Computers and Chemical Engineering 31 (2007) 206–218

Considerations on the crystallization modeling: Population balance solution

Caliane Bastos Borba Costa ∗ , Maria Regina Wolf Maciel, Rubens Maciel Filho
Laboratory of Optimization, Design and Advanced Control (LOPCA), School of Chemical Engineering, State University of Campinas (UNICAMP),
Cidade Universitária Zeferino Vaz, Caixa Postal 6066, CEP 13081-970 Campinas, SP, Brazil
Received 29 April 2005; received in revised form 15 May 2006; accepted 29 June 2006
Available online 17 August 2006

Abstract
In crystallization processes, the need to improve the product quality and to minimize production cost requires understanding and optimization
on how to obtain an appropriate particle-size distribution. In order to achieve this goal, a model of the process is necessary and the distribution
of particles must be represented, which is made through the population balance. This latter constitutes a strongly nonlinear hyperbolic partial
differential equation and, in most cases, an analytical solution is not possible, requiring the development and adaptation of numerical techniques.
In the present work, it is proposed to analyze the positive and negative aspects of some methods chosen as a tool in the treatment of the population
balance equation in crystallization processes. Numerical problems arising from transformation of the partial differential equation, computational
effort, as well as the incorrect prediction of the total number of particles and/or no mass conservation of the dispersed phase are discussed.
© 2006 Elsevier Ltd. All rights reserved.

Keywords: Crystallization; Population balance; Numerical methods; Computer simulation

1. Introduction but many of them have their applicability restricted to some

special cases/processes, while others are of more general appli-
Population balance is a well established approach as the math- cation.
ematical framework for dealing with particulate systems. These Some applications of the model numerical solution demand
kinds of processes involve formation of entities, growth, break- fast and efficient codes, as in model-based control algorithms,
age or aggregation of particles, as well as dispersion of one in which the model must be solved in real time. The control
phase in another one, and are, therefore, present in a large range objectives for many processes are different and the selection of
of applications, like polymerization, crystallization, bubble tow- a numerical method must look to it. In continuous crystallization
ers, aerosol reactors, biological processes, fermentation or cell problems, for example, the concern is in stabilization of a fre-
culture. Such mathematical approach in any of these systems quently oscillating operation, while, in batch mode, stability is
follows the number of entities, such as solid particles, bubbles not the issue. Here, the pressing question is whether the crystal
or cells in such way that their presence or occurrence may dictate size distribution (CSD) can be tailored to user specifications by
the behavior of the system under consideration (Puel, Févotte, applying an optimal manipulated variable trajectory to the sys-
& Klein, 2003a; Ramkrishna & Mahoney, 2002). The theoret- tem and by periodically on-line computing to correct noise and
ical treatment covered by the population balance, nevertheless, process disturbances (Rawlings, Witkowski, & Eaton, 1992).
encounters a practical barrier in most applications, since it repre- Bearing the exposed in mind, the present paper presents the
sents a hyperbolical partial differential equation with analytical original population balance framework, the hyperbolic partial
solutions in just few cases. In others, numerical methods must differential equation and reviews the main numerical methods
be applied to calculate the system response, or even to optimize already proposed to solve it. The methods are, then, analyzed
existing processes or to design new ones. A large number of and the advantages and drawbacks for application in crystalliza-
numerical methods have been proposed in the open literature, tion processes are discussed, focusing on possible numerical
problems, on the computational effort and on the correctness
of prediction of the total number of particles and of mass con-
∗ Corresponding author. Tel.: +55 19 3788 3971; fax: +55 19 3788 3965. servation. The idea is not to present a formal organization of
E-mail address: caliane@lopca.feq.unicamp.br (C.B.B. Costa). numerical methods based on any criteria, but, as mentioned,

0098-1354/$ – see front matter © 2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compchemeng.2006.06.005
 C.B.B. Costa et al. / Computers and Chemical Engineering 31 (2007) 206–218 207

Nomenclature 
βj,k agglomeration kernel for sizes xj , xk (=β (xj , xk ))
b (ε) breakage rate of a particle of size ε δ Dirac Delta function
bk defined as b(xk ) γ (ε) number of daughter particles in the breakage of a
B birth rate using length as internal coordinate particle of size ε
B birth rate using volume as internal coordinate μj moment j of a distribution
B̄ mean birth rate using length as internal coordinate μij cross-moment ij of a distribution
B0 nucleation rate ψi basis functions used in the weighted residuals
Ci ith granulometric class methods
D death rate using length as internal coordinate
D death rate using volume as internal coordinate Subscripts
D̄ mean death rate using length as internal coordi- agg aggregation
nate br breakage
f mesh function nucl nucleation
G growth rate
h net rate of production of particles
kv volume shape factor to stress the positive and negative features of methods in the
L characteristic length of the particles solution of crystallization problems. Cooling crystallization is
La lower limit of a subdomain a process in which the mechanisms of nucleation, growth and
Lb upper limit of a subdomain aggregation occur in a competitive fashion and, so, a numer-
Lmax maximum length of the particles in a truncated ical method suitable for solving PBE in crystallization prob-
domain lems should deal with the three mechanisms with acceptable
m number of properties of particles used in the PBE accuracy.
n number density function using length as internal The paper is organized as follows. The population balance
coordinate approach is first introduced, with discussions in its area of
n number density function using volume as internal application and analytical solution limitations. Then the for-
coordinate mal complete population balance equation and some definitions
N number of particles per suspension volume unit are presented, and some important issues are raised. Section 3
NE number of subdomains discusses in a general manner the three main groups in which
pi properties of particles used in the PBE the specific methods for solving the population balance equa-
p (v/ε) fraction of daughter particles with size between v tion may be classified, while Section 4 presents specific meth-
and v + dv generated from breakage of particles ods, with their applications in literature, advantages, drawbacks
of size ε and applicability in crystallization systems. The next section
Qk volumetric rate of kth stream presents other methods, which cannot be classified in none of
R residual in the weighted residuals methods the three main groups of Section 3. As an exemplification of a
RA net rate of agglomeration in Ci granulometric method application in a crystallization process, Section 6 gives
class (=Bagg − Dagg ) some numerical results and points out the notable drawbacks
RB net rate of breakage in Ci granulometric class and advantages. Finally, the last section provides the conclud-
(=Bbr − Dbr ) ing remarks.
S state vector
t time
2. The general population balance equation
ub , vb , wb components of the particle velocity vector vb
vb particle velocity vector
The population balance framework considers space as the
V suspension volume
physical one, as well as any important system property, which
wi weighting functions
changes due to the intrinsic mechanisms and external influences
x vector of internal coordinates
to the system. Concerning to the physical space, the state vector
xi representative volume of the ith size range, also
coordinates refer to spatial coordinates, like (x, y, z), and are
the ith grid point
denoted as external coordinates. The coordinates corresponding
x, y, z spatial coordinates
to the properties (internal coordinates) are the ones of relevance
Greek letters to the process and may be composed, for example, by any impor-
Ci width of ith class tant dimension of the particle or its age in the system (Hounslow,
β (L, λ) agglomeration kernel using length as internal 1998; Randolph & Larson, 1971). The state vector S can be, then,
coordinate expressed by S = [x, y, z, t, L, (pi )1≤i≤m ], where L is the char-
β (v, ε) agglomeration kernel using volume as internal acteristic dimension of the particles, besides m other properties
coordinate and t corresponds to time. The number density function n[x, t,
L, (pi )1≤i≤m ] can then be used to make a balance in the particles
208 C.B.B. Costa et al. / Computers and Chemical Engineering 31 (2007) 206–218

population, composing the population balance equation, PBE using volume as internal coordinate are given, respectively, by
(Pilon & Viskanta, 2003): Eqs. (2) and (3):
∂n ∂ ∂ ∂ ∂  v
+ (ub n) + (vb n) + (wb n) + (L̇n)  1
∂t ∂x ∂y ∂z ∂L Bagg (v) = β (v − ε, ε)n (v − ε, t)n (ε, t) dε (2)
2 0
m
 ∂
+ (ṗi n) = h (1)  ∞
∂pi 
Dagg (v) 
= n (v, t) β (v, ε)n (ε, t) dε (3)
i=1
0
where ub , vb and wb are the components of the particle veloc-
ity vector vb , L̇ is the time rate of change of the characteristic The agglomeration kernel, β (ν, ε), is a measure of the fre-
dimension, while ṗi is the time rate of change of the other prop- quency of collisions between particles of volumes ν and ε which
erties. The net rate of particles production of a particular state produces a particle of volume ν + ε, being a function of the size
at time t is denoted by h, which comprises death and birth domains of mother particles and of the agglomerate and the
mechanisms. hydrodynamic conditions.
The main reason for the unavailability of analytical solu- It is frequently easier to write the growth rate with length as
tions to the PBE is that the birth and death terms correspond- internal coordinate, and, therefore, a modification in the birth
ing to breakage and agglomeration are expressed as integral and death rate terms, based on length, is necessary:
functions. Apart from that, the density function may vary
in multidimensional space and may depend on external vari-  L β[(L3 1/3 1/3
L2 − λ3 ) , λ]n[(L3 − λ3 ) , t]n(λ, t) dλ
ables. These two last complications to the solution of the Bagg (L)= 2/3
PBE are normally neglected in most applications, by assum- 2 0 (L3 − λ3 )
ing perfectly mixed tank. In this way, the external coordi- (4)
nates are normally not considered and the PBE is expressed
only with the internal coordinates. An interesting work pre-  ∞
sented in literature dealing with imperfect mixing is the one Dagg (L) = n(L, t) β(L, λ)n(λ, t) dλ (5)
of Ma, Tafti, and Braatz (2002). The effect of spatial vari- 0
ation is evaluated using a compartmental model, which may
require the use of parallel computations, if the number of consid- David, Paulaime, Espitalier, and Rouleau (2003) divide
ered compartments is large, in order to have shorter simulation agglomeration in three levels, to know, Brownian, laminar and
times. turbulent agglomeration and develop the expressions for the
The hyperbolic differential operator of the PBE is the diver- agglomeration kernel for each one of them.
gence term, composed by all derivatives with respect to the According to Verkoeijen et al. (2002), the birth and death
coordinates, which is divided into two parts. The divergence terms in the PBE express the transfer of particles from one size
concerning to external coordinates expresses the change in class to the other by breakage or by agglomeration and, since
population in a particular volume due to the particles pass- they are not rate processes, cannot be differentiated with respect
ing through that volume with velocity vb . The second diver- to time.
gence term represents the change in population due to particle A different perspective in the modeling of agglomeration was
property, pi . If the property is the particle size, then the diver- proposed by Marchal, David, Klein, and Villermaux (1988), con-
gence term represents the growth (represented by G) or shrink- sidering this mechanism as chemical reactions between species
age of the particles (Verkoeijen, Pouw, Meesters, & Scarlet, in the many particle sizes. The final agglomeration rate expres-
2002). sion per size domain is based on an intrinsic rate, function of the
The kinetic mechanisms present in particulate processes are number of collision per time and volume unit and of supersatu-
growth (commonly associated with the rate of change of the char- ration.
acteristic dimension), nucleation (appearance of new entities), Equations analogous to Eqs. (2) and (3) can be formulated
agglomeration and breakage. In crystallization field, nucleation to express the birth and death terms due to the breakage mech-
and crystal growth are dominant, but in many systems, agglom- anisms, as given by Eqs. (6) and (7):
eration and breakage are present to a certain extent, that must be  ∞ v

considered into the modeling. Bbr (v) = γ(ε)b(ε)p n (ε, t) dε (6)
v ε
2.1. Expressions to birth and death terms

Dbr (v) = −b(v)n (v, t) (7)
The birth (B) and death (D) terms represented in Eq. (1) by h
correspond to breakage and agglomeration mechanisms, which where γ(ε) is the number of daughter particles generated from
are commonly expressed as integral functions. the breakup of a particle of size ε, b(ε) is the breakup rate of a
Aggregation results from the binary collision of mother par- particle with size ε and p(v/ε) is the fraction of daughter parti-
ticles and the sticking by growth of crystalline bridges between cles with size between v and v + dv, generated from breakup of
particles. Bearing aggregation in mind, the birth and death terms, particles of size ε.
 C.B.B. Costa et al. / Computers and Chemical Engineering 31 (2007) 206–218 209

2.2. Internal coordinate and density function When more than one crystal dimension is considered, cross-
considerations moments can be defined. For example, for two dimensions in
a distribution based on the width (L1 ) and the length (L2 ), the
Verkoeijen et al. (2002) present reasons why the population cross-moments can be defined as in Eq. (10):
balance should be expressed with particle volume as the basic
particle-size parameter. They argue, using the agglomeration  ∞ ∞
j
mechanism, that when length is used to express size, consis- μij = Li1 L2 n(L1 , L2 , t) dL (10)
tency with the mass balance cannot be achieved. However, as 0 0
pointed out by Mahoney and Ramkrishna (2002), singularity in
the number density may occur in certain treatments. In precipi- The total number is given by μ00 , while μ10 is the total width
tation systems, the length based growth rate is usually finite at and μ01 is the total length. The average width of crystals is
small particle sizes. This corresponds to a volumetric growth μ10 /μ00 and the average length is μ01 /μ00 .
rate that approaches zero at L = 0, and a consequent singular- Some methods for solving the PBE appeal to the moments
ity in the volume based number density. The singular behavior of the distribution in order to manipulate the original partial
creates very steep gradients in the solution that are difficult to differential equation. Special attention to these methods will be
capture with a discrete representation, besides being difficult to given in next section.
scale the solution, even if v = 0 is not included in the domain.
These problems are avoided by the formulation in length, with
the drawback that evaluation of aggregation kernel becomes con-
siderably more complex. 3. Numerical methods for PBE solution
Another point discussed by Verkoeijen et al. (2002) is the
proposal that the basic population distribution should be the The methods for solving the PBE can be classified into three
mass distribution of particle, instead of the common practice of main groups: the method of moments, the discretization of the
expressing the distribution of the particle size as a number distri- size domain interval and the weighted residuals.
bution. The reason is posed in terms that the number of particles The method of moments is one of the oldest ones and trans-
is not conserved, since the particles may break or agglomerate, forms the PBE into a set of ordinary differential equations
being the total mass of particles the amount that is conserved. As (ODEs) by multiplying the population balance equation by Lj (in
a result, the basic PBE should be expressed in terms of the mass a length based PBE) and integrating it, giving equations in terms
distribution, or the volume distribution if the density is constant, of moments (Randolph & Larson, 1971). The solution of the set
so the mass in the system is automatically conserved. Neverthe- of equations gives the moments of the distribution as function of
less, the authors deal with the description of grinding and aerosol time, considering size independent growth and mean birth and
processes. Conservation of total particles mass is not the case death terms:
for crystallization process, where the extraction of solute from   
∞ ∂n ∂
solution can occur with the growth of existing crystals or with
Lj + (Gn) + D − B = 0 (11)
the nucleation of new entities. So, this argument is not valid for 0 ∂t ∂L
a crystallization application and, historically, the number distri-
bution has been the preferred approach, as exposed by Puel et al.
dμ0
(2003a). = B0 + B̄ − D̄
dt
dμ1
2.3. Moments of the distribution = μ0 G + B̄ − D̄
dt (12)
dμ2
The moment j of the number distribution n is defined as in = 2μ1 G + B̄ − D̄
dt
Eq. (8): dμ3
= 3μ2 G + B̄ − D̄
 ∞ dt
μj = Lj n(L, t) dL (8)
0
In the set of Eq. (12), B0 is the nucleation rate. When only
For the number distribution based on volume as internal coor- nucleation and growth are being considered, the moments can
dinate, the moment j is given by Eq. (9): be generated by the ordinary differential equation given by Eq.
(13):
 ∞
μj = vj n (v, t) dv (9)
0 dμj
= 0j B0 + jGμj−1 (13)
The zeroth moment (j = 0) represents the total number of par- dt
ticles present in the system, while the total volume of particles
is inferred from the third moment (j = 3) defined by Eq. (8) or It is easy to derive the ODEs when PBE is a two-dimensional
from the first moment (j = 1) defined by Eq. (9). one:
210 C.B.B. Costa et al. / Computers and Chemical Engineering 31 (2007) 206–218

dμ00 may even be computationally attractive. Finite element methods

= B0
dt approximate the solution with piecewise low-order polynomials
(14)
dμij that are only locally nonzero, and are, thus, flexible and capable
= iG1 μ(i−1)j + jG2 μi(j−1) i+j>0 of capturing highly irregular solutions. (Rigopoulos & Jones,
dt
2003). According to Rawlings, Miller, and Witkowski (1993),
The great disadvantage of the method is the mathematical the computational time for the model solution with the weighted
complication in the equations when the growth rate is a size residuals method is modest, turning possible to consider it in an
dependent mechanism. In this case, an alternative method of on-line control scheme.
solution is to use the moments and an orthogonal polynomial In global methods, the population density function is approx-
to simulate the population density function. When aggregation imated as a linear combination of chosen basis function, Ψ i (L):
and/or breakage are included, the reduction to moment equations
∞

is impossible. Apart from this drawback, reconstructing the real
distribution from its moments is numerically unstable (Nallet, n(L, t) = ai (t)Ψi (L) (15)
Mangin, & Klein, 1998; Rigopoulos & Jones, 2003). This last i=1

disadvantage can represent a serious problem in certain model

The unknown ai ’s are determined by substituting the previous
applications. Many times, as in certain optimal control evalu-
equation into the population balance to define a residual, R (L,
ations, the entire particle size distribution may be necessary,
t). The sense in which the residual is made small determines the
which prevents the use of the method of moments (Ma et al.,
type of weighted residual method used. The idea of weighted
2002).
residuals is to find the ai ’s that force the residual to be orthogonal
The discretization technique, which is also referred to as
to a chosen set of weighting functions, wi (L):
discretization sizing technique, discretize the spectrum of the
independent variable into a number of intervals and subsequently  ∞
use the mean-value theorem to transform the continuous PBE wi (L)R(L, t) dL = 0 (16)
into a series of equations in terms of either number or aver- 0

age population density in each class. The method turns the PBE Possible choices of the weighting functions include Dirac’s
in the so-called discretized population balance (DPB) and the delta functions, resulting in collocation methods, or the
resulting set of ODEs has so many equations as the number of basis functions themselves (possibly weighted), resulting in
granulometric classes. Galerkin’s techniques (Mahoney & Ramkrishna, 2002).
The computational effort for the numerical solution can be In FE methods, the infinite domain is truncated to a finite
severely reduced if the grid assumes a geometric progression— one, L = [0, Lmax ). This truncated domain is then partitioned into
the number of combinations of particles that must be NE discrete subdomains. For all L ∈ (La , Lb ], a subdomain, the
considered to evaluate the aggregation terms is substan- solution to the PBE will be approximated as a linear combination
tially reduced (Rigopoulos & Jones, 2003). Furthermore, of interpolation functions:
coarse discretization methods are particularly amenable to
process control applications because of the speed with nc

which solutions can be obtained especially when fronts and n(L) ≈ neh (L) = nej Ψje (L) ∀L ∈ (Lea , Leb ] (17)
discontinuities are not present (Ramkrishna & Mahoney, j=1
2002).
Two drawbacks can be cited. Conservation of both number of where nc is the number of nodes in element e, nj ’s are the
particles and mass is only guaranteed in the limit of infinite reso- nodal values of the density distribution and the Ψ j (L)’s are
lution and a discontinuity can arise along the separatrix, which is the basis polynomials of order nc − 1 (Nicmanis & Hounslow,
the curve that divides states deriving from initial conditions from 1996).
those arising from boundary conditions. A sharp discontinuity The truncation results in an underestimation of the integrals of
can be created, which quickly broadens by numerical diffusion the death term due to agglomeration and breakage. In most prac-
in simulation (Mahoney & Ramkrishna, 2002). tical applications, the density distribution asymptotes towards
The weighted residuals comprise methods that retrieve the zero at sufficiently large particle sizes, so a maximum size can
distribution by approximating the solution with a series of trial be selected to be sufficiently large that underestimation is neg-
functions, whose coefficients are to be determined so that their ligibly small. However, care must be taken to avoid selection of
sum will satisfy the PBE. They are divided in weighted residuals unnecessarily large values of the upper limit, since tail regions
with global functions and finite element (FE) methods. Weighted can be difficult and/or computationally expensive to converge,
residual methods with global functions were among the first to be because of the very small values they can attain at large particle
tried in PBE numerical solution. It was soon realized, however, sizes (Nicmanis & Hounslow, 1998).
that global functions cannot capture the features of an arbitrar- In orthogonal collocation on finite elements (OCFE), the
ily shaped distribution, especially if it exhibits sharp changes residual is made small by forcing it to vanish at selected Li loca-
and discontinuities. If a priori knowledge of the shape of the tions, called the collocation points. The choice of these locations
resulting distribution is available, the trial functions can be tai- is vital, since they control the density of the collocation points
lored to accommodate it; in that case, the method converges and in the overall size domain. Thus, in regions with steep changes
 C.B.B. Costa et al. / Computers and Chemical Engineering 31 (2007) 206–218 211

a higher density of collocation points is usually required. The Summarizing, a comparison of the DPB techniques with FE
domain partition into finite elements can be performed either methods reveals that, although the latter are more difficult to
based on previous knowledge about the shape and characteris- implement and computationally demanding, they are very gen-
tics of the distribution or based on the satisfaction of a certain eral and flexible, coping with any possible formulation, like size-
error criterion (Krallis et al., 2004). dependent growth and different aggregation kernels. However,
The choice of the basis functions, both in collocation and it is interesting to note that, when mechanisms of aggregation
Galerkin’s methods, must be suitable to the specific problem, and breakage are present, DPBs are a more natural choice, since
either in global or FE applications. Laguerre polynomials are these mechanisms are essentially discrete. DPBs are, on the other
relevant for continuous flow crystallizers, since the steady-state hand, methods easy to use and computationally efficient, but less
solution is exponential in the range [0, +∞) and the cyclic behav- general, driven to specific problems, and may incur in severe
ior frequently present can be satisfactorily represented (Puel et errors of certain distribution moments.
al., 2003a; Rawlings et al., 1992).
Advantages are convergence properties on the entire distri- 4. Application of numerical methods in literature
bution and the availability of adaptive mesh techniques. Major
drawbacks include the inability to capture discontinuities, as 4.1. Method of moments
may arise along the separatrix, and the computational overhead
arising from double integral evaluations in Galerkin’s formu- As already exposed, this method, although providing a simple
lations. Furthermore, the number of system states required to manipulation of the PBE, has a lot of restrictions and draw-
represent the solution grows exponentially with the number of backs. One of the most recent papers in crystallization area
dimensions (Mahoney & Ramkrishna, 2002). Here, DPBs are that uses this method is the one of Ma and Braatz (2003).
much easier and more robust to be extended in a second (or The main proposal was an approach for robust identification
third, if necessary) dimension (Puel et al., 2003a). and control of batch and semibatch processes and the two-
The major advantages of FE numerical solutions are their dimensional crystallization was used to demonstrate it. The
robustness, easier convergence and open scale of the size authors appealed to the cross-moments in order to reduce the
domain. Apart from that, FE method does not incur any trun- PBE, with only nucleation and size-independent growth, to a set
cation errors at the lower volume range and uses significantly of ODEs.
smaller amounts of memory when solving the PBE as well
as CPU time needed to convergence than some DPB methods 4.2. Discretized population balances
(Nicmanis & Hounslow, 1998).
The large number of schema that must be evaluated for a given Marchal et al. (1988) introduced the Method of Classes, appli-
system, as well as the difficulty to implement varying grid reso- cable to a general case, making possible to consider agglomer-
lutions are possible disadvantages of the FE approach, although ation, breakage and length-dependent growth rate. The method
existing theory makes extension to multiple internal coordi- discretizes the size domain interval in a free of choice grid,
nates reasonably straightforward. Under conditions of excessive generating granulometric classes or bins (Ci ). The mean size in
nucleation, the stiffness characteristics of the OCFE can cause each class is assumed as the characteristic size for all particles
integration problems (Immanuel & Doyle III, 2003). belonging to that class. The model assumes a constant value for
the density function in each bin (zero order method).
⎧
⎪ dN1 1 dV Qout N1 − Qin N1e G(L1 ) G(L1 )
⎪
⎪ dt + V dt N1 + + N2 + N1 = B0 + RA,1 − RB,1
⎪
⎪ V 2C2 2C1
⎪
⎨ dN
i 1 dV Qout Ni − Qin Nie G(Li ) G(Li ) − G(Li−1 ) G(Li−1 )
+ Ni + + Ni+1 + Ni − Ni−1 = RA,i − RB,i (18)
⎪
⎪ dt V dt V 2C i+1 2C i 2Ci−1
⎪
⎪
⎪
⎪ dNNE 1 dV Qout NNE − Qin NNEin −G(LNE−1 ) G(LNE−1 )
⎩ + NNE + + NNE − NNE−1 = RA,NE − RB,NE
dt V dt V 2CNE 2CNE−1
In their original work, Marchal et al. (1988) solved the formed
set of ODEs (Eq. (18)) with a fourth order Runge–Kutta code
and considered the mechanisms of nucleation, length depen-
An approach also used, which differs from the determinis- dent growth and agglomeration. The Method of Classes was
tic approach of the population balance considered so far, is the vastly used in literature to solve the PBE in crystallization and
probabilistic Monte Carlo method. In this technique, the solu- precipitation field, for simulation of experimental results, deter-
tion is determined using the Markov conditional probability, mination of kinetic parameters, modeling of process or optimiza-
which is the probability of a particular state proceeding to any tion of operation conditions (Blandin, Mangin, Nallet, Klein,
other state. Besides predicting the ensemble average properties, & Bossoutrot, 2001; David & Bossoutrot, 1996; David et al.,
it also predicts the fluctuations. This technique solves more com- 2003; David, Villermaux, Marchal, & Klein, 1991; Frank, David,
plex multi-dimensional population balance models (Immanuel Villermaux, & Klein, 1988; Matthews, Miller, & Rawlings,
& Doyle III, 2003). 1996; Matthews & Rawlings, 1998; Miller & Rawlings, 1994;
212 C.B.B. Costa et al. / Computers and Chemical Engineering 31 (2007) 206–218

Monnier, Févotte, Hoff, & Klein, 1996; Monnier, Févotte, Hoff, vi+1 /vi = 2) and is also a zero order method, that is, the particle-
& Klein, 1997; Nallet et al., 1998; Puel et al., 2003a; Puel, size density is considered constant in each bin. The discretization
Févotte, & Klein, 2003b). is chosen to prevent the high computational complexity associ-
The major drawback of the Method of Classes is its depen- ated with the aggregation term. The authors identified four types
dency of the number or density functions on the adopted grid. of aggregation interactions that can change the total population
Since the number of discretized equations increases with the in a size range and derived expressions for each one of them
number of classes, high computational times can be neces- separately. The derived discretized expressions for the number
sary for the solution of a fine grid. Self-adaptive discretization of particles due to the mechanisms of aggregation, growth and
has already been proposed in order to reduce the number of nucleation are given, respectively, by Eqs. (19)–(21). When the
differential equations, without affecting the results. Another dis- three mechanisms are present in the process, Eqs. (19)–(21) must
advantage of the method is that the feasibility of the computation be combined in order to compute the evolution of the number of
is, likely, strongly dependent on the complexity and location of particles in each size range during time.
individual birth and death processes.  i−1

Numerical problems described in literature arise from the dNi 1
= Ni−1 2j−i+1 βi−1,j Nj + βi−1,i−1 Ni−1
2
possible loss of crystals, leading to an unbalanced mass bal- dt agg 2
j=1
ance during the crystal growth. In coagulation (aggregation)
processes, the method conserves particle mass, but underesti- i−1
 ∞

mates the number of entities (Kostoglou & Karabelas, 1994). − Ni 2j−i βi,j Nj − Ni βi,j Nj (19)
Kumar and Ramkrishna (1997) stress that the discretization j=1 j=i
equations for the Method of Classes have the potential to produce 
dNi
negative values for the number of particles in each class and
that the results from this technique oscillate about the analytical dt growth
solution, apart from the fact that the numerical solutions around 
2G r r
the location of a discontinuity are highly diffused, turning the = Ni−1 + Ni − 2 Ni+1 ,
(1 + r)Li r2 − 1 r −1
method not adequate to include growth. This is quite natural
Li+1 √
for finite-difference-type approximation as they have an infinite where r = = 2
3
(20)
velocity of propagation for the signal. Li
Puel et al. (2003a) extended the Method of Classes to a
bi-dimensional population balance, arguing that some organic 
dN1
systems can not be represented solely by one characteristic = B0 (21)
dimension, being necessary the consideration of two particle dt nucl
dimensions. The bi-dimensional Method of Classes predicts the The set of equations derived was internally consistent with
time-variation of the crystal habit, which must be a point of regard to mass and total numbers (zeroth and, ultimately, third
concern in some applications. They suggest the implementation moments, in a length basis).
of an adaptive bi-dimensional algorithm to improve the com- The DPB proposed by Hounslow et al. (1988) was extended
putation. In order to speed it up, the LSODAR was adopted by Litster, Smit, and Hounslow (1995) in order to permit the
to solve the resulting ODEs. In order to cope with possible use of an adjustable discretization with the form vi+1 /vi = 21/q ,
unbalanced mass balance, the area of the each class in the grid q ≥ 1, integer. The expression for the contribution of aggregation
was reduced, requiring an enormous number of classes (6 × 105 with the adjustable discretization is given by Eq. (22).
 i−S(q)−1
 q i−S(q−k+1)−k
dNi 2(j−i+1)/q   2(j−i+1)/q − 1 + 2−(k−1)/q
= βi−1,j Ni−1 Nj 1/q + βi−k,j Ni−k Nj
dt agg 2 −1 21/q − 1
j=1 k=2 j=i−S(q−k+2)−k+1
q i−S(q−k+1)−k+1
 
1 −2(j−i)/q + 21/q − 2−(k−1)/q
+ βi−q,i−q Ni−q
2
+ βi−k,j Ni−k+1 Nj
2 21/q − 1
k=2 j=i−S(q−k+2)−k+2
i−S(q)
 ∞

2(j−i)/q
− βi,j Ni Nj − βi,j Ni Nj (22)
21/q − 1
j=1 j=i−S(q)+1
q
classes), which would prevent the use of this model for use in where S(q) = s=1 s.
real time applications. The numerical solution is as more accurate as greater the
Hounslow, Ryall, and Marshall (1988) have devoted many value of q, but this demands more computational time.
works in the modeling of kidney stones formation, a process in The major drawback to the application of the method pro-
which only growth and aggregation are involved. The method posed by Litster et al. (1995) is its limitation in combined growth
developed by them discretize the particle-size
√ domain in a fixed and aggregation mechanisms at large times, although nucleation
geometric progression (Li+1 /Li = 3 2, which is equivalent to problems can be successfully solved with such method.
 C.B.B. Costa et al. / Computers and Chemical Engineering 31 (2007) 206–218 213

Kumar and Ramkrishna (1996a) propose a discretization the variation of the relevant density in the ith size range. In this
method for the PBE solution for breakage and agglomeration way, if the density in a section changes from steeply decreas-
mechanisms, in which the size (volume) domain is divided into ing to nearly uniform due to the particulate process, the pivot
sections of arbitrarily scales and a representative volume (pivot) moves from the lower end of the section to the middle. Eq. (25)
is chosen in each section. The main feature is not an attempt to expresses the movement of the pivots in each section and Eq.
approximate the continuous number density function on a suit- (26) computes the evolution of the number of particles.
ably fine scale, but rather to target calculation of properties of the
population of specific interest to an application without seeking j≥k

dxi 1
the complete number density function resulting from the process =
dt Ni
events. In the method, called fixed-pivot (FP) technique, events j, k
among particles that lead to the formation of particles of sizes vi ≤ (xj + xk ) < vi+1
other then the pivot ones are incorporated in the set of discretized  
equations, in such a way that the properties corresponding to two 1   
× 1 − δj,k (xj + xk ) − xi βj,k N j Nk
moments of interest are exactly preserved. The moments to be 2
preserved determine the two values of j in Eq. (9), here assigned
1  
(v) (1)
to ζ and ξ. These last two variables determine the value of η − b(xj )Nj B̄i,j − xi B̄i,j (25)
in Eq. (23), used in Eq. (24), which computes the evolution of Ni
j≥i
number of particles in each section.
⎧ ζ ζ ζ
⎪
⎪ v xi+1 − vξ xi+1 j≥k  
⎪
⎪ , xi ≤ v ≤ xi+1 dNi  1
⎨ xζ xξ − xξ xζ 
= 1 − δj,k βj,k Nj N k
i i+1 i i+1 dt 2
η= (23)
⎪
⎪
ζ ζ
vζ xi−1 − vξ xi−1 j, k
⎪
⎪ , xi−1 ≤ v ≤ xi
⎩ ζ ξ ξ ζ vi ≤ (xj + xk ) < vi+1
xi xi−1 − xi xi−1
M
 
 (1)
j≥k
  − Ni βj,k Nk + b(xj )Nj B̄i,j − b(xi )Ni (26)
dNi 1 
= 1 − δj,k ηβj,k N j Nk k=1 j≥i
dt 2
j, k (v) v
where B̄i,j = vii+1 vβ(v, xj ) dv.
xi−1 ≤ (xj + xk ) ≤ xi+1 The new technique, called moving-pivot (MP) technique,
M
 M
 shows that the correct discrete representation of the birth term

− Ni βj,k Nk + ni,k bk Nk − bi Ni (24) due to the coalescence of smaller entities (which determines the
k=1 k=1 movement of the front) is crucial for the accuracy of the numer-
ical solutions.
Kotoulas and Kiparissides (2004) and Krallis et al. (2004) The disadvantage of the two previous techniques of not
used the FP technique for the numerical solution of the PBE including nucleation and growth terms was overcome with the
applied to polymerization reactors. Stork, Tousain, Wieringa, work of Kumar and Ramkrishna (1997), which combines the
and Bosgra (2003) also used the FP technique with a geometric developed discretized technique with the method of character-
grid of factor 1.2, but in a process of emulsification. istics. The scale boundaries move with the growth rate of the
The main advantages of the method lie in the feature of pre- particles belonging to the respective interval (Eqs. (27) and (28)).
serving any two free of choice properties of the distribution The presence of nucleation poses a new difficulty because a sit-
and in the possibility of varying the coarseness of discretiza- uation arises when some or all of the nuclei become smaller than
tion. Furthermore, for appropriate values of involved parame- the smallest particle size represented in the bins, due to the bins
ters, it produces solutions that are exactly identical to those of movement. This problem is overcome by adding new bins, one
Hounslow et al. (1988) and Litster et al. (1995) with less com- at a time (Eqs. (29) and (30)), with zero population at the small
putation time. Nevertheless, the method is restricted to aggre- size end at regular intervals and the renumber of the old bins as
gation and breakage events, which makes the method useful in increasing sequence of integers. The number of equations that
polymerization–depolymerization and aerosol dynamics prob- need to be solved increases very rapidly, making the technique
lems (and indeed a lot of work which uses this numerical method computation-intensive. The effectiveness is restored by elimi-
is found in the open literature), but not in crystallization ones. nating some of the pivots to make the grid coarser in larger sizes
Apart from that, according to Kumar and Ramkrishna (1996a), and assigning them to the adjoining pivots in a way that the
over prediction is always present with the discretized equations desired properties of the population at the pivot in question are
for pure aggregation, unless special effort is made to prevent it. preserved. Eq. (31) provides for the calculation on the number
Due to this over prediction, Kumar and Ramkrishna (1996b) of particles in each section.
developed a technique that accounts for the variation of number
density in a size range, proposing the definition of the pivots in dvi
= G(vi ) (27)
such a way that their location in size range {vi , vi+1 } reflects dt
214 C.B.B. Costa et al. / Computers and Chemical Engineering 31 (2007) 206–218

dvi+1 populations of new ith and (i + 1)th meshes should satisfy Eqs.
= G(vi+1 ) (28)
dt (37) and (38).
 new  new  
dv1 dv0 (t) kv xi+1 xi+1 − kv xiold xiold old
= (29) Ni =  new  new
new   N (37)
dt dt kv xi+1 xi+1 − kv xinew xinew i

dx1 1 dv1 dv2    
= + (30) kv xiold xiold − kv xinew xinew old
dt 2 dt dt Ni+1 =  new  new
new   N (38)
kv xi+1 xi+1 − kv xinew xinew i
i≥j≥k
  The adaptive mesh method was applied to a potassium sulfate
dNi (t) 1 
= 1 − δj,k ηβj,k Nj (t)Nk (t) crystallization system, showing that the method is suitable for
dt 2 the simulation of crystallization processes.
xi−1 ≤(xj +xk )≤si+1

 Simultaneity of nucleation, growth and aggregation of parti-

M
 vi+1
 cles is essential to characterize processes such as precipitation
− Ni (t) βj,k Nk (t) + B0 dv (31)
vi and crystallization, among others, and the possibility of mod-
k=1
eling this processes are one of the advantages of the method.
Furthermore, the presence of moving discontinuities, which is
An adaptive mesh method is proposed by Lee et al. (2001) to
unavoidable due to the hyperbolic nature of the governing equa-
be used combined with the technique of Kumar and Ramkrishna
tion, is addressed with no additional difficulty.
(1997), conserving crystal number and mass during mesh adap-
The developments of Kumar and Ramkrishna (1996a, 1996b,
tation, being able to handle stiff nucleation problems of the char-
1997) allow the density function (particle distribution) evalua-
acteristic method. For the decision tolerance to adapt meshes,
tions as a consequence of distribution moments calculations.
the authors use the mesh function based on the first derivative
This fact makes the procedure interesting to broader applica-
of the solution, as in Eq. (32).
tions. For example, in many processes, like polymerization,
  crystallization and aerosol, the control of the full particle-size
f (v, t) = (v)2 + (vNv )2 = v 1 + Nv2 distribution (PSD) may be necessary, due to the strong depen-

 dence of the physico-chemical and mechanical properties of
∂N 2 products on the characteristics of the corresponding PSD (Kalani
= v 1 + (32)
∂v & Christofides, 2002). In order to control these processes, great
effort is made to determine off-line and on-line state variables
The meshes will be adapted to equidistribute the mesh func- values that lead to the achievement of a target distribution. These
tion. Therefore, the meshes having mesh function values larger calculations, which include optimization and control, require a
than the tolerance defined in Eq. (33) or in Eq. (34), a relative system model able to predict the full PSD.
equation in order to consider the size range where density func- Nopens, Beheydt, and Vanrolleghem (2005) compared the
tion are very small, will be refined. methods of Hounslow et al. (1988), the FP technique and the
M  MP technique applied to activated sludge flocculation, a pro-
i=1 v i 1 + Nv,i
2 cess in which only aggregation and breakage mechanisms have
TOL = (33) significant dynamics. The authors concluded that when both
M
mechanisms are present, the MP technique with a geometric
M 
grid of factor 2 is the most accurate method. A finer grid in the
i=1 v i (1/vref )2 + (Nv,i /Nref )2
TOLrel = (34) FP technique leads to an improvement in accuracy, but it is still
M worse than the one provided by the MP technique.
The idea is to eliminate meshes when the mesh function of the
solution is much smaller than the tolerance, in order to reduce 4.3. Method of weighted residuals
the calculation time. Mesh elimination, for conservation of both
number and mass, is made through Eqs. (35) and (36). Rawlings et al. (1992) used the global orthogonal collocation
i+1       for continuous crystallizers and orthogonal collocation on finite
k=i−1 Nk
old k xnew xnew − k xold xold
v i i v k k elements for batch mode. All simulations in this work were
Ni−1 =
new     (35)
kv xinew xinew − kv xi−1
new xnew
i−1
carried out using Petzold’s differential–algebraic equation
i+1       solver DASSL. Collocation and Galerkin FE algorithms were
k=i−1 Nk
old k xold xold − k xnew xi−1
new used by Nicmanis and Hounslow (1996, 1998) to solve the
v v
Ninew = k
 new  new k   i−1 (36) steady state population balance equation, using Lagrange cubic
kv x i xi − kv xi−1
new xnew
i−1
interpolation polynomials and evenly spaced nodes within
The study adopted the strategy of keeping the mass ratio in each element. Simulations were performed for a large range
the nuclei mesh to all meshes under a pre-specified constant. of indices of aggregation, and included, besides aggregation,
In this way, a new mesh is added when the mass in the nuclei breakage, nucleation and particle growth. The algorithms were
mesh is no longer negligible compared to the total mass. The found to be able to accurately predict the number density with
 C.B.B. Costa et al. / Computers and Chemical Engineering 31 (2007) 206–218 215

reasonably small numbers of elements and the collocation The integral approach can have more applications and it is
and Galerkin methods made almost identical predictions and easier to understand, having intuitive derivation in many cases.
converged in similar numbers of iterations. The Galerkin algo- It can be employed to any residence-time distributions in contin-
rithm, however, requires additional integrations that make this uous operations. Furthermore, it does not require the evaluation
method computationally much more expensive than collocation of the CSD in each time step, which can mean great computa-
method. The authors claim that the FE method proves to be tional time saving in applications where the detailed CSD is not
able to predict two orders of magnitude more accurate than the needed in each instant.
results of the DPB developed by Litster et al. (1995) and that The great disadvantage of this approach lies in its difficulty
a mixed collocation-Galerkin formulation can be used to avoid to cope with more complicated phenomena, like agglomeration.
the ill-conditioned matrices associated with growth problems.
The OCFE was also used by Crowley, Meadows, Kostoulas, 5.2. Kinetics decomposition
and Doyle III (2000), Immanuel et al. (2002) and Doyle III,
Harrison, and Crowley (2003), applied to the emulsion poly- Immanuel and Doyle III (2003) proposed other different
merization field, using Lagrange and Legendre interpolation approach based on employing individual rates of nucleation,
polynomials. growth and coagulation to update the PSD in a hierarchical
Mahoney and Ramkrishna (2002) improved two refinements framework, with a semi-analytical solution in the coagulation
in the Galerkin’s method on FE for the solution of PBEs for kernel, which reduces the computational requirement. Stiff-
precipitation systems, in order to reduce the traditional draw- ness is claimed to be eliminated by the decomposition of the
backs of this method: the time required for computation of the fast and the slow kinetics. The algorithm follows a two step
two-dimensional integrals arising from the aggregation integrals strategy—one to calculate the individual rates of nucleation,
and the difficulty in handling discontinuities that often arise in growth and coagulation, and the other to update the PSD based
simulations of seeded crystallizers. Linear basis functions were on these individual rates. The method discretizes the size particle
chosen. Computational costs were enormously reduced with the domain, but the proposed approach turns the method a com-
careful arrangement of invariant integrals of separable aggrega- pletely different mathematical perspective other than a strictly
tion models. The discontinuities are proposed to be specifically DPB.
tracked by the method of characteristics. It is found that the The method is applied to emulsion polymerization, but can
computation speed may be sufficient for dynamic online opti- be extended to other PBE systems. According to the authors,
mization. efficacy of the method in on-line feedback applications is guar-
A FE scheme was used by Rigopoulos and Jones (2003) with anteed from a computational standpoint, while not losing any
collocation linear elements and an upwind propagation of growth process information. In addition to this advantage, the current
term. The method is claimed to be computationally much faster code is very robust to a variety of operating conditions that one
than higher-order FE collocation methods, such as cubic splines, might encounter in industrial practice, including very high nucle-
or integral FE schemes, such as the Galerkin’s. ation rates.

5. Alternative numerical methods for solving the PBE 5.3. Closed-form solution with the method of
characteristics
This section deals with some alternative methods to solve
the PBE that are not included in the general classification Lee, Lee, Yang, and Mahoney (2002) used a closed-form of
of method of moments, discretization and weighted residual solution of the PBE, when the growth rate takes on a separa-
methods. ble form of Gt GL in which Gt is the time-dependent part of
the growth rate and GL is the size-dependent part. The solu-
5.1. Integral method tion involves an integral along the characteristic curve and the
analytical solution is given by the set of Eq. (39):
A different approach in the modeling of particulate processes
V (t0 ) GL (Lb )
was made by Wynn and Hounslow (1997) by following the n(L, t) = nt (Lb ) for Lb (L, t) ≥ 0
progress of particles since their insertion (through the feed flow V (t) 0 GL (L)
(39)
V (tb ) GL (0)
or generation as nuclei) and computing their contribution to the n(L, t) = nL (tb ) for tb (L, t) ≥ t0
eventual crystal size distribution (CSD). This approach leads to V (t) 0 GL (L)
integrals from earlier times and smaller sizes until the present where t0 is the start time of growth reaction. Lb and tb are the birth
time and size, opposing to the traditional treatment of deriva- size and birth time of the particle of size L at time t, respectively,
tives. It reduces to the characteristics method for a batch case, and can be obtained by solving the following Eqs. (40) and (41):
without nucleation and with length-dependent growth: particles  L  t
1
in an interval are described by their characteristic curve, defined dl = Gt (t  ) dt  (40)
by the growth rate equation. Due to the differences in the growth Lb GL (l) t0
rates between the lowest and the highest interval boundaries, the  L  t
1
interval width changes, while the number of particles remains dl = Gt (t  ) dt  (41)
unchanged. 0 GL (l) tb
216 C.B.B. Costa et al. / Computers and Chemical Engineering 31 (2007) 206–218

This method is limited to absence of aggregation and break-

age and to a separable form of the growth rate. The simulations
are managed quickly and accurately, turning possible its use in
on-line model-based control.

5.4. Parallel parent and daughter classes (PPDC)

A novel approach has been suggested by Bove, Solberg, and

Hjertager (2005), but the technique is still limited to breakage
and agglomeration problems, although the authors promise to
extend it to include nucleation and growth in order to be appli-
cable to wider classes of processes in which PBE is used.
The name of the technique comes from the generation of
several distinct grids, parallel in the internal coordinate domain,
each grid depending on the specific mechanism (aggregation,
breakage). Breakage daughter classes are also parallel in the
internal coordinate domain. In the PPDC method, the PBE is
discretized in time using an explicit Eulerian time marching
method and the density function is approximated by a finite set
of Dirac’s functions in order to assign function values just in the
pivot sizes in each bin of the partition domain.
The advantages are the need to track very few classes, a better
closure of the agglomeration processes and the improvement
of the breakage mechanism accuracy with the increase in the
number of breakage daughter classes.

6. Numerical results with the method of classes

This section is devoted to present a case study as an exempli-

fication of a method application, in order to outline its numerical
issues. The Method of Classes was selected, for its ease of Fig. 1. Evolution of the number of particles per volume unit of suspension for
implementation, and a FORTRAN code was written in order to the first and last classes.
simulate the batch cooling crystallization of adipic acid. Details
about the code structure and the system features can be found in with larger time steps and at the smallest granulometric classes.
Costa, da Costa, and Maciel Filho (2005). In the 25th granulometric class, the oscillation is of the order of
In order to test the numerical ability, many runs were made, 103 times smaller than the ones in the 1st class. Furthermore,
which allowed acquisition of skill at the method. Some remarks for a time step of 0.5 s, there is no noticeable difference between
may be, then, outlined about a few numerical issues. the Euler and the Runge–Kutta solution (results are superim-
As already pointed out by Kumar and Ramkrishna (1997), posed). It is evident, therefore, that the choice of an inappropriate
the set of ODEs produced with the Method of Classes (Eq. (18)) integrator method could cause serious errors in crystallization
has the potential to produce negative values of the number of calculations. The Fortran LSODAR code, developed by Petzold
crystals in each class, when the kinetics that drives crystals out and Hindmarsh, is even better than the Runge–Kutta technique,
of a specific granulometric class is greater than the one respon- since the former is a solver for ODEs with automatic method
sible for the appearance of them. This is a point of concern that switching for stiff and nonstiff problems and with root-finding.
was not made clear in the development of the method. In the Details about the solver can be found in Petzold (1980).
developed code (Costa et al., 2005), it was dealt merely by set- In order to evaluate the influence of the number of granu-
ting negative values of number of particles to zero. Apart from lometric classes on the numerical result, the same system was
this point, the set of ODEs constitutes a very stiff-nonlinear sys- simulated, with the same range of crystal size (from 0.01 to
tem, which can cause oscillation in the numerical solution. It 1735.25 ␮m) and with the same amount (in mass) and size of
was already remarked in Costa et al. (2005) that some integra- seeds, but differing in the chosen number of classes (and, as a
tors are not suitable for the system solution. Fig. 1 illustrates this consequence, in the values of the characteristic sizes in each
feature, showing the evolution of the number of crystals during class). The set of ODEs was solved using LSODAR. Fig. 2
batch time in the first and in the 25th granulometric classes in depicts the final CSD, both in mass and number percentage,
a run with 25 granulometric classes. The results were gener- for 5, 25 and 50 classes. The particle size in the right graphic
ated with the Euler method with three different time steps and of Fig. 2 was presented only up to 600 ␮m in order to increase
with Runge–Kutta. The Euler integrator produces a lot of oscil- its visualization in the smaller sizes range. It is clear that the
lation in the solution and this oscillation is more pronounced number of classes is a key variable in the correct representation
 C.B.B. Costa et al. / Computers and Chemical Engineering 31 (2007) 206–218 217

Fig. 2. Final CSD for different number of classes.

of the system. As the number of considered classes increases great advantage of being general and able to handle any for-
in the same size range, the solution is more accurate. The solu- mulation of the PBE. The computational burden was improved
tions with 25 and 50 classes are closer to each other than that in the work of Mahoney and Ramkrishna (2002), turning the
of just 5 granulometric classes, but, since they are not superim- computational speed sufficient for dynamic online optimization.
posed, 50 classes is still not a enough number of classes for the As other presented methods, the integral approach has its
method convergence. The main idea in the selection of optimal limitations in the comprised kinetic mechanisms. The integral
number of granulometric class is to increase it to the limit that approach can save great computational time when detailed CSD
produces no more change in the results (system minimization is not needed in each instant, but cannot easily handle with
realization). Furthermore, the greater the number of granulomet- agglomeration and breakage.
ric classes, the better is the accuracy in the mass balance, a fact
easily observed with the application of the Method of Classes. Acknowledgments
Though the previous remarks represent disadvantages
observed in the applicability of the method, it is easy to be imple- The authors would like to thank the Unicamp Scholarship
mented, leading to a very fast code, which took approximately Program and CNPq.
35 s to simulate a batch with 3000 time iterations in a AMD
XP 2.4 GHz Processor with 50 granulometric classes. This is an References
important feature for on-line applications.
Blandin, A. F., Mangin, D., Nallet, V., Klein, J. P., & Bossoutrot, J. M. (2001).
7. Conclusions Kinetics identification of salycilic acid precipitation through experiments
in batch stirred vessel and a T-Mixer. Chemical Engineering Journal, 81,
91–100.
The present paper made a brief review in the numerical solu- Bove, S., Solberg, T., & Hjertager, B. H. (2005). A Novel algorithm for
tion of the population balance model, discussing the difficulty solving population balance equations: The parallel parent and daughter
in solving it, particularly when a complete set of mechanisms classes. Derivation, analysis and testing. Chemical Engineering Science,
(nucleation, growth, agglomeration and breakage) is present, 60, 1449–1464.
like in crystallization systems. Some important numerical meth- Costa, C. B. B., da Costa, A. C., & Maciel Filho, R. (2005). Mathematical
modeling and optimal control strategy development for an adipic acid crys-
ods developed in the literature in DPBs and in weighted residuals tallization process. Chemical Engineering and Processing, 44, 737–753.
methods, besides different approaches, were presented and the Crowley, T. J., Meadows, E. S., Kostoulas, E., & Doyle III, F. J. (2000). Control
major advantages and drawbacks were raised. Attention was of particle size distribution described by a population balance model of semi-
given to the issues of numerical problems, computational effort batch emulsion polymerization. Journal of Process Control, 10, 419–432.
and the correctness of prediction of the total number of particles David, R., & Bossoutrot, J.-M. (1996). Crystallization and precipitation
engineering—VII. The modelling of sodium perborate tetrahydrate crys-
and mass conservation. The discussion of the most used numer- tallization from solution. Chemical Engineering Science, 51, 4939–
ical methods and their limitations intends to be useful for the 4947.
reader dealing with crystallization problems. David, R., Paulaime, A.-M., Espitalier, F., & Rouleau, L. (2003). Modeling
DPBs are a more natural choice when aggregation and break- of multiple-mechanism agglomeration in a crystallization process. Powder
age are included in the PBE. Notable developments are the ones Technology, 130, 338–344.
David, R., Villermaux, J., Marchal, P., & Klein, J. P. (1991). Crystallization and
of Kumar and Ramkrishna (1996a, 1996b, 1997), which allow precipitation engineering—IV. Kinetic Model of adipic acid crystallization.
to a free choice of grid and preserve two properties of interest Chemical Engineering Science, 46, 1129–1136.
(moments). Their method that includes nucleation and growth Doyle III, F. J., Harrison, C. A., & Crowley, T. J. (2003). Hybrid model-based
requires an adaptive grid and Lee et al. (2001) proposed an adap- approach to batch-to-batch control of particle size distribution in emulsion
tive mesh method able to handle stiff nucleation problems, being polymerization. Computers and Chemical Engineering, 27, 1153–1163.
Frank, R., David, R., Villermaux, J., & Klein, J. P. (1988). Crystallization and
able to deal with crystallization problems. precipitation engineering—II. A chemical reaction engineering approach to
The weighted residuals methods are of more difficult imple- salicylic acid precipitation: Modelling of batch kinetics and application to
mentation and are computationally demanding, but have the continuous operation. Chemical Engineering Science, 43, 69–77.
218 C.B.B. Costa et al. / Computers and Chemical Engineering 31 (2007) 206–218

Hounslow, M. J. (1998). The population balance as a tool for understanding Matthews, H. B., Miller, S. M., & Rawlings, J. B. (1996). Model identification for
particle rate processes. Kona Powder and Particle, 16. crystallization: Theory and experimental verification. Powder Technology,
Hounslow, M. J., Ryall, R. L., & Marshall, V. R. (1988). A discretized popu- 88, 227–235.
lation balance for nucleation, growth and aggregation. AIChE Journal, 34, Matthews, H. B., & Rawlings, J. B. (1998). Batch crystallization of a photo-
1821–1832. chemical: Modeling, control, and filtration. AIChE Journal, 44, 1119–1127.
Immanuel, C. D., Cordeiro, C. F., Sundaram, S. S., Meadows, E. S., Crowley, Miller, S. M., & Rawlings, J. B. (1994). Model identification and control strate-
T. J., & Doyle III, F. J. (2002). Modeling of particle size distribution in gies for batch cooling crystallizers. AIChE Journal, 40, 1312–1327.
emulsion co-polymerization: Comparison with experimental data and para- Monnier, O., Févotte, G., Hoff, C., & Klein, J. P. (1996). An advanced calori-
metric sensitivity studies. Computers and Chemical Engineering, 26, 1133– metric approach for population balance modelling in batch crystallization
1152. processes. Thermochimica Acta, 289, 327–341.
Immanuel, C. D., & Doyle III, F. J. (2003). Computationally efficient solu- Monnier, O., Févotte, G., Hoff, C., & Klein, J. P. (1997). Model identification
tion of population balance models incorporating nucleation, growth and of batch cooling crystallizations through calorimetry and image analysis.
coagulation: Application to emulsion polymerization. Chemical Engineer- Chemical Engineering Science, 52, 1125–1139.
ing Science, 58, 3681–3698. Nallet, V., Mangin, D., & Klein, J. P. (1998). Model identification of batch
Kalani, A., & Christofides, P. D. (2002). Simulation, estimation and control of precipitations: Application to salicylic acid. Computers and Chemical Engi-
size distribution in aerosol process with simultaneous reaction, nucleation, neering, 22, S649–S652.
condensation and coagulation. Computers and Chemical Engineering, 26, Nicmanis, M., & Hounslow, M. J. (1996). A finite element analysis of the steady
1153–1169. state population balance equation for particulates systems: Aggregation and
Kostoglou, M., & Karabelas, A. J. (1994). Evaluation of zero order methods for growth. Computers and Chemical Engineering, 20, S261–S266.
simulating particle coagulation. Journal of Colloid and Interface Science, Nicmanis, M., & Hounslow, M. J. (1998). Finite-element methods for steady-
163, 420–431. state population balance equations. AIChE Journal, 44, 2258–2272.
Kotoulas, C., & Kiparissides, C. (2004). Time evolution of particle size distri- Nopens, I., Beheydt, D., & Vanrolleghem, P. A. (2005). Comparison and pitfalls
bution in suspension polymerization reactors. In Proceedings of BatchPro of different discretised solution methods for population balance models: A
Symposium. simulation study. Computers and Chemical Engineering, 29, pp. 367, 377.
Krallis, A., Meimaroglou, D., Saliakas, V., Chatzidoukas, C., Kotoulas, C., & Petzold, L. R. (1980) Automatic selection of methods for solving stiff and nons-
Kiparissides, C. (2004). Dynamic optimization of molecular weight distri- tiff systems of ordinary differential equations. Sandia National Laboratories
bution in batch free radical polymerization reactors using the orthogonal Report SAND80-8230.
collocation on finite elements and the fixed pivot technique. In Proceedings Pilon, R., & Viskanta, R. (2003). Modified method of characteristics for solving
of BatchPro Symposium. population balance equations. International Journal for Numerical Methods
Kumar, S., & Ramkrishna, D. (1996a). On the solution of population balance in Fluids, 42, 1211–1236.
equations by discretization—I. A fixed pivot technique. Chemical Engineer- Puel, F., Févotte, G., & Klein, J. P. (2003a). Simulation and analysis of indus-
ing Science, 51, 1311–1332. trial crystallization processes through multidimensional population balance
Kumar, S., & Ramkrishna, D. (1996b). On the solution of population balance equations. Part 1: A resolution algorithm based on the method of classes.
equations by discretization—II. A moving pivot technique. Chemical Engi- Chemical Engineering Science, 58, 3715–3727.
neering Science, 51, 1311–1332. Puel, F., Févotte, G., & Klein, J. P. (2003b). Simulation and analysis of indus-
Kumar, S., & Ramkrishna, D. (1997). On the solution of population balance trial crystallization processes through multidimensional population balance
equations by discretization—III. Nucleation, growth and aggregation of par- equations. Part 2: A study of semi-batch crystallization. Chemical Engineer-
ticles. Chemical Engineering Science, 52, 4659–4679. ing Science, 58, 3729–3740.
Lee, K., Lee, J. H., Yang, D. R., & Mahoney, A. W. (2002). Integrated run- Ramkrishna, D., & Mahoney, A. W. (2002). Population balance modeling.
to-run and on-line model-based control of particle size distribution for a Promise for the future. Chemical Engineering Science, 57, 595–606.
semi-batch precipitation reactor. Computers and Chemical Engineering, 26, Randolph, A. D., & Larson, M. A. (1971). Theory of particulate processes:
1117–1131. Analysis and techniques of continuous crystallization. New York: Academic
Lee, G., Yoon, E. S., Lim, Y.-II, Le Lann, J. M., Meyer, X.-M., & Joulia, X. Press.
(2001). Adaptive mesh method for the simulation of crystallization pro- Rawlings, J. B., Miller, S. M., & Witkowski, W. R. (1993). Model identification
cesses including agglomeration and breakage: The potassium sulfate system. and control of solution crystallization processes: A review. Industrial and
Industrial and Engineering Chemistry Research, 40, 6228–6235. Engineering Chemistry Research, 32, 1275–1296.
Litster, J. D., Smit, D. J., & Hounslow, M. J. (1995). Adjustable discretized pop- Rawlings, J. B., Witkowski, W. R., & Eaton, J. W. (1992). Modelling and control
ulation balance for growth and aggregation. AIChE Journal, 41, 591–603. of crystallizers. Powder Technology, 69, 3–9.
Ma, D. L., & Braatz, R. D. (2003). Robust identification and control of batch Rigopoulos, S., & Jones, A. G. (2003). Finite-element scheme for solu-
processes. Computers and Chemical Engineering, 27, 1175–1184. tion of the dynamic population balance equation. AIChE Journal, 49,
Ma, D. L., Tafti, D. K., & Braatz, R. D. (2002). Optimal control and simula- 1127–1139.
tion of multidimensional crystallization processes. Computers and Chemical Stork, M., Tousain, R. L., Wieringa, J. A., & Bosgra, O. H. (2003). A MILP
Engineering, 26, 1103–1116. approach to the optimization of the operation procedure of a fed-batch emul-
Mahoney, A. W., & Ramkrishna, D. (2002). Efficient solution of population sification process in a stirred vessel. Computers and Chemical Engineering,
balance equations with discontinuities by finite elements. Chemical Engi- 27, 1681–1691.
neering Science, 57, 1107–1119. Verkoeijen, D., Pouw, G. A., Meesters, G. M. H., & Scarlet, B. (2002). Pop-
Marchal, P., David, R., Klein, J. P., & Villermaux, J. (1988). Crystallization ulation balances for particulate processes—A volume approach. Chemical
and precipitation engineering—I. An efficient method for solving popula- Engineering Science, 57, 2287–2303.
tion balance in crystallization with agglomeration. Chemical Engineering Wynn, E. J., & Hounslow, M. J. (1997). Integral population balance equations
Science, 43, 59–67. for growth. Chemical Engineering Science, 52, 733–746.