Chemical Engineering Science 58 (2003) 3715 – 3727

www.elsevier.com/locate/ces

Simulation and analysis of industrial crystallization processes through

multidimensional population balance equations. Part 1: a resolution
algorithm based on the method of classes
F. Puel∗ , G. F,evotte, J. P. Klein
LAGEP UMR CNRS 5007, Universite Lyon 1, ESCPE. Bât. 308G, 43 bld. du 11 novembre 1918. F-69622 Villeurbanne, France
Received 19 July 2002; received in revised form 31 March 2003; accepted 14 May 2003

Abstract

In order to obtain constant solid properties with particles exhibiting a low order of symmetry, it is necessary to monitor and to control
several distributed parameters characterising the crystal shape and size. A bi-dimensional population balance model was developed to
simulate the time variations of two characteristic sizes of crystals. The nonlinear population balance equations were solved numerically
over the bi-dimensional size domain using the so-called method of classes. An e9ort was made to improve usual simulation studies
through the introduction of physical knowledge in the kinetic laws involved during nucleation and growth phenomena of complex organic
products. The performances of the simulation algorithm were successfully assessed through the reproduction of two well-known theoretical
and experimental features of ideal continuous crystallization processes: the computation of size-independent growth rates from the plot of
the steady-state crystal size distribution and the possibility for MSMPR crystallizers to exhibit low-frequency oscillatory behaviours in
the case of insu;cient secondary nucleation.
? 2003 Elsevier Ltd. All rights reserved.

Keywords: Bi-dimensional population balance; Modelling; Dynamic simulation; Morphology; Nucleation kinetics; Crystallization

1. Introduction during the process have their origins on the one hand in the
physico-chemical structure of the crystals (which is induced
Crystallization is the @rst and probably the main unit op- by the molecular structure of the product), on the other hand
eration dealing with solids in chemical plants. There is a real in the presence of solvents, impurities and by-products dis-
challenge to produce crystals with speci@c properties allow- solved in the mother-liquors. These chemical compounds
ing to perform downstream operations more easily thanks modify both the nucleation and growth mechanisms even
to mastered @lterability, low @nes concentration, low trend at very low concentrations (some hundreds of ppm). Con-
to cake on storage, good Aow ability, etc. Consequently, sequently, in order to obtain constant solid properties with
in the industrial practice, the control of the chemical pu- particles exhibiting a low order of symmetry, it is necessary
rity is no more su;cient and the control of both crystal to monitor and to control several characteristic sizes and/or
habit and crystal size distribution (CSD) can be critically internal shape factors of the crystals.
important. Most past studies were performed using prod- Modelling particulate processes such as crystallization
ucts which simple and isotropic shape (spherical, cubic, oc- is possible using the formalism of population balance
tahedric, etc.) could be described by a single characteristic equations (PBE) which was introduced by Hulburt and
dimension. Obviously, such approaches become question- Katz (1964), and developed later by Randolph and Larson
able in the case of the crystallization of organics products (1988). Although these authors were primarily interested
presenting anisotropic morphologies that sometimes vary in crystallization, their book describes a generic approach.
during the crystallization (Puel, Marchal, & Klein, 1997). Ramkrishna (2000) also considers the biological cell growth
The phenomena leading to modi@cations of crystal habits process as a particulate process and @rmly establishes pop-
ulation balance modelling as a basic theoretical framework
∗ Corresponding author. for all particle processes. In the @eld of industrial crystal-
E-mail address: puel@lagep.univ-lyon1.fr (F. Puel). lization, previous works usually consider mono-dimensional

0009-2509/03/$ - see front matter ? 2003 Elsevier Ltd. All rights reserved.
doi:10.1016/S0009-2509(03)00254-9
3716 F. Puel et al. / Chemical Engineering Science 58 (2003) 3715 – 3727

particles (Garside, 1985), and most reported applications examples, using a constant agglomeration kernel, where
deal with simple inorganic products such as sodium nitrate analytical solutions of the PBEs can be derived. Patil and
or potassium alum. In the eighties, few authors focused their Andrews (1998) presented an analytical solution to contin-
e9orts on organic crystals: Franck, David, Villermaux, and uous population balance models describing Aoc coalescence
Klein (1988) with salicylic acid, or Marchal, David, Klein, and breakage.
and Villermaux (1988) with adipic acid. Here again, simula- Generally, the complexity of the crystallization systems
tions were successfully performed using mono-dimensional encountered with organic products precludes the existence
PBEs. Nevertheless, such approaches are not suitable to of analytic solutions and requires the development and adap-
describe the evolution of populations of organic anisotropic tation of numerical techniques. As far as such techniques
crystals which indeed require multidimensional population are concerned, three approaches are more frequently used:
balances. method of moments, discretization techniques and @nite ele-
In the past 10 years, the capacity to design population ments methods. Hulburt and Katz (1964) suggested the ana-
balance models from experimental data was consider- lytical methods of moments. The PBEs are @rst transformed
ably enhanced, due to the appearance of sophisticated on into a set of ordinary di9erential equations—the moment
line measurement techniques through laser technologies equations—which can be solved numerically. The moments
or image analysis. Measuring CSD in-line or monitoring of order zero, one and two can be assessed experimentally.
habit transient is now possible, even though the physical In this case, the calculation can be compared with experi-
meaning of the observed characteristic size(s) sometimes ence. This technique was used in organic precipitation for
remains unclear. Since few years, the use of multidimen- the estimation of kinetic parameters (Franck et al., 1988).
sional PBEs comes out for di9erent particulates processes. Hounslow and Wynn (1993) may also be mentioned; who
In the @eld of granulation, Iveson (2002) states that current have developed short cut techniques based on the method
one-dimensional models of wet granulation are inadequate, of moments, and applied them to hydrocyclones and to con-
the author presents a general four-dimensional population tinuous crystallizers.
balance equation model which remains to be solved nu- A second class of numerical methods to solve PBEs is
merically. Ramkrishna and Mahoney (2002) underline that referred to as discretized population balance which in the
solving such multidimensional integro-di9erential equations @eld of crystallization is also referred to as “discretization
is exceedingly di;cult and that this has impeded their ap- sizing technique”. The technique consists in transforming
plication. Few authors have already solved bi-dimensional the PBEs into a set of ODEs through the discretization of
population balance models. Xiong and Pratsinis (1993) de- the particle sizes. The set of ODEs is then solved along
velop a bi-dimensional population balance equation to relate with the coupled solute balance equations by an iteration
the evolution of both volume and surface of aerosol titania technique at each time step. There are as many equations
particles during gaz phase powder production. Ma, Tafti, as discretized size domains considered and two main for-
and Braatz (2002) use a multidimensional CSD in order mulations of these ODEs exist. The evolution of the CSD
to compute optimal control strategies. Heinrich, Peglow, can be simulated using either a population density function
and Morl (2002) develop a bi-dimensional balance model (L; t) (David, Marchal, & Marcant, 1995) or population
for a Auidized bed granulator, validated on a continuous number function N (L; t) (Marchal et al., 1988; Hounslow
semi-industrial scale. and Wynn, 1993). In the @rst case, the spatial derivatives of
Population balance equations are strongly nonlinear and the population density functions are approximated by @nite
their coe;cients often depend indirectly on number density di9erences. The main di;culty of this formulation arises
functions. Nevertheless, these equations admit analytical from the nucleation step because the calculation of the dis-
solutions in the simpler cases: the steady-state continu- tributed nucleation rates is delicate. In the case of the pre-
ous mixed suspension mixed product removal crystallizer cipitation of calcium oxalate and silver bromide, David et al.
(MSMPR), with clear feed, negligible agglomeration and (1995) and Muhr, David, Villermaux, and Jezequel (1996)
breakage events and size-independent growth (Mac Cabe’s reported numerical recipes to circumvent this problem, The
law) is a well-known theoretical example. In this case, the occurrence of negative values of the density function during
mono-dimensional density function presents an exponential the computation may be observed due to @nite di9erence
form (Mullin, 2001; Mersmann, 2001). In Section 4, this equations coupled with inadequate discretized sizes. Muhr
example is addressed in the case of bi-dimensional popula- et al. (1996) overcame this numerical issue using a method
tion balance equations. The R-z analytical model was also of upwind di9erencing originally developed in Auid me-
developed by Randolph and Larson (1988) to represent clas- chanics. An analysis of the performance of 12 @nite di9er-
si@cation and @nes dissolution continuous crystallizers. This ence schemes for the resolution of a precipitation problem
model is often used to analyze the behaviour of industrial involving growth and nucleation was presented by Wojcik
plants. More recently Pan and Mersmann (1992) have pro- and Jones (1998). The second formulation of the discretized
posed few simpli@cations in order to transform the involved population balance methods is referred to as the method
PDEs into ODEs, using nucleation models proposed by Kind of classes; it consists in integrating the PBEs over short
and Mersmann (1990). Marchal (1989) also quotes two intervals of size. The population balances are expressed in
 F. Puel et al. / Chemical Engineering Science 58 (2003) 3715 – 3727 3717

terms of number population functions instead of number (Fig. 1b). An internal elongation shape factor can also be
population density functions, which allows to compute the de@ned as the ratio of the length L to the width L
. To take
number of crystals for each class of size. The main interest into account either two sizes or one of these two sizes and
of this formulation over the previous one is that the numer- the elongation shape factor, a bi-dimensional population bal-
ical processing of the nucleation step is made easier. The ance has to be established and solved.
method was successfully applied by Marchal et al. (1988)
to both semi-batch and continuous crystallization of adipic 2.1. Bi-dimensional population balance equations
acid in water. However, a major drawback arises from these
latter discretization techniques: a dependence of the number In the following, the population of crystals is described
or density functions on the size spacing can easily be by the number population density function (L; L
; t). It is
observed. Self adaptive one dimensional size class dis- assumed that the CSD does not depend on spatial coordinates
cretization was then proposed in order to reduce the num- in the well-mixed crystallizer. During the time-interval dt,
ber of di9erential equations, without a9ecting the results the balance of crystals in the bi-dimensional size domain
(Puel, 1994; Muhr et al., 1996). delimited by L; L + dL and L
; L
+ dL
can be written as
The last group of methods assumes that the mathematical follows:
form of the distribution function over the size domain is 1 @ @
known. Due to the extensive size domain encountered in [ (L; L
; t)VT (t)] + [G(L; t) (L; L
; t)]
VT (t) @t @L
crystallization, the function is divided in segments and the
technique is referred to as @nite elements method. Finite el- @
+ [G
(L
; t) (L; L
; t)]
ements use local basis functions, Rawlings, Witkowski, and @L

Eaton (1992) recommend the choice of the basis functions F(t) (L; L
; t) − Fin (t)

in (L; L ; t)
to depend on the features of the dynamics to be character- +
VT (t)
ized. For example, Laguerre polynomials are relevant for
continuous Aow crystallizers since the steady-state solution =RN (L; L
; t) + RA (L; L
; t) − RB (L; L
; t): (1)
is exponential in the range [0 + ∞[. Such crystallizer For the crystallization of hydroquinone, primary and sec-
con@guration can lead to cyclic behaviours that Laguerre ondary nucleation, and growth mechanisms should be con-
polynomials can satisfactorily represent. The same au- sidered while in the following, breakage and agglomeration
thors indicate that Laguerre polynomials are irrelevant for were found to be of second order of importance. Hence-
modelling batch crystallization operations. Other possible forth, these phenomena were not taken into account in the
choices are the Dirac delta functions, resulting in collocation PBE. Eq. (1) was then simpli@ed to yield the following
methods, or the weighted functions resulting in Galerkin’s expression:
techniques. Nicmanis and Hounslow (1998) have com- 1 @ @
pared Galerkin and collocation formulations for steady-state [ (L; L
; t)VT (t)] + [G(L; t) (L; L
; t)]
VT (t) @t @L
systems using adaptive mesh generation. Mahoney and
Ramkrishna (2002) gave a recent overview on the resolu- @
tion of PBEs using @nite elements for precipitation systems. + [G
(L
; t) (L; L
; t)]
@L

This paper presents a bi-dimensional model applied to
F(t) (L; L
; t) − Fin (t)

in (L; L ; t)
the crystallization of hydro-quinone in water. Section 2 de- +
scribes the theoretical formulation of the population bal- VT (t)
ance equations, the nucleation and growth kinetic equations = RN (L; L
; t) (2)
and the mass balance equations. Section 3 presents the res-
olution of the bi-dimensional population balance using the with

∞
∞
L∗
L∗
method of classes. An evaluation of the simulation program
RN dL dL
= RN dL dL

is then presented and discussed in Section 4. Such evalu- 0 0 0 0
ation is based on the ability of the program to reproduce
= rN 1 + r N 2 : (3)
well-known theoretical features of continuous crystallization
processes. Eq. (3) expresses possible primary and secondary nucle-
ation rates, rN1 and rN2 , as the generation of new crystals of
∗
critical size L∗ ; L
upon formation, so that, in the follow-
2. Modelling multidimensional crystallization processes ing, the smallest class considered in the population balance
∗
is delimited by 0; L∗ ; L
.
Hydroquinone raises many interesting industrial prob-
lems which were partly addressed by Puel et al. (1997). 2.2. Kinetic nucleation and growth equations
As displayed in Fig. 1a, hydroquinone crystals exhibit a
rod-like habit which can rather satisfactorily be represented If no crystal or any other particles (e.g. solid impurities)
as bi-dimensional parallelepipeds with dimensions L and L
is present in solution, homogeneous nucleation takes place
3718 F. Puel et al. / Chemical Engineering Science 58 (2003) 3715 – 3727

Fig. 1. (a) SEM picture of a typical crystal of hydroquinone and (b) bi-dimensional approximation of the corresponding rod-like particles.

when a relatively high level of supersaturation is obtained. The growth of crystals in the supersaturated solution is a
The presence of foreign surfaces such as solid impurities or complex process where both surface integration and di9u-
items of equipment in the crystallizer allows heterogeneous sive mechanisms should be considered. In almost all pub-
nucleation to occur with lower supersaturation values. In the lished studies, the relationship between the growth rates
event that particles crystallizing from the solution are ab- and the relative supersaturation is assumed to be described
sent, the formation of particles on such surfaces is referred through simple power laws. Such approximation is far from
to as primary nucleation. Moreover, if crystals are already being realistic:
dispersed in the crystallizing medium, secondary nucleation
can occur at supersaturation levels that are signi@cantly 1. It is well-known that several solid integration mecha-
lower than those at which primary nucleation takes place. nisms can be observed during the crystal growth (i.e.
In this case, many various elementary processes are likely Birth and Spread Model, Hillig (1966); Screw Disloca-
to be involved, including micro-breakage and abrasion, con- tion Models, Burton, Cabrera, and Frank (1951)) which
tacts with existing particles and/or crystallizer surfaces— cannot easily be reduced to basic power laws,
all of which might result in the production of secondary 2. As usual for mass transfer phenomena, di9usive limita-
nuclei. tions are likely to happen if the rate of solid integration
For the system of concern, primary nucleation is assumed is very high. The overall growth rate can therefore be
to initiate the crystallization process according to the fol- more or less limited by di9usion during the crystalliza-
lowing kinetic expression: tion process. In such conditions, the same growth model
  will neither be valid during the whole crystallization pro-
B cess, nor for the various crystal surfaces considered in
rN 1 = A exp − 2 ; (4)
ln  the PBEs.
where A; B are coe;cients that can be determined exper- According to Fig. 1b, two di9erent solid faces can be consid-
imentally and have complex physical meaning (see e.g. ered in the modelling of the crystallization of hydroquinone,
Mersmann, 1996). which have their own growth rates G and G
:
In addition to an obvious dependence on the level of su- dL dL

persaturation, secondary nucleation is generally considered G= and G
= : (6)
dt dt
to be promoted by the suspension density:
The well-known @lm model can be used to relate the
rN 2 = kN · n MTk ; (5) growth as a two step mechanism where the molecules of so-
lute are transported by di9usion and/or convection, and then
where the kinetic parameter kN is generally assumed incorporated onto the solid surface through several possible
to be related to the stirring power and to exhibit a mechanisms. It has been demonstrated that limitations due
temperature-dependency according to Arrhenius’s law. to heat transfer are generally negligible, especially when
Exponent n which does not depend on temperature, lies the solvent is water (Mersmann, 2001).
between 0.5 and 2.5. Exponent k is generally assumed to (1) Di9usive/convective mass transfer through a @lm
be of the order of 1, and to slightly depend on temperature around the surface of the growing crystal is assumed to be
(Garside, 1985). driven by the di9erence between the bulk concentration C
 F. Puel et al. / Chemical Engineering Science 58 (2003) 3715 – 3727 3719

and the concentration at the crystal-solution interface, de- $ and $
decrease if the crystal growth is strongly


noted below Ccs (t) and Ccs (t), respectively, for both faces: di9usion-controlled, while $ and $
= 1 if the growth rate
only depends on the kinetics of solids incorporation.
t (t) = kd (C(t) − Ccs (t)) and One can easily show that the growth rates are @nally given
t
(t) = kd (C(t) − Ccs


(t)); (7) by the following expressions which assimilate the real par-
ticles to bi-dimensional parallelepipeds:
where the mass transfer coe;cient, kd , is assumed to be Ms
the same for both growth rates G(t); G
(t) as it mainly de- G(t) = 2 $ · kint (C(t) − C ∗ (T (t)) jint ;
#s
pends upon the hydrodynamic conditions around the crystal
surface. kd can be assessed using various published exper- Ms


G
(t) = 2 $ kint (C(t) − C ∗ (T (t)) jint ; (12)
imental correlations such as the dimensionless relationship #s
developed by Levins and Glastonbury (1972):
where the two e9ectiveness factors are solutions of the fol-
Sh = 2:0 + 0:47 Sc 0:36
Re 0:62
[DS =DC ] 0:17
; (8) lowing equations:
 
kint ∗ jint −1
where Sh is the Sherwood number (Sh = kd Leq =D), Sc is [C(t) − C (T (t))] $ + $1=jint − 1 = 0;
the Schmidt number (Sc = =D), Re is a solution/particle kd
4=3 1=3 

Reynolds number (Re = (Leq )=) and Leq is the equiva- kint
∗ 
jint −1 1=j 
lent diameter of the sphere having the same crystal area [C (t) − C (T (t))] $
+ $
int − 1 = 0: (13)
kd
(2) The incorporation of solid is represented with the as-
sumption that the solution is saturated at the crystal surface. If the kinetic parameters are known, it follows that the e9ec-
The driving force of such incorporation is now given by tiveness factors are given by Eq. (13) for which numerical
the di9erence [Ccs (t) − C ∗ (T (t))]; [Ccs

(t) − C ∗ (T (t))], solutions may be obtained regardless of the value of the ex-


respectively. ponents jint and jint . Moreover, if the exponents are equal
As mentioned above, various growth mechanisms are to = 1 or 2, analytical solutions are available.
available in the literature according to di9erent structures It is worth noting that the ratio G=G
may depend:
of crystal–solution interface. However, for practical pur-

• on the level of supersaturation (if jint = jint ),
poses, empirical formulations such as the following are more


• both sizes L and L (through kd , $ and $ ),
tractable:
• any parameter e;cient in modifying selectively the indi-


s (t) = kint (Ccs (t) − C ∗ (T (t)) jint and vidual kinetic constants kint and kint (e.g. the concentra-

tion of impurities in the solution).
s
(t) = kint



(Ccs (t) − C ∗ (T (t)) jint ; (9)
where s ; s
are the Aux density of the integration of solid 2.3. Mass balance equations


on both crystal faces, kint ; kint are the corresponding kinetic


constant and jint and jint are the order of integration, gener- As an example, Eq. (14) describes the overall mass bal-
ally found between 1 and 2, depending on the mechanism ance during an isothermal semi-continuous crystallization
in question. operation fed with clear solution:
The linear growth rates are directly related to the Aux d[V (t)C(t) + VT (t)CS (t)]
fin (t)Cin (t) = ; (14)
density of integration by dt
#s #s
where the total volume of solution can easily be computed if
s = G and s
= G: (10)
2Ms 2Ms one neglects the variation of volume related to the generation
As for chemically reacting systems, it can be seen from of solid in the well-mixed crystallizer:

t
the previous equations that the rate of particle growth is a
function of both di9usion and solid incorporation rates. In V (t) = V (0) + fin (t) dt: (15)
0
the context of mono-dimensional modelling, an e9ectiveness
For a given time the solid concentration CS can be computed
factor was introduced by Garside (1971) to measure the real
from the bi-dimensional CSD, as follows:
mass Aux density real , with respect to the maximum mass


Aux density max that would be obtained in the absence of #s 2
CS (t) = LL
· (L; L
; t) dL dL
(16)
di9usive limitations. For the bi-dimensional case, we de@ne Ms L L 
two e9ectiveness factors, $ and $
, respectively : and the total volume of suspension is given by
real s  
$= = and Ms
max kint (C − C ∗ ) jint V (t) = VT (t) 1 − CS (t): (17)
#s


real s
Finally, the solute concentration C(t) is computed through
$
=

=  ;

(C
− C ∗ ) jint
(11)
max kint the integration of the balance equations (15) to (17).
3720 F. Puel et al. / Chemical Engineering Science 58 (2003) 3715 – 3727

3. Solving the bi-dimensional population balances lem of aerosol aggregation—where no particle growth is
involved—considering both particle volume and surface as
Population balance models are the natural framework independent of particle coordinates. Heinrich et al. (2002),
for systems where a discrete population is distributed over above mentioned, used a formulation similar to the method
one or more variables including CSD. The resulting equa- of classes. Ma et al. (2002) have presented a high resolu-
tions are partial integro-di9erential equations and allow tion @nite di9erence algorithm taken from the @eld of com-
the calculation of population number density functions. putational physics, which was applied to a model inorganic
Mass was suggested as the distribution parameter in order crystallization system.
to make the consistency of the mass balance easier (see
Ramkrishna’book (2000); Scarlett, 2002). The integral of 3.2. Solving bi-dimensional PBEs using the method of
such populations then represents the total mass of particles classes
in some chosen volume or stream, whereas the total num-
ber of particles is constantly changing. In the present study 3.2.1. De;nitions
a classical formulation based on crystal number density The spatial domain of crystals length and width (respec-
functions was used. tively L and L
) is @rst discretized and the smallest class of
size is assumed to @t the characteristic nuclei represented by
∗
its two dimensions, L∗ and L
.
3.1. Numerical methods for solving bi-dimensional PBEs: Let L0 ; : : : ; Li ; : : : ; Lim be a suite of length. L0 = L∗ is the
selection smallest length of crystals, i.e the length of the nuclei. Lim
is the length of the largest crystals. These lengths de@ne
The existing solving techniques, which were presented in im classes quoted Cli , the extent of the class is TCli =
the previous paragraph, focus primarily on models in a sin- Li − Li−1 and the characteristic length of the class Cli is
gle distributed dimension. As far as advanced applications Si = (Li−1 + Li )=2.
of these techniques should deal with multidimensional PBEs ∗
Let L
0 ; : : : ; L
j ; : : : ; L2; jm be a suite of width. L0 = L
is
describing the time evolutions in size and habit of a popu- the smallest width of crystals, i.e. the width of the nuclei.
lation of particles (hydroquinone in the present study), the L
jm is the width of the largest crystals. These widths de@ne
question arises about the selection of the more appropriate
jm classes quoted Cl
j , the extent of the class is TCl
j =
solving numerical technique.
L
j − L
j−1 and the characteristic length of the class Cl
j is
Ramkrishna and Mahoney (2002) indicate that Monte
Sj
= (L
j−1 + L
j )=2.
Carlo methods have traditionally been among the @rst
Such discretization de@nes bi-dimensional classes Cli; j ,
applied techniques, due to the ease of formulation of
delimited by Li ; Li−1 and L
j ; L
j−1 of area Ai; j = Cli Cl
j .
multidimensional problems. However, the required com-
These classes are @xed and of constant size. The size do-
putational speed and accuracy represent an important
main is divided in a system of im by jm bi-dimensional
drawback.
classes. The method of classes also requires the introduc-
Even if @nite elements methods are among the most pow-
tion of the population number function N (L; L
; t) de@ned as
erful solving techniques, the number of states required to
follows:
represent the solution increases exponentially with the num-
ber of dimensions. Discretization sizing techniques appear d 2 N (L; L
; t)
to be robust and seem rather easy to extend in a second = (L; L
; t); (18)
dL dL

dimension. This is why such method was selected in the
present work. However two main drawbacks of the method where N (L; L
; t) is the number of crystals with length less
should be outlined. Firstly, the feasibility of the computation than or equal to L and width less than or equal to L
. The
is likely to strongly depend on the complexity and location number of crystals belonging to the class Cli; j can therefore
of individual birth and death processes. For the crystalliza- be expressed as:
tion of hydroquinone, only birth through nucleation at low
Li
Lj
sizes has to be considered. Secondly, the number of dis- Ni; j (t) = (L; L
; t) dL dL
; (19)
cretized equations increases as the number of classes to the Li−1 Lj−1
square, leading to high computational times. Implementing
an adaptive bi-dimensional algorithm may probably improve im by jm number functions Ni; j (t) are thus de@ned.
the situation.
Finally, it should be noted that the choice of the numer- 3.2.2. Application of the method of classes
ical technique was, to a certain extent, con@rmed by three To express the population balance equations, the pop-
di9erent papers dealing with bi-dimensional balance equa- ulation number function N (L; L
; t) are now used, rather
tions where discretization methods were used, despite dif- than the population density function (L; L
; t). Through a
ferences in the formalism and the application area. Xiong two steps process, the balance equation (1) is discretized
and Pratsinis (1993) have applied the method to the prob- according to the bi-dimensional grid presented above (see
 F. Puel et al. / Chemical Engineering Science 58 (2003) 3715 – 3727 3721

Growth along L' direction

i with fi;Oj (t) and fi;Oj (t) denoting outlet crystal Aows from
L'j+1 the Cli; j class in the L and L
directions, respectively, and
L'j-2 L'j-1 L'j I
fi;I j (t) and f
i; j (t), inlet crystal Aows from the Cli; j class in
Li-2 the L and L
directions, respectively.
Ni-1,j
The calculation of these inlet and outlet crystal Auxes are
Growth
w
Li-1
f I
i,j carried out using a @rst-order Tailor series expansion:
along L
direction f 'Ii,j
Ni,j
f 'Oi,j
fi;Oj (t) = G(Si ; t)[ai Ni; j (t) + bi Ni+1; j (t)]; (24)
Ni,j-1 Ni,j+1
Li
f O
i,j fi;I j (t) = G(Si−1 ; t)[ai−1 Ni−1; j (t) + bi−1 Ni; j (t)] (25)
Ni+1,j
Li+1 with
TCli+1
ai = ; (26)
Fig. 2. Bi-dimensional class Cli; j , inlet Aows of crystal numbers fi;I j ; TCli (TCli+1 + TCli )
f Ii; j and outlet Aows of crystal numbers fi;Oj ; f O
i; j; (t). TCli
bi = (27)
TCli+1 (TCli+1 + TCli )
also Fig. 2): and
(1) The PBE is integrated after combining expressions O
(19) and (2) f
i; j (t) = G
(Sj
; t)[cj Ni; j (t) + dj Ni; j+1 (t)]; (28)
1 d
[N (L; L
; t)VT (t)] I
f
i; j (t) = G
(Sj−1


; t)[cj−1 Ni; j−1 (t) + dj−1 Ni; j (t)] (29)
VT (t) dt

Lmax
Lmax with
@
+ [G(L; t) (L; L
; t)] TClj+1
L∗ L∗ @L cj = ; (30)

Lmax
Lmax TClj (TClj+1 + TClj )
@ TClj
+

[G
(L
; t) (L; L
; t)] dj = : (31)
L ∗ L ∗ @L TClj+1 (TClj+1 + TClj )
F(t)N (L; L
; t) − Fin (t)Nin (L; L
; t) The previous expressions are valid for 2 ¡ i ¡ im − 1 and
+
VT (t) 2 ¡ j ¡ jm − 1. Only four neighbouring classes of Cli; j are

Lmax
Lmax considered in the design of the algorithm, particular cases
= RN dL dL
(20) are considered for the classes which are set at the boundaries
L∗ L∗ of the size domain:
with

Lmax
Lmax • The lower classes cannot accept growing crystals form
RN dL dL
= rN 1 + rN 2 : (21) the previous ones:
L∗ L∗
(2) The population balance around the bi-dimensional f1;I 1; j (t) = f2;I i; 1 (t) = 0: (32)
class Cli; j leads to a set of im by jm ordinary di9erential • Crystals in the upper classes cannot grow in the next ones
equations:
1 d f1;I im; j (t) = f2;I i; jm (t) = 0: (33)
[Ni; j (t)VT (t)] + fi; j (t) + fi;
j (t)
VT (t) dt
3.2.3. Numerical procedure
F(t)N (L; L
; t) − Fin (t)Nin (L; L
; t) At each time step, the discretized population balance
+ = r N 1 + rN 2 :
VT (t) equations are solved simultaneously with the mass balance
(22) in order to estimate the supersaturation in the crystallizer.
fi; j (t) and fi;
j (t) are the net inlet and outlet Aows of crystals Using a set of kinetics parameters, one can then compute
in class Cli; j , in the length and width directions induced nucleation and growth kinetic rates for each crystal size
by growth, respectively. Each crystals Aow being divided (see kinetic expressions in Section 2) and then integrate the
in inlet and outlet Aows for each direction (see arrows ODEs system. Originally, these integrations were carried
on Fig. 2) out using a fourth-order Runge–Kutta routine. In order to
speed up the computation, the “LSODAR” software was
fi; j; (t) = fi;Oj ; (t) − fi;I j ; (t) and adopted, which selects Adams–Moulton predictor–corrector
O I
or a Gear method with time step adjustment. Fig. 3 sum-
fi;
j ; (t) = f
i; j ; (t) − f
i; j ; (t) (23) marizes the structure of the main computer program.
3722 F. Puel et al. / Chemical Engineering Science 58 (2003) 3715 – 3727

sizes of generated nuclei and - = VT =F:

  
∗
rN 2 L − L∗ L − L

= exp − exp − : (35)
4GG
- 2G- 2G
-
Introducing a new coordinate X allows formulating expres-
sion (35) as follows:
  
G
X = (L − L∗ ) cos Arctan
G

  
∗ G
+ (L
− L
) sin Arctan ; (36)
G

 
rN 2 cos(Arctan(G=G
))
= exp − X (37)
4-GG
-G
and
 r   cos(Arctan(G=G
)) 
N2
ln = ln − X
4-GG
-G
Fig. 3. Structure of the main simulation program.
 r 
N2
= ln − /X; (38)
4-GG

Preliminary tests were carried out, in order to track poten- where /, the slope in Eq. (38), is a constant for given growth
tial numerical problems described in the literature. The main rates and residence time.
di;culty arises from the possible loss of crystals, leading Expressions (37) and (38) provide a means of evaluating
to an unbalanced mass balance during the crystal growth. the convergence of the simulation algorithm through the as-
Without signi@cant bene@t, a third-order Tailor series ex- sessment of the @nal steady-state CSD computed after tran-
pansion was used for the calculation of the Aows of growing sient conditions (i.e. when time exceeds about six average
crystals, using eight neighbouring classes for Cli; j . In order residence times, -). In order the continuous crystallizer to be
to cope with the di;culty, a simple and e;cient way was stabilized, the competition between the consumption of so-
to reduce the area of Cli; j . For the system under investiga- lute through growth mechanisms and the generation of new
tion, the maximal area was set to 1 m × 0:5 m which, crystal surfaces through nucleation should keep supersatura-
for a size domain of 1000 m × 300 m, requires 6 × 105 tion from increasing excessively, and should therefore avoid
classes. The crystallization process was then simulated with- cyclic bursts of nucleation. In order to simulate such condi-
out signi@cant inconsistencies in the population balances tions a speci@c set of kinetic data was used in the present
involved. test (see Table 1).
Fig. 4 represents the results obtained using the method of
classes, in steady-state conditions, using the kinetic data set
4. Basic simulation tests given in Table 1. As expected, the plot = f(L; L
) shows
the exponential shape of the bi-dimensional CSD, and in
4.1. Steady-state CSD in ideal MSMPR crystallizers semi-logarithmic coordinates, the same plot gives a straight
line which can be drawn with respect to the X coordinate
The mass and population balance equations describing to provide the slope / introduced in Eq. (38). Using /, it
the behaviour of ideal “mixed suspension mixed product re- is a straightforward exercise to re-compute the steady-state
moval” (MSMPR) crystallizers have widely been published values of G; G
and rN2 .
(see e.g. Mullin, 2001). At the steady state, one can easily The “input” kinetic data for the simulation of the continu-



write ous crystallization are the values of kint ; kint and jint ; jint (see
Eq. (9)) displayed in Table 1. The simulation was carried
@ out assuming negligible di9usive limitations. Steady-state
[G · (L; L
)]
@L values for G and G
are thus obtained and Eq. (5) is used to
@ (L; L
) compute rN2 . The simulation leads to the plots in Fig. 4, and
+

[G
· (L; L
)] + = RN : (34) new values of G; G
and rN2 are computed as follows: the
@L -
projection of the straight line given by Eq. (40) on both axes
Indeed, at the steady state, only secondary nucleation L and L
allows to estimate G and G
, respectively, while
mechanisms have to be considered. the intercept gives rN2 . As one can see on Table 1, the com-
Assuming that the growth rates are size-independent, one parison between simulated “input” kinetic parameters and
can demonstrate that the following expression is the analyt- the values given by the @nal CSD is rather satisfactory as,
∗
ical solution of Eq. (34), where L∗ ; L
are the characteristic even though the smoothness of the plot is not perfect, the
 F. Puel et al. / Chemical Engineering Science 58 (2003) 3715 – 3727 3723

Table 1
Kinetic equations and parameters selected for the simulation of steady-state continuous well mixed crystallizers

G (m s−1 ) G  (m s−1 ) rN 2 ([nb] m−3 s−1 )

Secondary nucleation and growth rates 5:02 × 10−9 1:25 × 10−9 3:247 × 109
computed using the kinetic data set

Secondary nucleation and growth rates 5:03 × 10−9 1:256 × 10−9 3:241 × 109
computed from the CSD in Fig. 2
(10,000 size classes were used)

Relative error (%) with 10,000 classes 0.2 0.5 0.3

Secondary nucleation and growth rates 5:46 × 10−9 1:41 × 10−9 3:276 × 109
computed when 168 size classes are used

Relative error (%) with 168 classes 8.8 12.3 1.1

Nucleation kinetic law: rN2 = kN · 2 MT


 ·  jint
Growth laws: G = kint ·  jint and G2 = kint
−1
Input kinetic data: kN = 2 × 1015 [nb] mol s−1 ; kint = 7 × 10−5 m s−1 ; kint
 = 1:75 × 10−5 m s−1 j = j  = 1
int int

Fig. 4. Simulation of a continuous crystallization of hydroquinone: steady-state bi-dimensional population density functions.

relative error on the @nal kinetic parameters never exceeds between nucleation and growth processes (Nyvlt and
0.5% when 10,000 size classes were used to simulate the Mullin, 1970; Randolph, Beckmen, and Kraljevich, 1977;
CSD. Table 1 also shows that the relative uncertainty be- Garside, 1985). At an early stage, the criterion of insta-
comes signi@cant (i.e. up to 12.8%) if only 168 size classes bility has been expressed in terms of critical values of the
are set to solve the PBEs. ratio i∗ = d(ln rN )=d(ln G) (see e.g. Randolph and Larson,
1965; Lei, Shinnar, & Katz, 1971). However, further ex-
perimental and theoretical results showed that the reported
4.2. Oscillatory CSD in ideal MSMPR crystallizers theoretical critical values of i∗ were unlikely. Finally, it
turned out that instability can be explained as a result of
In order to validate the simulation program, it was also discontinuities in the nucleation mechanisms which may
interesting to reproduce oscillatory behaviours in MSMPR occur when various metastable limits are crossed dur-
crystallizers. Such unsteady-state behaviours of industrial ing the time variations of supersaturation (Garside, 1985;
continuous crystallizers have been widely investigated in Puel et al., 1997).
the past. “Low and high Order Cycling” was described and In industrial continuous perfectly mixed crystallizers, the
theoretically explained by complex competition phenomena occurrence of oscillatory behaviours can be observed with
3724 F. Puel et al. / Chemical Engineering Science 58 (2003) 3715 – 3727

Table 2
Kinetic equations and parameters selected for the simulation of oscillatory continuous well mixed crystallizers

A ([nb] m−3 s−1 ) B kN ([nb] mol−1 s−1 ) kint (m s−1 )  (m s−1 )

kint

Nucleation and growth 1020 0.4 2 × 104 7 × 10−7 1:75 × 10−7

kinetic parameters
 
−B
Nucleation kinetic laws: rN1 = A exp rN 2 = k N ·  2 M T
ln2 

Growth laws: G = kint ·  jint and Gint
 = ki; 2 ·  jint
 =1
Input kinetic data: jint = jint

oscillatory dynamics during the crystallization process is

encouraging.

4.3. Transient propagation of the CSD during batch

desupersaturation operation

In order to check the ability of the program to reproduce

“ideal” propagation of the CSD in well-mixed crystallizers,
it was assumed that size-independent growth rate takes place
after the introduction of seed crystals in the supersaturated
region. As displayed in Fig. 6, the CSD of the seed was
considered as Gaussian. Secondary nucleation was assumed
to be negligible, it was therefore expected that the Gaus-
sian CSD should reproduce the crystal growth in both direc-
tions (i.e. length and width) without any deformation (i.e. no
increase in the width of the distribution). The program sim-
ulates the decrease in supersaturation according to the over-
all mass balance of the crystallization operation and solves
the PBEs until the solubility curve is reached. As one can
see in Fig. 6, the shape of the CSD is satisfactorily repro-
duced by the simulation program when 10,000 size classes
are used as no signi@cant uncertainty was observed in the
@nal CSD.

Fig. 5. Simulation of oscillatory behaviour of continuous well mixed

5. Conclusions and perspectives
crystallization of hydroquinone (low secondary nucleation rate).
The method of classes was presented and developed to
simulate the time variations of multidimensional crystal size
hydroquinone, especially if additives reducing the growth distributions. Using the method of classes, the accurate sim-
rate G are used. Even though the kinetic parameters in- ulation of the very beginning of the generation of solid
volved are not well-known, instability can be simulated if (i.e. at very short size and time scales) is bypassed as nuclei
low secondary nucleation rates are speci@ed through the are simply considered as a number of crystals belonging to
kinetic data set. Fig. 5 shows the simulation results ob- the smaller size class of the size grid. The main drawback of
tained using the data in Table 2. As in the previous case, the technique is the requirement for small spacing in each
di9usive limitations were assumed to be negligible. The size direction, leading to a very large number of classes,
results are not expected to accurately reproduce the dy- which may result in high computational time.
namics of industrial crystallizers since, as demonstrated An e9ort was made in the present work to improve pre-
by Puel et al. (1997), habit transients may be observed viously reported simulation studies through the introduction
during the crystallization process, which make both nu- of detailed physical knowledge in the kinetic laws involved
cleation and growth kinetics time- and size-dependent. during nucleation and growth phenomena. For example, dif-
Nevertheless, as far as the performances of the algorithm fusive limitations, which are likely to occur in the moment
are concerned, the successful representation of complex following primary homogeneous nucleation (i.e., at very
 F. Puel et al. / Chemical Engineering Science 58 (2003) 3715 – 3727 3725

Fig. 6. Propagation of an initial Gaussian CSD during the simulation of a desupersaturation experiment (a) initial; (b) @nal. The growth rates were
computed according to Eq. (12).

high levels of supersaturation), are taken into account in the CS concentration of solid in the suspension,
modelling approach. mol m−3
The performances of the simulation algorithm were D di9usivity of the solute in solution, m2 s−1
successfully assessed through the reproduction of two DC internal diameter of the crystallizer, m
well-known theoretical and experimental features of ideal DS stirrer diameter, m
continuous crystallization processes: the computation fin Aow rate of inlet clear solution, m3 s−1
O
of size-independent growth rates from the plot of the fi;Oj (t); f
i; j (t) outlet crystal number Aow in the L and L

steady-state CSD and the possibility for MSMPR crystal- directions, [nb] m3 s−1
I
lizers to exhibit low-frequency oscillatory behaviours, in fi;I j (t); f
i; j (t) inlet crystal number Aow in the L and L

the case of insu;cient secondary nucleation. directions, [nb] m3 s−1
New experimental results related to the crystallization of F Aow rate of suspension, m3 s−1
hydroquinone in water will be more extensively described Fin Aow rate of inlet suspension, m3 s−1
in the next issue of this series of papers. More attention will G; G
growth rate of crystal face of type 1 and
be focused on both nucleation and growth mechanisms. The 2, m s−1
relevancy of both primary and secondary kinetic laws used

jint ; jint index related to the law of solid integration
in this paper will be assessed and the reality of di9usive on faces 1 and 2
limitations during the growth of organic industrial products k kinetic orders for secondary nucleation
such as hydroquinone will be demonstrated in the next issue kd mass transfer coe;cient, m s−1
of this series of papers.

kint ; kint kinetic parameter of the integration law
for faces 1 and 2, mol−j m3j+1 s−1
kN secondary nucleation rate coe;cient, [nb]
Notation mol−j m3−3j s−1
Leq: average equivalent size of crystal, m
A; B empirical coe;cients in the primary nu- L; L
length and width of crystal, m
cleation law, A, (nb) s−1 m−3 L∗ ; L

∗
length and width of nuclei, m
Ai; j area of class Cli; j , m2

Li ; Lj discretized length and discretized width of
ai ; bi tailor development coe;cient, m−1 crystal, m
C solute concentration, mol m−3 Ms molecular weight of solid, kg mol−1
C∗ solubility, mol m−3 MT mass of crystallized solid per unit volume
C
solute concentration at the solid interface, of suspension, kg m−3
mol m−3 n kinetic orders for secondary nucleation
cj ; dj tailor development coe;cient, m−1 Re solution/particle Reynolds number
Cli ; Cl
j Mono-dimensional class on length and RA distribution of agglomeration rates, [nb]
width directions m−2 s−1 m−3
Cli; j bi-dimensional class RB distribution of breakage rates, [nb] m−2


Ccs ; Ccs solute concentration at the solid–solution s−1 m−3
interface for the RN distribution of nucleation rates, [nb] m−2
face of type 1 and 2, mol m−3 s−1 m−3
3726 F. Puel et al. / Chemical Engineering Science 58 (2003) 3715 – 3727

rN 1 primary nucleation rate, [nb] s−1 m−3 Hulburt, H. M., & Katz, S. (1964). Some problems in particle technology.
rN 2 secondary nucleation rate, [nb] s−1 m−3 A statistical mechanical formulation. Chemical Engineering Science,
Sc Schmidt number, dimensionless 19, 555–574.
Iveson, S. M. (2002). Limitations of one-dimensional population balance
Sh sherwood number, dimensionless
models of wet granulation processes. Powder Technology, 124,
Si mean size of class Cli , m 219–229.
Sj
mean size of class Cl
j , m Kind, M., & Mersmann, A. (1990). The induction time of seeded
t time, s solutions—on its application for the design of crystallization processes.
T crystallizer temperature, ◦ C In A. Mersmann (Eds.), Proceedings of the 11th symposium on
industrial crystallization. Garmish Parten-Kirshen, West Germany (pp.
V volume of solution in the crystallizer, m−3 393–398).
VT total volume of suspension, m−3 Lei, S. J., Shinnar, R., & Katz, S. (1971). The regulation of continuous
crystallizer with @nes trap. A.I.Ch.E. Symposium Series, 110(67),
Greek letters 129–144.
Levins, D. M., & Glastonbury, J. R. (1972). Particle-liquid hydrodynamics
and mass-transfer in a stirred vessel. Part II: Mass transfer. Transaction
 supersaturation degree C=C ∗ of the Institution of Chemical Engineers, 50, 132–146.
TCli ; TCl
j extent of classes Cli and Cl
j , m Ma, D. L., Tafti, D. K., & Braatz, R. D. (2002). Optimal control and
stirring power dispersed per unit mass of simulation of multidimensional crystallization processes. Computers
slurry, W kg−1 and Chemical Engineering, 26(7–8), 1103–1116.
Mahoney, A. W., & Ramkrishna, D. (2002). E;cient solution of
$; $
e9ectiveness factor for the faces 1 and 2 population balance equations with discontinuities by @nite elements.
 kinematic viscosity of the solution, Chemical Engineering Science, 57, 1107–1119.
m2 s−1 Marchal, P. (1989). Genie de la cristallisation: Application a l’acide
#s density of the solid, kg m−3 adipique. Thesis, Institut Polytechnique de Lorraine.
Marchal, P., David, R., Klein, J.-P., & Villermaux, J. (1988).
 relative supersaturation (C − C ∗ )=C ∗ Crystallization and precipitation engineering—I. An e;cient method
s ; s
Aux density of solid integration on faces for solving population balance in crystallization with agglomeration.
1 and 2, mol s−1 m2 Chemical Engineering Science, 43(1), 59–67.

Mersmann, A. (1996). Supersaturation and nucleation. Transactions of
real ; real real mass Aux density on faces 1 and 2,
the Institution of Chemical Engineers, 74(Part A), 812–820.
mol s−1 m2
Mersmann, A. (2001). Crystallization technology handbook (2nd ed.).
t ; t
mass Flux density transferred through the New York: Marcel Dekker Inc.
@lm (faces 1and 2), mol s−1 m2 Muhr, H., David, R., Villermaux, J., & Jezequel, P. H. (1996).
number population density function in the Crystallization and precipitation engineering—VI. Solving population
suspension, [nb] m−1 m−3 balance in the case of the precipitation of silver bromide crystals
with high primary nucleation rates by using the @rst order upwind
di9erentiation. Chemical Engineering Science, 51(2), 309–319.
Mullin, J. W. (2001). Crystallization in the process industries (4th ed.).
London: Butterworth.
References Nicmanis, M., & Hounslow, M. J. (1998). Finite-elements methods for
steady state population balance equations. A.I.Ch.E. Journal, 44(10),
Burton, W. K., Cabrera, N., & Frank, N. C. (1951). The growth 2258–2272.
of the crystals and the equilibrium structure of their surfaces. Nyvlt, J., & Mullin, J. W. (1970). The periodic behaviour of continuous
Philosophical Transactions of the Royal Society of London, A243, crystallizers. Chemical Engineering Science, 25, 131–147.
299–358. Pan, D., & Mersmann, A. (1992). On an integro-PDE arising from
David, R., Marchal, P., & Marcant, B. (1995). Modelling of agglomeration population balance of crystal growth—Analytical solutions and
in industrial crystallization from solution. Chemical Engineering and simpli@cations. Applied Mathematics Modelling, 16, 214–218.
Technology, 18(5), 302–309. Patil, D. P., & Andrews, J. R. G. (1998). An analytical solution to
Franck, R., David, R., Villermaux, J., & Klein, J.-P. (1988). Crystallization continuous population balance model describing Aoc coalescence and
and precipitation engineering—II. A chemical reaction engineering breakage—A special case. Chemical Engineering Science, 53(3),
approach to salicylic acid precipitation: Modelling of batch kinetics and 599–601.
Puel, F. (1994). Bilan de population pour deux tailles caracteristiques de
application to continuous operation. Chemical Engineering Science,
particules: Application a la crystallisation de l’hydroquinone. Thesis,
43(1), 69–77.
Universit,e Claude Bernard Lyon 1 no. 162-94.
Garside, J. (1971). The concept of e9ectiveness factors in crystal growth.
Puel, F., Marchal, P., & Klein, J.-P. (1997). Habit transient analysis in
Chemical Engineering Science, 26, 1425–1431.
industrial crystallization using bi-dimensional crystal sizing technique.
Garside, J. (1985). Industrial crystallization from solution. Chemical Transactions of the Institution of Chemical Engineers, 75(Part A),
Engineering Science, 40(1), 3–26. 193–205.
Heinrich, S., Peglow, M., & MWorl, L. (2002). Unsteady and Ramkrishna, D. (2000). Population balances—theory and applications
steady state particle size distributions in batch and continuous to particulate systems in engineering. San Diego: Academic Press.
Auidized bed granulation systems. Chemical Engineering Journal, 86, Ramkrishna, D., & Mahoney, A. W. (2002). Population balances
223–231. modelling. Promise for the future. Chemical Engineering Science, 57,
Hillig, W. B. (1966). A derivation of classical bi-dimensional nucleation 595–606.
kinetics and the associated crystal growth laws. Acta Metallurgica, Randolph, A. D., Beckman, J. B., & Kraljevich, Z. I. (1977). Crystal size
14, 1868. distribution dynamics in a classi@ed crystallizer: Part I. Experimental
Hounslow, M. J., & Wynn, J. W. (1993). Short-cut models for particulate and theoretical study of cycling in a potassium chloride. Crystallizer.
processes. Computers Chemical Engineering, 17(5/6), 505–516. A.I.Ch.E. Journal, 23(4), 500–510.
 F. Puel et al. / Chemical Engineering Science 58 (2003) 3715 – 3727 3727

Randolph, A. D., & Larson, M. A. (1965). Transient and steady state size Wojcik, J. A., & Jones, A. G. (1998). Dynamics and stability of continuous
distribution in continuous mixed suspension. Crystallizers. A.I.Ch.E. MSMPR agglomerative precipitation: Numerical analysis of the
Journal, 19, 1140. dual co-ordinate model. Computers Chemical Engineering, 22(4 –5),
Randolph, A.D., & Larson, M.A. (1988). Theory of particulate processes: 535–545.
Analysis and techniques of continuous crystallization. Second Edition, Xiong, Y., & Pratsinis, S. E. (1993). Formation of agglomerate particles
Academic Press, London (ISBN = 01257966528). by coagulation and sintering—Part I. A bi-dimensional solution of
Rawlings, J. B., Witkowski, W. R., & Eaton, J. W. (1992). Modelling the population balance equation. Journal of Aerosol Science, 24(3),
and control of crystallizers. Powder Technology, 69, 3–9. 283–300.
Scarlett, B. (2002). Particle Population—to balance or not to balance,
that is the question. Powder Technology, 125, 1–4.