary particles
Mechanism b(L|z)

b
(k)
i
Symmetric
fragmentation
_
_
_
2 if L =
z
2
1}3
0 otherwise
2
(3k)}3
L
k
i
Erosion
_

_
1 if L = L
o
1 if L = (z
3
L
3
o
)
1}3
0 otherwise
L
k
o
+ (L
3
i
L
3
o
)
k}3
Uniform
_
_
_
6L
2
z
3
if 0 L z
0 otherwise
L
k
i
6
k+3
data, found that can assume values ranging between 1 and
3 and that [ = 1}2.
Serra and Casamitjana (1998b) t experimental data and
found that the relationship between the breakage kernel and
the turbulent dissipation rate c depends on the solid frac-
tion. When the solid fraction is low [ = 0.9, whereas for
higher values of the solid fraction [ is much higher. This
can be explained by the fact that particleparticle collisions
are more eective if the solid fraction is higher. Their results
conrm ndings by Spicer and Pratsinis (1996), who used
data from Oles (1991) and found [ = 0.8 for low values of
solid concentration.
In their investigation of nucleation, growth, agglomera-
tion and disruption kinetics for calcium oxalate, Zauner and
Jones (2000) found a linear dependency between the break-
age rate and the turbulent dissipation rate c (i.e., [ =1) and
the breakage kernel was size-independent (=0). In another
work (Wojcik & Jones, 1998) using a size-dependent break-
age kernel, a linear relationship between the kernel and the
turbulent dissipation rate ([ = 1) was also found.
The form of the fragment distribution function, b(L|z),
depends upon many factors including the particle properties,
such as strength and morphology, as well as the breakup
mechanism. Because the fragment distribution function is
likely to be highly system-specic, we consider a variety of
possibilities suggested by experiments on disruption kinetics
of animal cell aggregates undergoing hydrodynamic break-
age (Moreira, Cruz, Santana, Aunins, & Carrondo, 1995).
Some of these distribution functions include erosion of pri-
mary particles, formation of two equal fragments, formation
of two fragments with xed mass ratio (e.g., one to four),
and uniformdistribution of fragments. The distribution func-
tions corresponding to these cases are reported in Table 2.
3. Computational details and case study
The experimental data used in this study are taken from
Serra et al. (1997) and Serra and Casamitjana (1998a, b).
The reactor used in their work is a TaylorCouette device,
which consists of a uid conned to the annular region be-
tween two concentric cylinders (outer cylinder stationary,
inner cylinder rotating). The wetted diameters of the inner
and outer cylinders are d
1
=193 mm and d
2
=160 mm, re-
spectively, thereby giving an annular gap D=(d
1
d
2
)}2=
16.5 mm. The reactor length is H =360 mm, resulting in an
aspect ratio I
A
= H}D 22. Latex particles were used in
the experiments, and in order to avoid sedimentation eects,
the density of the solution was set at the same value as that
of the latex particles (j
s
=1.055 g}cm
3
) by adding an inor-
ganic salt to the solution (KCl). Samples were continuously
taken from the bottom of the reactor and passed through the
optical cell of a particle analyzer, and then pumped back
to the top of the reactor. The ux in the tubes was laminar
so that disruption of aggregates formed in the reactor could
be minimized. The residence time of the aggregates in the
tubes was also kept short in comparison with the time spent
in the reactor (about 300 times shorter).
Although aggregationbreakage experiments were car-
ried out with inner cylinder rotational speeds in the range
40200 rpm, we compare simulation results only with data
in the turbulent owregime (N=80200 rpm). An overview
of the various hydrodynamic regimes in the reactor can be
found in the review by Kataoka (1986), whereas validation
of ow eld predictions in turbulent regimes has been car-
ried out in our previous work (Marchisio, 2002).
In the absence of an applied axial ow, the hydrodynam-
ics in a TaylorCouette reactor depend upon the azimuthal
Reynolds number
Re =
r
1
c
1
D
v
, (22)
where r
1
= d
1
}2 is the radius of the inner cylinder, c
1
is
the angular velocity of the inner cylinder, D is the annular
gap, and v is the kinematic viscosity of the uid. When the
rotational speed of the inner cylinder is very small, circular
Couette ow is established. If the rotational speed of the in-
ner cylinder is increased past a critical value, this circular
ow becomes unstable and laminar toroidal vortices (Tay-
lor vortices) are formed. The critical azimuthal Reynolds
number for Taylor instability Re
c
, also depends upon the
specic reactor geometry and can be calculated by using
well known correlations (Kataoka, 1986). Other instabilities
occur at higher values of Re, and the azimuthal Reynolds
number ratio Re}Re
c
, can be used to determine which hy-
drodynamic resides in the reactor.
In this work, the ow eld in the reactor was modeled
by using the commercial CFD software FLUENT 6.0. Since
the solid particles are smaller than 2030 m, and the solid
concentration is smaller than 0.1%, the inuence of the solid
phase on the ow eld can be neglected and single-phase
turbulence models can be used. Thus, the Reynolds Stress
Model with standard wall function was used to model the
ow eld. In our previous work (Marchisio, 2002) this com-
bination of turbulence model and near-wall treatment was
D. L. Marchisio et al. / Chemical Engineering Science 58 (2003) 33373351 3343
Fig. 1. Sketch of the TaylorCouette reactor and of the computational
domain used in the simulation (grey section).
found to give the best agreement with experimental data.
Simulations were carried out in two dimensions under the
hypothesis of axial-symmetry. Moreover, since the ow was
found to be symmetrical about the center of the reactor, only
one half of this section has been modeled.
A sketch of the reactor and of the computational do-
main used in the simulations is shown in Fig. 1. Several
grids were tested in order to verify that the solution was
grid-independent, and it was found that a grid with 18 nodes
in the radial direction and 181 in the axial direction for a to-
tal of 3060 computational cells was sucient. In our previ-
ous work (Marchisio, 2002) a comparison between 2D and
3D simulations was also carried out. It was found that with-
out axial ow and in a range of operating conditions similar
to this work no appreciable dierence was detected between
2D and 3D predictions.
The QMOM was implemented with three nodes, and as a
consequence the rst six moments were tracked. This choice
is justied by previous ndings (Marchisio et al., 2003b)
where N
d
= 3 was found to be the best conguration in
terms of accuracy and computational costs. The transport
equations for the rst six moments were implemented by
using compiled user-dened functions, and details can be
found in Fluent Inc. (2002).
The quadrature approximation can be determined by
nding the roots of the polynomial of order N
d
belonging
to the sequence of orthogonal polynomials with respect
to the measure n(L) (see for details Dette & Studden,
1997). Since root nding of a polynomial is a notori-
ously ill-conditioned problem, weights and abscissas of the
quadrature approximation can be calculated by using the
PD algorithm. Following this procedure a tridiagonal matrix
of rank N
d
is calculated from the rst 2N
d
moments, and
the eigenvalues and the square of the rst component of
the eigenvectors of this matrix are respectively the abscis-
sas (L
i
) and weights (w
i
) of the quadrature approximation.
The eigen-value/eigen-vector problem was implemented in
FLUENT via user-dened-subroutine by using the C sub-
routine TQLI (Press, Teukolsky, Vetterling, & Flannery,
1992).
The computational procedure was as follows. First, the
ow eld was solved until a steady-state solution was
reached. The convergence criteria required that all normal-
ized residuals be smaller than 10
6
. Next, the moments were
determined (m
0
, . . . , m
5
). Their initial values were dened to
be uniform throughout the computational domain and then
a time-dependent simulation was initiated with a xed
time step of 10 s. The convergence criteria for all scalars
required that the normalized residuals be smaller than 10
6
.
Several cases were investigated using dierent combina-
tions of breakage kernels and fragment distribution func-
tions. In all the cases investigated, aggregation was modeled
using an aggregation kernel consisting of the summation of
the Brownian and hydrodynamic kernels reported in Eqs.
(14) and (15). As already mentioned, experiments were con-
ducted in saline solutions in order to match the uid density
with that of the latex particles. Since the double-layer re-
pulsion is very sensitive to the value of ionic strength [and
at high salt concentration is practically negligible (Serra
et al., 1997)], the aggregation eciency can be consid-
ered to be unity. Knowledge of the solid volume fraction
[
t
=2.510
5
and of the diameter of the latex primary par-
ticles L
o
=210
6
m is sucient for determining the initial
conditions of the six moments (m
0
=3.12510
12
m
3
; m
1
=
6.25 10
6
m
2
; m
2
= 12.5 m
1
; m
3
= 2.5 10
5
; m
4
=
5.0 10
11
m; m
5
= 1.0 10
16
m
2
). A summary of the
aggregation and breakage functions used in each of the dif-
ferent cases is reported in Table 3.
The exponents of the power-law breakage kernel [see
cases (1, 2, 3) in Table 3 or kernel (5) in Table 1] were found
through a dimensional analysis, forcing a linear dependence
for the particle size (i.e., =1). The resulting exponent for
the turbulent dissipation rate ([ = 3}4) is in good agree-
ment with the value found by tting with experimental data
([ =0.51). Moreover, direct experimental evidence seems
to conrm the linear dependence with respect to the particle
size ( = 1). As concerns the exponential breakage kernel
we rst need to identify some quantities such as the size of
3344 D. L. Marchisio et al. / Chemical Engineering Science 58 (2003) 33373351
Table 3
Cases tested for the comparison with experimental data
Case a(L) b(L|z)
(1) c
1
c
3}4
v
5}4
L Symmetric fragmentation
(2) c
1
c
3}4
v
5}4
L Erosion
(3) c
1
c
3}4
v
5}4
L Uniform
(4) c
A
c
1}2
exp
_
c
B
[(L)
2.2
c
1}2
_
Symmetric fragmentation
(5) c
A
c
1}2
exp
_
c
B
[(L)
2.2
c
1}2
_
Erosion
(6) c
A
c
1}2
exp
_
c
B
[(L)
2.2
c
1}2
_
Uniform
For all cases investigated, L
o
= 2 10
6
m and the solid volume
fraction [
t
= 2.5 10
5
.
the primary particles (L
o
=2 10
6
m), the fractal dimen-
sion of the particles d
[
=2.6 (Serra & Casamitjana, 1998a;
Sonntag & Russel, 1986), the value of the Hamaker constant
(A=10
21
J, for details see Fitch, 1997), and the interparti-
cle distance that can be approximated as H
o
=2410
9
m.
Hence, the breakage constants [compare kernel (4) in Table
1 and cases (4, 5, 6) is Table 3] are given by c
A
=1}

12v
258.2 and c
B
= (t
[
v
1}2
)}(j[
2.2
) 17.9. However, since A
and H
o
are properties of the system that are dicult to esti-
mate accurately, the constant c
B
will be expressed in terms
of the parameter group A}H
2
o
as follows
c
B
= 2.24 10
5
A
H
2
o
. (23)
As already mentioned, although the numerical value of the
group A}H
2
o
can be calculated by using values of A and H
o
found in the literature (Ayazi Shamlou et al., 1994; Fitch,
1997), the eect of changes in A}H
2
o
on the nal mean aggre-
gate size will be investigated. It should be highlighted here
that this approach presents an inconsistency due to the fact
that the QMOM in the form presented in this work, assumes
non-fractal aggregates. However, since these aggregates are
compact (i.e., d
[
is very close to 3), the application of the
QMOM (assuming d
[
=3) and the calculation of the break-
age kernel (assuming d
[
= 2.6) seems acceptable. Indeed,
the introduction of the fractal dimension as an independent
parameter in the QMOM formulation represents one of the
next steps of our work, but it requires the formulation of the
population balance using at least two internal coordinates.
In Fig. 2 breakage frequency rates as a function of particle
size for the exponential and power-law kernels are shown
for two rotational speeds, N=125 and 165 rpm (Re}Re
c
=96
and 127), which correspond to volume-averaged turbulent
dissipation rates c = 0.035 and 0.070 m
2
s
3
, respectively.
It is evident that the exponential kernel exhibits a much
stronger dependency on particle size. In fact, it scales with
the sixth power for low c and with the third power for high c.
L/L
o
0 10 20 30 40
N = 165 rpm
N = 125 rpm
L/L
o
a
(
L
/
L
o
)
0 10 20 30 40
10
-3
10
-2
10
-1
10
0
10
1
10
a
(
L
/
L
o
)
10
-5
10
-3
10
-1
10
1
Fig. 2. Comparison between breakage kernels for c = 0.035 m
2
s
3
(N = 125 rpm; Re}Re
c
= 96) and for c = 0.070 m
2
s
3
(N = 165 rpm;
Re}Re
c
= 127); open squares: exponential kernel with c
1
= 6.0 10
4
,
open circles: power-law breakage kernel with d
[
= 2.6.
a
x
i
a
l
c
o
o
r
d
i
n
a
t
e
,
m
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
radial coordinate, m
Fig. 3. Velocity vectors in a meridian section of the TaylorCouette
reactor at several rotational speeds of the inner cylinder. From left to
right: 75, 125, 165 and 211 rpm (Re}Re
c
= 58, 96, 127, 162).
4. Results and discussion
In Fig. 3, the mean velocity vectors in a meridian section
of the TaylorCouette reactor are shown for 75, 125, 165,
and 211 rpm. For the four rotational speeds reported above
we nd that Re}Re
c
falls in the range between 50 and 160,
which corresponds to the Turbulent Vortex Flow regime
(see Marchisio, 2002).
In all cases, the contour plots show the expected counter-
rotating vortical structure. Although no experimental
D. L. Marchisio et al. / Chemical Engineering Science 58 (2003) 33373351 3345
t, s
0 2500 5000
t, s
0 2500 5000
t, s
d
0 2500 5000
0
5
10
15
20
N = 211 rpm
t, s
0 2500 5000
0
5
10
15
20
25
t, s
0 2500 5000
0
5
10
15
20
25
t, s
0 2500 5000
0
5
10
15
20
25
t, s
0 2000 4000 6000 8000
0
5
10
15
20
25
30
t, s
0 2000 4000 6000 8000
0
5
10
15
20
25
30
t, s
0 2000 4000 6000 8000
0
5
10
15
20
25
30
t, s
0 5000 10000 15000
0
10
20
30
40
t, s
d
0 5000 10000 15000
0
10
20
30
40
t, s
0 5000 10000 15000
0
10
20
30
40
t, s
0 5000 10000 15000
0
10
20
30
40
N = 75 rpm
t, s
d
0 2000 4000 6000 8000
0
5
10
15
20
25
30
N = 125 rpm
t, s
d
0 2500 5000
0
5
10
15
20
25
N = 165 rpm
t, s
0 2500 5000
Fig. 4. Time-evolution of the normalized mean particle size for case 1 (circle: experimental data; dotdashed line: c
1
= 0.3 10
3
; dashed line:
c
1
= 0.6 10
3
; continuous line: c
1
= 1.0 10
3
).
velocity data are available for these specic cases, the re-
sults are consistent with our previous ndings (Marchisio,
2002). Because the grid used in the CFD simulation only
represents half of the axial length of the actual reactor,
the predicted number of vortex pairs in the reactor must
be multiplied by a factor of two. At the lowest rotational
speed shown in Fig. 3, the CFD simulation predicts eight
vortex pairs, whereas at higher rotation rates, seven vortex
pairs are predicted. It is well-known that the axial wave
number in the turbulent Taylor vortex regime depends upon
the azimuthal Reynolds number history, but in general the
number of vortices decreases to a minimum value as Re
is increased beyond the critical value for the onset of tur-
bulent Taylor vortex ow (Lewis & Swinney, 1999). The
change in the number of vortices in mainly caused by the
increase in size of the top and bottom vortices, which is
caused by a well-known end eect, whereas the axial length
of the central vortices seems to be quite constant. The CFD
simulations are consistent with this trend.
Comparison of the aggregationbreakage simulations
with experimental data is made through the normalized
mean particle size d, which is the ratio between the mean
particle size L
43
and the size of primary particles L
o
. The
mean size of the aggregate L
43
is calculated as the ratio be-
tween m
4
and m
3
and corresponds to the volume-averaged
particle size. Notice that the particle size evolution reported
in the comparison is the spatial average over the entire
reactor volume.
4.1. Power-law breakage kernel
Case 1. In this case the aggregationbreakage pro-
cess was modeled by using a power law breakage ker-
nel and assuming that particles break into two equal
fragments.
The comparison between experimental data at four rota-
tional speeds of the inner cylinder with model predictions
for case 1 (see Table 3) for dierent values of the con-
stant c
1
is reported in Fig. 4.The model predictions and
experimental data both show that the steady-state mean
particle size decreases with increasing rotational speed.
However, the simulations predict a faster approach to
steady state, especially at the lower rotational speeds in-
vestigated. The best agreement between simulations and
experiments is found when c
1
=0.610
3
. The constant c
1
is a dimensionless quantity that accounts for physical prop-
erties of the particles and attractive forces that maintain the
aggregate.
Case 2. In this case the aggregationbreakage process was
modeled by using a power-law breakage kernel and assum-
ing that particles undergo erosive breakage. The comparison
with experimental data is shown in Fig. 5. The combina-
tion of the strong hydrodynamic aggregation kernel with the
relatively weak erosion fragment distribution function leads
to simulation predictions of a rapid accumulation of most
of the system mass into a few very large particles. This is
not surprising, since erosion of large aggregates results in
two particles, one of which is a monomer and the other of
which is still a large aggregate. Changes in the value of c
1
are unable to reconcile the simulation predictions and ex-
perimental data, which suggests that erosion cannot be the
only breakage mechanism at play in the experiments. This
does not imply that erosion is not playing any role in the
fragmentation process. It may be possible to explain this
experimental behaviour in terms of a mixture of breakage
mechanisms.
3346 D. L. Marchisio et al. / Chemical Engineering Science 58 (2003) 33373351
1 10 100 1000 10000
t, s
1
10
100
d
Fig. 5. Time-evolution of the normalized mean particle size for case
2 (experimental data: circles N = 75 rpm; squares N = 125 rpm;
diamonds N = 165 rpm; triangles N = 211 rpm; CFD predictions
with c
1
= 0.3 10
3
: dotdotdashed line N = 75 rpm; dotdashed
line N = 125 rpm; dasheddashed line N = 165 rpm; continuous line
N = 211 rpm).
Case 3. In this case the breakage process was modeled
by using a power-law breakage kernel and assuming that
aggregates break into fragments with a uniform probability
distribution of sizes. In this case agreement between experi-
ment and simulation is quite good for a value of the constant
c
1
in the range between 0.610
3
and 1.010
3
(see Fig.
6). The results are also qualitatively and quantitatively com-
parable with case 1 (symmetric fragmentation), but in this
case it seems that the limiting particle size is reached after
a longer period of time, thereby improving the agreement
with the experimental data.
t, s
d
0 5000 10000 15000
0
10
20
30
40
50
N = 75 rpm
t, s
d
0 5000 10000 15000
0
10
20
30
40
50
t, s
d
0 2500 5000
0
5
10
15
20
25
30
N = 211 rpm
t, s
d
0 2000 4000 6000 8000
0
10
20
30
40
N = 125 rpm
t, s
d
0 2000 4000 6000 8000
0
10
20
30
40
t, s
d
0 2500 5000
0
5
10
15
20
25
30
N = 165 rpm
t, s
d
0 2500 5000
0
5
10
15
20
25
30
t, s
d
0 2500 5000
0
5
10
15
20
25
30
t, s
d
0 2500 5000
0
5
10
15
20
25
30
t, s
0 2500 5000
t, s
d
0 2500 5000
0
5
10
15
20
25
30
t, s
0 2500 5000
t, s
d
0 2000 4000 6000 8000
0
10
20
30
40
t, s
0 2000 4000 6000 8000
t, s
d
0 5000 10000 15000
0
10
20
30
40
50
t, s
0 5000 10000 15000
Fig. 6. Time-evolution of the normalized mean particle size for case 3 (circle: experimental data; dotdashed line: c
1
= 0.3 10
3
; dashed line:
c
1
= 0.6 10
3
; continuous line: c
1
= 1.0 10
3
).
4.2. Exponential breakage kernel
Case 4. This case corresponds to the exponential breakage
kernel coupled with a fragment distribution function that re-
quires the formation of two equal-sized fragments. In Fig. 7
the comparison between experimental data and exponential
breakage model predictions for the four rotational speeds
of the inner cylinder are reported. As was already men-
tioned, model predictions were calculated using dierent
values of the group of constants A}H
2
o
, due to the uncertainty
in determining the Hamaker constant A and the interparticle
distance H
o
. Neither the approach to steady state nor the
steady-state mean particle size are well predicted by the
exponential breakage kernel. This could be due to the fact
that in the beginning of the simulation the aggregates are
not fractal (i.e., d
[
is practically equal to three) and thus the
exponential breakage kernel does not give good predictions.
Case 5. In this case the exponential breakage rate function
was used with the erosive fragment distribution function.
Fig. 8 shows that A}H
2
o
= 0.8 10
4
J}m
2
does not give
good agreement with experimental data. In fact, by using
this value a gelling transition is detected. Gelation occurs
when mass is no longer being conserved, and in a nite time,
innite-sized particles are detected. In terms of the moments
this is signaled by negative values of the zeroth moment.
By decreasing A}H
2
o
to 0.4 10
4
J}m
2
, the gelling
transition is only delayed, and A}H
2
o
has to be decreased to
0.210
4
J}m
2
in order not to have gelation within the rst
5000 s of the simulation. The existence of a steady-state
solution for this case is very dicult to be proved, in fact, it
might be possible that the transition is again only delayed.
However, even if the solution for A}H
2
o
= 0.2 10
4
J}m
2
was the actual steady-state solution, the predicted limiting
D. L. Marchisio et al. / Chemical Engineering Science 58 (2003) 33373351 3347
t, s
d
0 2500 5000
0
5
10
15
t, s
d
0 2500 5000
0
5
10
15
t, s
d
0 2500 5000
0
5
10
15
t, s
d
0 2000 4000 6000 8000
0
5
10
15
20
t, s
d
0 2000 4000 6000 8000
0
5
10
15
20
t, s
d
0 2000 4000 6000 8000
0
5
10
15
20
t, s
0 5000 10000 15000
0
5
10
15
20
25
t, s
0 5000 10000 15000
0
5
10
15
20
25
t, s
0 5000 10000 15000
0
5
10
15
20
25
N = 75 rpm
t, s
d
0 2000 4000 6000 8000
0
5
10
15
20
N = 125 rpm
t, s
d
0 2500 5000
0
5
10
15
N = 165 rpm
t, s
d
0 2500 5000
0
2
4
6
8
10
N = 211 rpm
t, s
d
0 2500 5000
0
2
4
6
8
10
t, s
d
0 2500 5000
0
2
4
6
8
10
t, s
d
0 2500 5000
0
2
4
6
8
10
t, s
d
0 5000 10000 15000
0
5
10
15
20
25
Fig. 7. Time-evolution of the normalized mean particle size for case 4 and for d
[
= 2.6 (circle: experimental data; continuous line: A}H
2
o
= 0.7 10
4
;
dashed line A}H
2
o
= 0.8 10
4
; dotdashed line A}H
2
o
= 0.9 10
4
).
Fig. 8. Time-evolution of the normalized mean particle size for case 5 and
for d
[
=2.6 (circle: experimental data; continuous line: A}H
2
o
=0.810
4
;
dashed line: A}H
2
o
= 0.4 10
4
; dotdashed line: A}H
2
o
= 0.2 10
4
).
particle size established is much smaller than what is exper-
imentally observed.
Case 6. The last case considered pairs the exponential
breakage rate with the uniform fragment distribution func-
tion. Results in this case are very similar to case 4, but the
predictions are more sensitive to the value of the parameter
A}H
2
o
(see Fig. 9).
4.2.1. Comparison with the homogeneous model
It is interesting to compare the predictions of the CFD
model with the predictions of a spatially homogeneous
model. Fig. 10 shows the CFD predictions of the distribu-
tion of turbulent dissipation rate over the reactor volume for
two rotational speeds. Note that there is little spread in the
distribution, which is one of the most important and attrac-
tive aspects of turbulent TaylorCouette ow. Nevertheless,
if simulations are performed neglecting spatial inhomo-
geneities and using volume-averaged values of turbulent
properties, signicant dierences arise in the predictions of
the homogeneous and spatially heterogenous models. For
example if spatial inhomogeneities are neglected, then the
mean particle size can be overestimated by 3050%, de-
pending on the choice of kernels. If the reactor is modeled
by using a small number of perfectly mixed reactors with
dierent turbulent properties, according to the hystogram
reported in Fig. 10, the agreement may improve without
resorting to the direct solution of the population balance
equation in the CFD code. However, if for example the
solid particles are produced by a fast chemical reaction, or
if the solid concentration is higher and the ow eld has to
be described by using a multiphase model, this full-CFD
approach seems to be necessary (Baldyga et al., 2002;
Marchisio et al., 2002).
5. Conclusions
Simultaneous aggregation and breakage of particles in a
TaylorCouette reactor was simulated by implementing the
QMOM in a commercial CFD code (FLUENT). Experi-
mental data taken from the literature was compared with the
CFD predictions for a range of operating conditions and for
several combinations of aggregation and breakage kernels.
The implementation of the QMOM in FLUENT was
found to be very convenient. In fact, by tracking only six
scalars and using the quadrature approximation the mo-
ments of the PSD were tracked with very small errors. It is
important to emphasize that achieving the same accuracy
with a DPB method requires use of at least 2040 classes
3348 D. L. Marchisio et al. / Chemical Engineering Science 58 (2003) 33373351
t, s
0 5000 10000 15000
0
5
10
15
20
25
30
t, s
0 5000 10000 15000
0
5
10
15
20
25
30
t, s
0 5000 10000 15000
0
5
10
15
20
25
30
t, s
d
0 5000 10000 15000
0
5
10
15
20
25
30
N = 75 rpm
t, s
d
0 2000 4000 6000 8000
0
5
10
15
20
N = 125 rpm
t, s
d
0 2000 4000 6000 8000
0
5
10
15
20
t, s
d
0 2000 4000 6000 8000
0
5
10
15
20
t, s
d
0 2000 4000 6000 8000
0
5
10
15
20
t, s
d
0 2500 5000
0
5
10
15
N = 165 rpm
t, s
d
0 2500 5000
0
5
10
15
N = 165 rpm
t, s
d
0 2500 5000
0
5
10
15
N = 165 rpm
t, s
d
0 2500 5000
0
5
10
15
N = 165 rpm
t, s
d
0 2500 5000
0
2
4
6
8
10
N = 211 rpm
t, s
d
0 2500 5000
0
2
4
6
8
10
N = 211 rpm
t, s
d
0 2500 5000
0
2
4
6
8
10
N = 211 rpm
t, s
d
0 2500 5000
0
2
4
6
8
10
N = 211 rpm
Fig. 9. Time-evolution of the normalized mean particle size for case 6 and for d
[
= 2.6 (circle: experimental data; continuous line: A}H
2
o
= 0.7 10
4
;
dashed line: A}H
2
o
= 0.8 10
4
; dotdashed line: A}H
2
o
= 0.9 10
4
).
turbulent dissipation rate, m
2
s
-3
v
o
l
u
m
e
%
0 0.25 0.5 0.75 1
0
20
40
60
80
N = 165 rpm
turbulent dissipation rate, m
2
s
-3
v
o
l
u
m
e
%
0 0.1 0.2 0.3 0.4 0.5
0
20
40
60
80
N = 125 rpm
Fig. 10. Histogram of the distribution of the turbulent dissipation rate
over the reactor volume at 125 and 165 rpm (Re}Re
c
= 96 and 127).
(Marchisio et al., 2003b). For this kind of calculation the
controlling step is the solution of the convection and turbu-
lent diusion terms and thus the global CPU time is mainly
inuenced by the number of scalars to be tracked. Although
a general rule for nite-volume codes does not exist, it is
reasonable to expect that the increase in the CPU time with
the number of scalars is stronger than a linear function. For
this reason we can infer that an increase in the number of
scalars from six to 2040 would cause a drastic increase
in the CPU time, thereby rendering the CFD approach im-
practical, especially for complex geometries and if coupled
with multiphase models. A detailed comparison between
the QMOM and a DPB approach with a multiphase model
for bubble columns is under investigation, and rst results
appear to conrm this statement. Moreover, previous results
showed that by using the QMOM the specic CPU time per
scalar solved is constant for dierent kernels (Marchisio
et al., 2003a). Further conrmation is provided by the fact
that with all the tested combinations of kernels used in this
work, the global CPU time was approximately constant (i.e,
about 34 h of wall time for simulating two hours of real
experiment). This is not true for most DPB approaches,
where the discretization of the internal coordinate can
generate mathematical stiness.
Although the QMOM like all moment methods, does
not provide a detailed description of the PSD, recent work
by Diemer and Olson (2002) shows that knowledge of
the lower-order moments is sucient to infer the shape of
the PSD. Moreover, in many applications knowledge of the
PSD is not required. It is important to point out that since
the information needed during the simulation is directly
obtained from the moments, the PSD reconstruction repre-
sents a post-processing step at the end of the simulation,
and thus does not increment the global CPU time.
Concerning the choice of breakage kernels, the exponen-
tial kernel is attractive since its formulation does not require
the invocation of unknown constants. The exponential ker-
nel can in fact be expressed in terms of physical properties of
the aggregate and of the suspending solution. The power law
kernel has no physical basis and proportionality constants
and exponents can be found only through semi-theoretical
D. L. Marchisio et al. / Chemical Engineering Science 58 (2003) 33373351 3349
analysis and comparison with experimental data. However,
use of the power-law kernel improved the agreement with
the transient behavior. As already mentioned, this can be
caused by the fact that for short times the aggregates are
very compact and small and therefore the fractal description
underpredicts the breakage rates. Concerning the daughter
distribution function, it appears that the experimental behav-
ior might be explained in terms of a mixture of breakage
mechanisms, for example erosion for small aggregates and
symmetric breakage for large ones.
It is very dicult to assess the suitability of dierent ker-
nels since the comparison with experiments is made in terms
of the volume-averaged mean particle size. For this reason,
the experimental investigation should be carried out using a
continuous reactor and an in-situ image analysis system to
gather local information for comparison with CFD predic-
tions. Such an investigation is under way (Marchisio et al.,
2002). Moreover the approach has been also applied in a di-
rect formulation (direct quadrature method of moments) for
the investigation of multiphase systems (Marchisio & Fox,
2003). Results from these studies will be reported in greater
detail in future communications.
Notation
a(L) breakage kernel
A Hamaker constant
b(L|z) daughter distribution function

b
(k)
i
kth moment of the daughter distribution function
for L = L
i
B(L; x, t) birth term due to aggregation and breakage
B
a
k
(x, t) kth moment transform of the birth term due to
aggregation
B
b
k
(x, t) kth moment transform of the birth term due to
breakage
d dimensionless mean particle size
d
1
diameter of the inner cylinder of the Taylor
Couette reactor
d
2
diameter of the outer cylinder of the Taylor
Couette reactor
d
[
fractal dimension
D annular gap between the inner and outer cylinder
of the TaylorCouette reactor
D(L; x, t) death term due to aggregation and breakage
D
a
k
(x, t) kth moment transform of the death term due to
aggregation
D
b
k
(x, t) kth moment transform of the death term due to
breakage
F inter-particle force of the fractal aggregate
G molecular growth rate
H length of the TaylorCouette reactor
H
o
primary particle distance in the aggregate
J(x, t) nucleation rate
k
B
Boltzmann constant
k
c
co-ordination number of the fractal aggregate
L particle size
L
i
abscissa (or node) of the quadrature approxima-
tion
L
o
primary particle size
L
43
mean particle size
m
k
(x, t) kth moment of the PSD
n(L; x, t) particle size distribution function
N rotational speed of the inner cylinder of the
TaylorCouette reactor
r
1
inner cylinder radius of the TaylorCouette re-
actor
Re Reynolds number
Re
c
critical Reynolds number for the transition to the
laminar Taylor Vortex Flow
t time
u
i
Reynolds-averaged velocity in the ith direction
w
i
weight of the quadrature approximation
x position vector
x
i
ith component of the position vector
Greek letters
: exponent for the kinematic viscosity in the
power-law breakage kernel
[ exponent for the turbulent dissipation rate in the
power-law breakage kernel
[(L, z) aggregation kernel
exponent for the particle size in the power-law
breakage kernel
I molecular diusivity
I
A
aspect ratio of the TaylorCouette reactor
I
t
turbulent diusivity
c turbulent dissipation rate
z particle size
j viscosity of the suspending uid
v kinematic viscosity of the suspending uid
j
s
solid particle density
t
[
aggregate strength
[(L) volume fraction of solid within the aggregate
c
1
angular velocity of the inner cylinder of the
TaylorCouette reactor
Acknowledgements
This work has been nancially supported the US
Department of Energy (Project award number DE-FC07-
01ID14087).
Appendix A.
The procedure used to nd weights (w
i
) and abscissas
(L
i
) from the moments is based on the PD algorithm. The
rst step is the construction of a matrix P with components
3350 D. L. Marchisio et al. / Chemical Engineering Science 58 (2003) 33373351
P
i, )
starting from the moments. The components in the rst
column of P are
P
i, 1
= o
i1
, i 1, . . . , 2N + 1, (A.1)
where o
i1
is the Kronecker delta. The components in the
second column of P are
P
i, 2
= (1)
i1
m
i1
, i 1, . . . , 2N + 1. (A.2)
Since the nal weights can be corrected by multiplying by
the true m
0
, the calculations can be done assuming a nor-
malized distribution (i.e., m
0
=1). Then the remaining com-
ponents are found from the PD algorithm:
P
i, )
= P
1, )1
P
i+1, )2
P
i, )2
P
i+1, )1
,
) 3, . . . , 2N + 1 and i 1, . . . , 2N + 2 ). (A.3)
If, for example, N = 2 then P becomes
P =
_

_
1 1 m
1
m
2
m
2
1
m
3
m
1
m
2
2
0 m
1
m
2
m
3
+ m
2
m
1
0
0 m
2
m
3
0 0
0 m
3
0 0 0
0 0 0 0 0
_

_
.
(A.4)
The coecients of the continued fraction (:
i
) are gener-
ated by setting the rst element equal to zero (:
1
= 0), and
computing the others according to the following recursive
relationship:
:
i
=
P
1, i+1
P
1, i
P
1, i1
, i 2, . . . , 2N. (A.5)
A symmetric tridiagonal matrix is obtained from sums and
products of :
i
:
a
i
= :
2i
+ :
2i1
, i 1, . . . , 2N 1 (A.6)
and
b
i
=

:
2i+1
:
2i1
, i 1, . . . , 2N 2, (A.7)
where a
i
and b
i
are, respectively, the diagonal and the
co-diagonal of the Jacobi matrix. Once the tridiagonal ma-
trix is determined, generation of the weights and abscissas
is done by nding its eigenvalues and eigenvectors. In fact,
the eigenvalues are the abscissas and the weights can be
found as follows:
w
)
= m
0
t
2
)1
, (A.8)
where t
)1
is the rst component of the )th eigenvector v
)
.
References
Adachi, Y., Cohen Stuart, M. A., & Fokkink, R. (1994). Kinetics
of turbulent coagulation studied by means of end-over-end rotation.
Journal of Colloid and Interface Science, 165, 310317.
Ayazi Shamlou, P., Stravrinides, S., Titchener-Hooker, N., & Hoare, M.
(1994). Growth-independent breakage frequency of protein precipitates
in turbulently agitated bioreactors. Chemical Engineering Science, 49,
26472656.
Baldyga, J., Jasinska, M., & Orciuch, W. (2002). Barium sulphate
agglomeration in a pipean experimental study and CFD modelling.
Proceedings of the 15th international symposium on industrial
crystallization, Sorrento, Italy, 1518 September 2002. Chemical
Engineering Transactions, 1, 6570.
Baldyga, J., & Orciuch, W. (2001). Barium sulphate precipitation
in a pipean experimental study and CFD modelling. Chemical
Engineering Science, 56, 24352444.
Barret, J. C., & Webb, N. A. (1998). A comparison of some approximate
methods for solving the aerosol general dynamic equation. Journal
Aerosol Science, 29, 3139.
Brakalov, L. B. (1987). Connection between the orthokinetic coagulation
capture eciency of aggregates and their maximum size. Chemical
Engineering Science, 42, 23732383.
Brown, D. L., & Glatz, C. E. (1987). Aggregate breakage in protein
precipitation. Chemical Engineering Science, 42, 18311839.
Coulaloglou, C. A., & Tavlarides, L. L. (1977). Description of interaction
processes in agitated liquidliquid dispersions. Chemical Engineering
Science, 32, 12891297.
Dette, H., & Studden, W. J. (1997). The theory of canonical moments
with applications in statistics, probability, and analysis. New York:
Wiley.
Diemer, R. B., & Olson, J. H. (2002). A Moment methodology
for coagulation and breakage problems: Part IIMoment models
and distribution reconstruction. Chemical Engineering Science, 57,
22112228.
Filippov, A. V., Zurita, M., & Rosner, D. E. (2000). Fractal-like
aggregates: Relation between morphology and physical properties.
Journal of Colloid and Interface Science, 229, 261273.
Fitch, R. M. (1997). Polymer colloids: A comprehensive introduction.
San Diego, CA: Academic Press.
Fluent Inc. (2002). FLUENT 6.0 UDF manual. Lebanon, NH: Fluent
Inc.
Frisch, U. (1995). Turbulence. Cambridge: Cambridge University Press.
Gordon, R. G. (1968). Error bounds in equilibrium statistical mechanics.
Journal Mathematical Physics, 9, 655672.
Gruy, F. (2001). Formation of small silica aggregates by turbulent
aggregation. Journal of Colloid and Interface Science, 237, 2839.
Hansen, P. H. F., Malmsten, M., Bergenstahl, B., & Bergstrom, L.
(1999). Orthokinetic aggregation in two dimensions of monodisperse
and bidisperse colloidal systems. Journal of Colloid and Interface
Science, 220, 269280.
Hounslow, M. J., Mumtaz, H. S., Collier, A. P., Barrick, J. P., & Bramley,
A. S. (2001). A micro-mechanical model for the rate of aggregation
during precipitation from solution. Chemical Engineering Science, 57,
25432552.
Hulburt, H. M., & Katz, S. (1964). Some problems in particle technology.
Chemical Engineering Science, 19, 555574.
Jlang, Q., & Logan, B. E. (1991). Fractal dimension of aggregates
determined from steady-state size distributions. Environmental Science
and Technology, 25, 20312038.
Kataoka, K. (1986). Taylor vortices and instabilities in circular Couette
ows. In N. P. Cheremisino (Ed.), Encyclopedia of uid mechanics
(Vol. 1, pp. 237273). Houston: Gulf Publishing.
Kramer, T. A., & Clark, M. M. (1999). Incorporation of aggregate breakup
in the simulation of orthokinetic coagulation. Journal of Colloid and
Interface Science, 216, 116126.
D. L. Marchisio et al. / Chemical Engineering Science 58 (2003) 33373351 3351
Krutzer, L. L. M., van Diemen, A. J. G., & Stein, H. N. (1995). The
inuence of the type of ow on the orthokinetic coagulation rate.
Journal of Colloid and Interface Science, 171, 429438.
Lee, K., & Matsoukas, T. (2000). Simultaneous coagulation and break-up
using constant-N Monte Carlo. Powder Technology, 110, 8289.
Lewis, G. S., & Swinney, H. L. (1999). Velocity structure functions,
scaling, and transitions in high-Reynolds-number CouetteTaylor ow.
Physical Review A, 59, 54575467.
Luo, H., & Svendsen, H. F. (1996). Theoretical model for drop and bubble
breakup in turbulent dispersion. A.I.Ch.E. Journal, 42, 12251233.
Marchisio, D. L. (2002). Precipitation in turbulent uids. Ph.D.
dissertation, Politecnico di Torino, Italy.
Marchisio, D. L., Barresi, A. A., & Garbero, M. (2002). Nucleation,
growth and agglomeration in barium sulphate turbulent precipitation.
A.I.Ch.E. Journal, 48, 20392050.
Marchisio, D. L., & Fox, R. O. (2003). Direct quadrature method
of moments: Derivation, analysis and applications. Journal of
Computational Physics, submitted.
Marchisio, D. L., Pikturna, J. T., Fox, R. O., Vigil, R. D., & Barresi, A.
A. (2003a). Quadrature method of moments for population balances
with nucleation, growth and aggregation. A.I.Ch.E. Journal, 49, 1266
1276.
Marchisio, D. L., Pikturna, J. T., Wang, L., Vigil, R. D., & Fox, R. O.,
(2002). Quadrature method of moments for population balance in CFD
applications: Comparison with experimental data. Proceedings of 14th
International Symposium on Industrial Crystallization, September
1518, 2002, Sorrento, Italy, pp. 305310.
Marchisio, D. L., Vigil, R. D., & Fox, R. O. (2003b). Quadrature method
of moments for aggregationbreakage processes. Journal of Colloid
and Interface Science, 258, 322334.
McGraw, R. (1997). Description of aerosol dynamics by the quadrature
method of moments. Aerosol Science and Technology, 27, 255265.
Moreira, J. L., Cruz, P. E., Santana, P. C., Aunins, J. G., & Carrondo,
M. J. T. (1995). Formation and disruption of animal cell aggregates in
stirred vessels: Mechanisms and kinetic studies. Chemical Engineering
Science, 50, 27472764.
Narsimhan, G., Gupta, J. P., & Ramkrishna, D. (1979). A model for
transitional breakage probability of droplets in agitated lean liquid
liquid dispersion. Chemical Engineering Science, 34, 257265.
Oles, V. (1991). Shear-induced aggregation and breakup of polystyrene
latex particles. Journal of Colloid and Interface Science, 154,
351362.
Pandya, J. D., & Spielman, L. A. (1982). Particle breakage: Theory and
processing strategy. Journal of Colloid and Interface Science, 90,
517531.
Peng, S. J., & Williams, R. A. (1994). Direct measurement of oc
breakage in owing suspension. Journal of Colloid and Interface
Science, 166, 321332.
Press, W. H., Teukolsky, S. A., Vetterling, W. T., & Flannery, B. P.
(1992). Numerical recipes in C (2nd ed.). New York: Cambridge
University Press.
Ramkrishna, D. (1985). The status of population balances. Review
Chemical Engineering, 3, 4995.
Ramkrishna, D. (2000). Population balances: Theory and applications
to particulate systems in engineering. New York: Academic Press.
Ramkrishna, D., & Mahoney, A. W. (2002). Population balance modeling.
Promise for the future. Chemical Engineering Science, 57, 595606.
Randolph, A. D., & Larson, M. A. (1988). Theory of particulate processes
(2nd ed.). San Diego, CA: Academic Press.
Raphael, M., & Rohani, S. (1996). Isoelectric precipitation of sunower
protein in a MSMPR precipitator: Modeling of PSD with aggregation.
Chemical Engineering Science, 51, 43794384.
Rosner, D. E., & Pyykonen, J. J. (2002). Bivariate moment simulation of
coagulation and sintering nanoparticles in ames. A.I.Ch.E. Journal,
48, 476491.
Rosner, D. E., & Yu, S. (2001). MC simulation of aerosol aggregation
and simultaneous spheroidization. A.I.Ch.E. Journal, 47, 545561.
Saman, P. G., & Turner, J. S. (1956). On the collision of drops in
turbulent clouds. Journal of Fluid Mechanics, 1, 1631.
Serra, T., & Casamitjana, X. (1998a). Structure of the aggregates during
the process of aggregation and breakup under shear ow. Journal of
Colloid and Interface Science, 206, 505511.
Serra, T., & Casamitjana, X. (1998b). Eect of the shear and volume
fraction on the aggregation and breakup of particles. A.I.Ch.E. Journal,
44, 17241730.
Serra, T., Colomer, J., & Casamitjana, X. (1997). Aggregation and
breakup of particles in shear ows. Journal of Colloid and Interface
Science, 187, 466473.
Smith, M., & Matsoukas, T. (1998). Constant-number Monte Carlo
simulation of population balances. Chemical Engineering Science, 53,
17771786.
Smoluchowski, M. Z. (1917). Versuch Einer Mathematischen Theorie Der
Koagulationskinetik Kolloider Losunger. Zeitschrift fur Physikalische
Chemie, 92, 129142.
Sonntag, R., & Russel, W. B. (1986). Structure and breakup of ocs
subjected to uid stresses. I. Shear experiments. Journal of Colloid
and Interface Science, 113, 339349.
Spicer, P. T., & Pratsinis, S. E. (1996). Coagulation and fragmentation:
Universal steady-state particle-size distribution. A.I.Ch.E. Journal, 42,
16121620.
Tontrup, C., Gruy, F., & Cournil, M. (2000). Turbulent aggregation
of titania in water. Journal of Colloid and Interface Science, 229,
511525.
Vanni, M. (2000a). Approximate population balance equations for
aggregationbreakage processes. Journal of Colloid and Interface
Science, 221, 143160.
Vanni, M. (2000b). Creeping ow over spherical permeable aggregates.
Chemical Engineering Science, 55, 685698.
Wojcik, J. A., & Jones, A. G. (1998). Particle disruption of precipitated
CaCO
3
crystal agglomerates in turbulently agitated suspensions.
Chemical Engineering Science, 53, 10971101.
Wright, D. L., McGraw, R., & Rosner, D. E. (2002). Bivariate extension
of the quadrature method of moments for modeling simultaneous
coagulation and sintering of particle populations. Journal of Colloid
and Interface Science, 236, 242251.
Zauner, R., & Jones, A. G. (2000). Determination of nucleation, growth,
agglomeration and disruption kinetics from experimental precipitation
data: The calcium oxalate system. Chemical Engineering Science, 55,
42194232.

Potrebbero piacerti anche