Abstract
E!orts to model precipitation processes in chemical reactors need to combine #uid mechanics with precipitation processes
description. The main di$culty is to consider together micromixing phenomena, which are known to play an important role, with
complex reactions kinetics and particles size distribution in the #ow. In the present study, a probability density function (PDF)
method is presented to calculate the evolution of the particles size distribution. The main advantage of the method is to include
a micromixing model based on a Lagrangian frame where chemical reactions are treated without modelling and which requires
minimal computational resources. The micromixing model considered is the interaction and exchange with the mean (IEM).
Precipitation reactions are nucleation, growth and aggregation. Because of the nature of the powerful numerical technique used, based
on Monte-Carlo simulations, the method is able to produce, at any point in the #ow, reactants concentration and supersaturation
"elds, and the particles size distribution "eld, by simple moments method. The main advantage of the method over traditional
population balance solvers is its capability to treat multi-dimensional (e.g. size, crystal morphology, chemical properties, etc.)
population balances just as e$ciently as it treats high-dimensional PDF. The PDF code is coupled with the commercial CFD
package FLUENT which is used for the turbulent calculation of the #ow (classical k} turbulence model), where the presence of the
solid phase is assumed not to a!ect the #ow "eld. An example is presented in the case of silica particles aggregation. The in#uence of
initial reactants concentration and hydrodynamics parameter are investigated. 2001 Elsevier Science Ltd. All rights reserved.
0009-2509/01/$ - see front matter 2001 Elsevier Science Ltd. All rights reserved.
PII: S 0 0 0 9 - 2 5 0 9 ( 0 0 ) 0 0 4 4 7 - 4
2446 L. Falk, E. Schaer / Chemical Engineering Science 56 (2001) 2445}2457
such as feed addition mode, feed rate, feed location, feed species at point X and time t, between C and C#dC.
concentration, stirring speed, and stirrer type have been The transport equation of f can be derived from the
A
shown to a!ect the resulting average particle size conservation equations for passive scalars and can be
(Mersmann & Kind, 1988; Aslund & Rasmuson, 1992; written as (Pope, 1985)
Franke & Mersmann, 1995). However, majority of these
previous works have considered simple phenomenologi- ! 0
cal models of the #ow, based on residence time distribu- CDE CFDFE
tion (RTD). They describe the #ow pattern by perfect f f
A #; A # (S (C)f )
reactors associations, as plug #ow or perfect stirred tank t G x C H A
reactors, coupled with micromixing models. This ap- G H
proach is simple to develop and may give tendency re- 2 +
sults but does not, in general, re#ect details of the real CFFDFFE CFFFDFFFE
#ow patterns in a complex 3D apparatus. With the devel- 1
"! (u Cf ) ! D C C f . (1)
opment of computational #uid dynamics (CFD) codes, it x G A C A H A
G H
is now possible to calculate satisfactorily the #ow struc-
ture and mixing in complex precipitators. However, the The left-hand side terms of this expression include the
description of turbulent reacting #ows is still di$cult accumulation of probability, the probability transport in
because of the randomness of the #ow "eld and the real space by the mean velocity "eld ; and the prob-
complex chemical processes involved. Studies have been G
ability transport in composition space of species j by
carried out in this "eld (Van Leeuwen, Bruinsma, & Van chemical reactions S . These terms are calculated exactly
Rosmalen, 1996; Wei & Garside, 1997) but perfect mi- H
without any assumptions. On the right-hand side are the
cromixing was assumed to exist in each computational unclosed terms which include probability transport in
cell. As nucleation has been shown to be very sensitive to real space by the velocity #uctuations u and the prob-
micromixing, such a strong assumption may lead to G
ability transport in composition space by molecular dif-
considerable errors in particle size distribution predic- fusion (micromixing). These terms must be modelled.
tion. Therefore, for such mixing-sensitive reactions, all
scales must be considered.
Traditionally, reacting #ows equations are solved us- 2.1. Turbulent diwusivity model
ing the Reynolds decomposition, introducing the concen-
tration moment which deals with averaged quantities of Classically, the Boussinesq assumption, i.e. the gradi-
the #ow and concentration "elds. The mean scalar "eld ent di!usion model, is used to close the turbulent trans-
equation contains unclosed terms as the mean reaction port term:
rate which is the most di$cult term to close and, unfortu-
f
nately, it is of great importance in reactive #ow model- u Cf "!D A , (2)
ling. When several reactants are involved with complex G A R x
G
"nite rate chemistry as in precipitation processes, the
where D is the turbulent di!usion coe$cient. Typically,
di$culties associated with moment closures are insur- R
D can be estimated by using the turbulence k} model
mountable and other methods must be employed. R
D "(C /0.7)k/ where C "0.09, k is the turbulent
Full probability density functions (PDF) methods are R I I
tractable methods for industrial complex reactive #ows. kinetic energy and the turbulent energy dissipation.
They combine an exact treatment of chemical reactions
with a closure of turbulence "eld and are an e$cient tool 2.2. Micromixing model
for complex reaction processes. The aim of this paper is
to present PDF methods for precipitation processes In this study, we closed the micromixing phenomena
simulation, including nucleation, growth and agglomer- by the classical Lagrangian interaction by exchange with
ation kinetics together with micromixing phenomena. An the mean (IEM) model proposed by Villermaux and
application example will be presented in the case of Devillon (1972), also known in turbulent combustion as
agglomeration of silica particles. the LMSE model proposed by Dopazo (1975):
1 (C !C )
D C C f " H H f, (3)
2. PDF methods A H A t A
I
In the composition PDF approach, adopted here, the where C is the concentration of species j in a #uid
H
vector of species concentrations C is described by aggregate and where C " C f dC is the local
H H A H
a multi-dimensional probability density function f , so mean concentration of species j. The micromixing time
A
that f dC is the probability to have the concentration of t is estimated using the k} model by t "k/.
A I I
L. Falk, E. Schaer / Chemical Engineering Science 56 (2001) 2445}2457 2447
1 T
tallizing compound in a supersaturated solution and in B(v)" (v!u, u) n (v!u)n(u) du,
the crystal. The chemical potential of the dissolved com- 2
pound can be de"ned by way of activity while the super-
saturation ratio is de"ned as D(v)"n(v) (v, u)n(u) du. (9)
a CJ>CJ\
J J CJ>CJ\ J
S" ! " > \ !A " > \
. The aggregation kernel , expressed as a function of
a CJ> CJ\
J K !A
! > \ ! A particle volume, is a measure of the frequency of collision
(4)
between particles of volumes u and v!u that are suc-
Thus it is necessary to know the stoichiometric mean cessful in producing a particle of volume v.
activity coe$cient
in order to calculate the super-
!A
saturation ratio. In the case of relatively dilute solutions, 3.2. Moment transformation
it is approximated to one.
3.2.1. Kinetic expression
3.1.2. Nucleation The size distribution of the particles that are trans-
The nucleation rate of particles of volume <H can be ported by the #ow is characterised by the moment equa-
represented, according to the classical primary nuclea- tions. The moments of the distribution represent the
tion theory, by average or total properties of the solid phase and can be
used to calculate properties such as mean particle size or
A
B"B exp ! , (5) distribution variance. Following the treatment proposed
ln S by Hulburt and Katz (1964) (i.e. multiplying by vH and
where integrating each term with respect to v over the entire
4
N< particle volume range), the moment transformation of
A" ? K (6) the population balance yields for the moment's kinetic
27
(R¹)
T expression
and the critical volume <H of the nucleated particles is
S H "B<HH#jG
deduced from the Gibbs}Thomson relation I H\
1
4
< (v#u)H!vH (v, u)n(u)n(v) du dv. (10)
<H"
(¸H)"
? K . (7) #
T T 3
R¹ ln S
2
T
Moments equations for the "rst three moments are then:
3.1.3. Growth
The overall growth rate relating the particle volume 1
S "B! (v, u)n(u)n(v) du dv,
variation against time is frequently expressed by the I 2
relation
S "B<H#G ,
G"k (S!1)E. (8) I
E
The growth rate is here taken to be independent of the S "B<H#2G # uv(v, u)n(u)n(v) du dv. (11)
I
particle volume.
The "rst moment (which represents the total volume
3.1.4. Agglomeration of the particles per volume of solution) is conserved
Agglomeration is here taken as the uni"cation of par- during agglomeration. For analytically convenient ker-
ticles which are cemented later by chemical forces, for nels (i.e. kernels that yield analytical solutions of the
2448 L. Falk, E. Schaer / Chemical Engineering Science 56 (2001) 2445}2457
Table 1
Solutions of the agglomeration term for analytically convenient kernels
Form of kernel (u, v)" (u, v)" (uv) (u, v)" (u#v)
(constant kernel) (product kernel) (sum kernel)
Moment's order
1 1 !
0 ! !
2 2
1 0 0 0
2 2
agglomeration term, see Table 1) analytical solutions for model is used. The results are the mean velocity compo-
the "rst three moments can be found. nents "eld, the turbulent kinetic energy k and the turbu-
lent energy dissipation "elds.
3.2.2. Concentration variation In the second step, the PDF simulation is performed
The variation of the solution-phase precipitating com- by a home-made code. It is based on a set of notional
pound may be described by the variation of the solid particles moving in the computational domain, where the
mass produced: Monte-Carlo code mesh is the same as the CFD mesh.
Each of the notional particles follows a set of stochastic
dC dC d
> > "
\ \ "! N . (12) di!erential equations which mimic the PDF transport in
dt dt M dt physical and phase spaces, where each particle transports
N
species concentration and particles distribution mo-
Finally, the resolution of the coupled concentration and
ments.
moments transport equations, including micromixing, at
Two types of Monte-Carlo codes can be used, a Lag-
any point of the #ow, yields the reactants concentration,
rangian or an Eulerian frame. Our code has an Eulerian
supersaturation and distribution moments "elds.
form and in that case, the PDF transport equation is very
The average and total properties of the distribution
similar to the classical population balance of notional
can directly be obtained from these position-dependent
particles over each computational cell. We can consider
moments. For example, the average particle volume and
inter-cell or macromixing processes * such as spatial
distribution variance are here characterised by
transport by mean convection and by turbulent di!usion
exchange notional particles between cells * and intra-
< " , " !1 . (13)
T cell processes * such as chemical reaction and microm-
ixing. The intra-cell processes are commonly subdivided
The resolution has been here described for the number into two di!erent steps. For each notional particle, the
population density expressed as a function of particle chemical reaction step is "rst considered where the time
volume. As it will be seen in the next part, the method can evolution of the composition vector C (concentration
H
also be applied for the moments expressed as a function and particles distribution moments) is calculated by in-
of particle diameter. tegration of the balance equation over a small time inter-
val t:
dC
4. Stochastic solution method H "S . (14)
dt H
The composition PDF is a multidimensional distribu- The micromixing step is then considered where the time
tion of (n #dim) independent variables, with n being evolution of the composition vector C is calculated by
G G H
the number of composition variables (generally more integration of the IEM balance equation over a small
than 5, 2 for chemical species and 3 for the particle size time interval t:
distribution moments) and dim, the spatial dimension. As
dC (C !C )
the computational e!ort to solve the equation using H" H H , (15)
a "nite-di!erence grid grows exponentially with the num- dt t
I
ber of variables, more e$cient techniques have to be used where the micromixing time t is calculated using the
to solve the problem. For multi-species PDF a Monte- I
local values of k and . The mean value of the composi-
Carlo technique is used as the solving method (Pope, tion vector C is computed by averaging the composition
1985; Fox, 1996). The procedure is split in two steps: H
values over the N notional particles:
The turbulent #ow "eld, without any reaction, is "rst
solved using the CFD commercial code FLUENT, based 1 ,
C " C . (16)
on a "nite-volume method. The standard k} turbulent H N HI
I
L. Falk, E. Schaer / Chemical Engineering Science 56 (2001) 2445}2457 2449
Table 2
Flow con"guration parameters
Con"guration 1 Con"guration 2
Compared to classical methods which do not solve From dimensional analysis, it can be shown that the
microscales processes, PDF methods are considerably #ow in such a coaxial mixing jet can be described by
more computationally expensive. To keep an acceptable a dimensionless number, named the Craya-Curtet num-
statistical noise level, it is necessary to consider not less ber (Becker, Hottel, & Williams, 1963):
than 100 notional particles per computational cell. For
(u /u )((D/d)!1
coarse 2D meshes * 2000 cells, for instance * and C" . (17)
a composition vector of 5 species (2 chemical species and R (1!(u /u )
3 particles size distribution moments), 10 di!erential
A critical value of the number C "0.976 (Barchilon
equations have to be solved. For stationary reactive R
#ows, the PDF transport equation is integrated over time & Curtet, 1964) has been experimentally determined for
which a recirculation loop appears when C (C ,
until the calculated composition "elds reach stationary R R
values. Typically, in an open system, integration time is inducing a back-mixing e!ect in the reactor.
of the order of several residence times. On a Pentium In order to illustrate the in#uence of mixing on the
300 MHz PC, stable results are reached after several aggregation reaction, two di!erent #ow con"gurations
CPU hours of calculation. characterised by a di!erent value of the Craya-Curtet
number (cf. Table 2) are compared:
Con"guration 1: C "0.296,
5. Application R
Con"guration 2: C "1.414.
R
5.1. Fluid yow and passive tracer validations The reactor is designed for a speci"c production of ag-
gregated silica particles. As our objective is to reach the
5.1.1. Flow conxguration same silica production for both #ow con"gurations, the
The aggregation is carried out in a coaxial mixing jet inlet concentrations of silica and sodium are adjusted to
reactor as illustrated in Fig. 1. The tubular reactor has an the inlet #ow rates.
internal diameter of 5;10\ m and the central injection The characteristics of both #ows are given in Table 2.
tube 1;10\ m in diameter. The central in#ow (A) feeds
the reactor within a silica solution, i.e. one of the react- 5.1.2. Passive tracer validation
ants that contains particles, whereas the annular in#ow In order to validate the CFD and the Monte-Carlo
(B) is composed of sodium ions, that will increase the calculations, simulation results for #ow con"guration
ionic strength of the solution, promoting aggregation of 1 are compared with experimental data. These data are
the particles. Mixing between the two streams is realised provided from Guiraud, Bertrand, and Costes (1991),
by the shear created by the high-velocity di!erence. who characterised turbulent #uid #ow by laser Doppler
2450 L. Falk, E. Schaer / Chemical Engineering Science 56 (2001) 2445}2457
Fig. 2. (a) Mean axial velocity on the axis. (b) Kinetic turbulent energy. Points: experiments (Guiraud et al., 1991). Lines: CFD.
Fig. 3. (a) Mean concentration on the axis. (b) R.m.s. concentration. Points: experiments (Guiraud et al., 1991).
dm j!3
H" m m for 0)j)6,
dt 3 H
By choosing the following inlet concentration (C "1
and C "0) for both con"gurations, the calculated and
values of the mixture fraction are:
m
E at z"0 (inlet): f "f "0.04, ¸ " )250 nm (20)
m
E at in"nite length f (complete mixing): f "
with depending only on physico-chemical parameters:
0.4, f "0.14,
in order to compare both con"gurations, a dimensionless 9387
"1.88;10 exp 8.466;10\[Na>]! , (21)
mixture fraction is used: ¹
where is in s\, [Na>] in mol m\ and ¹ in Kelvin.
( f!f )
.
>" (19)
( f !f )
5.2.2. PDF modelling of silica aggregation
The longitudinal evolution of the dimensionless mixture From the above relations, the moment's kinetics ex-
fraction for the two con"gurations is drawn in Fig. 5. The pression becomes, in length coordinates:
mixing performances are very similar in both cases from
j!3
the inlet to about 3 cm. For greater axial positions, the S H" m m (22)
K 3 H
macromixing performance is signi"cantly higher for con-
"guration 1. More precisely, complete mixing is reached with depending on temperature and on sodium con-
when the axial position exceeds about 10 cm while for centration.
con"guration 2 the maximal value of > is about 0.9 at the Since no chemical reaction occurs, the silica and
outlet of the reactor. In conclusion, the back-mixing sodium concentration pro"les depend only on mixing.
occurring in con"guration 1 strongly enhances the mac- Silica aggregation is calculated in the same geometry as
romixing performance. presented in Section 5.1. For both #ow con"gurations,
silica particles are fed into the inner tube, and sodium
5.2. Application to silica aggregation ions introduced into the annular space. The sodium and
silica inlet concentrations are identical (6 M) for #ow
The in#uence of macro- and micromixing on nuclea- con"guration 1, and the molar #ow rates have been
tion and particle growth has been identi"ed by several conserved for both con"gurations ; the sodium and silica
researchers. Because of complexity, the study of the com- concentrations for #ow con"guration 2 are then (accord-
plete precipitation process undergoing nucleation, ing to the velocity values, see Table 2) respectively, 3 and
growth and agglomeration cannot be presented here. 12 M. The inlet length coordinate moments are those of
Therefore, for illustration of the PDF method, we choose a monodisperse, 12 nm in diameter (¸ "12 nm), par-
to focus on aggregation (i.e. in the absence of super- ticles solution:
saturation) alone. The studied aggregating system is sil-
[SiO ]
ica. m" (¸ )H\, (23)
H
T N
5.2.1. Kinetic modeling of silica particles aggregation where [SiO ] is the silica weight concentration.
The aggregation kinetics of precipitated silica particles If the local concentration of sodium is null, almost no
have been studied in a batch laboratory vessel (Schaer, aggregation takes place (for [Na>]"0 and
2452 L. Falk, E. Schaer / Chemical Engineering Science 56 (2001) 2445}2457
Fig. 6. (a) [Na>] "eld for #ow con"guration 1 ([Na>] values in mol m\). (b) m "eld for #ow con"guration 1 (m values in m m\). (c) [Na>] "eld
for #ow con"guration 2 ([Na>] values in mol m\). (d) m "eld for #ow con"guration 2 (m values in m m\).
m m m and the y-axis corresponds to the longitudinal coordi-
¸ " , " !1 . (24)
m * m nate (from 0 to 0.2 m).
The "elds of these two non-reacting quantities (since
From Eq. (24), it can be noticed that the particle's size the third moment is conserved during aggregation) trans-
distribution characteristics are de"ned by the "rst four late the mixing conditions of the two inlet #uids. For
moments, and that the calculation of the fourth to sixth z"0 (initial axial position), [Na>] is null in the inner
distribution moments is useless. Therefore, the illustra- tube and of course maximum in the annular space. It
ting results concern the "rst four moment's "elds at any decreases slowly as it mixes with the inner #ow. On the
point of the #ow. contrary, m is maximum in the central injection tube
L. Falk, E. Schaer / Chemical Engineering Science 56 (2001) 2445}2457 2453
Fig. 9. Average radial values of the mean number size ¸ against axial position: (a) #ow con"guration 1; (b) #ow con"guration 2.
Fig. 10. Average radial values of the variance against axial position: (a) #ow con"guration 1; (b) #ow con"guration 2.
*
The formation of larger particles induces a distribution ber size ¸ and of the variance . To reduce the
*
broadening. However, because of the micromixing limit- statistical noise level, more notional particles per com-
ing step, the size variation does not proceed as fast as in putational cell could be employed, at the cost of higher
perfect mixing conditions. computational time. As the power of computers is con-
tinuously increasing, complex and "ne-scale calculations
now possible on large supercomputer, will be e!ective in
6. Conclusion a few years on PC.
The presented method based on the uncoupling of
The PDF method coupled with CFD description of #ow "eld and reaction calculation, is only supportable
the #ow is an e!ective tool for studying turbulent reactive for dilute systems. However, for more concentrated
#ows, involving complex kinetics as in precipitation pro- systems, PDF methods remain applicable, but back-
cesses. All scales of mixing, including macro- and mi- action of the particle dynamics/reaction chemistry on the
cromixing may be included, which are of primary parameters describing the continuous phase #ow has to
importance for mixing sensitive reactions. Furthermore, be done.
the complex kinetics may be treated exactly without any
assumption. The method is well adapted to localise speci-
"c rate processes, such as nucleation, growth and ag- Notation
gregation in complex #ows and to handle the calculation
of the particle size distribution expressed in moments. A model constant in nucleation kinetic
The illustration of silica aggregation showed that mixing a mean ionic activity, mol m\
!
may drastically a!ect the aggregation process and that B nucleation rate, no m\ s\
all coupled phenomena must be considered to describe B model constant in nucleation kinetic,
properly fast reaction processes. The stochastic calcu- no m\ s\
lation procedure used for the PDF method may, how- B(v) birth rate function, expressed as a function
ever, lead to some #uctuations in PSD moments, which of particle volume, no m\ s\
can be a di$culty for accurate estimation of mean num- C vector of species concentrations
2456 L. Falk, E. Schaer / Chemical Engineering Science 56 (2001) 2445}2457
Mersmann, A., & Kind, M. (1988). Chemical engineering aspects of Van Leeuwen, M. L. J., Bruinsma, O. S. L., & Van Rosmalen, G. M.
precipitation from solution. Chemical Engineering Technology, 11, (1996). In#uence of mixing on the product quality in precipitation.
264}276. Chemical Engineering Science, 51(11), 2595}2600.
Pope, S. B. (1985). PDF methods for turbulent reactive #ows. Progress Villermaux, J., & Devillon, J. C. (1972). RepreH sentation de la coales-
in Energy and Combustion Science, 11, 119}192. cence et de la redispersion des domaines de seH greH gation dans un
Schaer, E. (1997). Conception d'un proceH deH pour la production de #uide par un modèle d'interaction pheH nomeH nologique. Proceedings
microparticules "ltrables et redispersables. Ph.D. thesis, INPL of the 2nd International symposium on chemical reaction engineering,
* Nancy, France. Amsterdam (pp. B1}13).
Smoluchowski, M. von. (1917). Versuch eine matematischen Theorie Wei, H., & Garside, J. (1997). Application of CFD modelling to precipi-
der Koagulationskinetik kolloider LoK sungen. Zeitschrift fu( r tation systems. Transactions of the Institution of Chemical Engineers,
physikalische Chemiee, 92, 129}168. 75(A), 219}227.