Professional Documents
Culture Documents
8575
The paper presents a novel methodology for the estimation of the shape of the crystal size distribution (CSD)
during a crystallization process. The approach, based on a combination of the quadrature method of moments
(QMOM) and the method of characteristics (MOCH), provides a computationally efficient solution of the population
balance equation (PBE) and hence a fast prediction of the dynamic evolution of the CSD for an entire batch.
Furthermore, under the assumption that for supersaturation-controlled crystallization the main phenomenon is growth,
an analytical CSD estimator is derived for generic size-dependent growth kinetics. These approaches are evaluated
for the crystallization of potassium alum in water. The model parameters are identified on the basis of industrial
experimental data, obtained using an efficient implementation of supersaturation control. The proposed methods
are able to predict and reconstruct the dynamic evolution of the CSD during the batch. The QMOM-MOCH
solution approach is evaluated in a model-based dynamic optimization study, which aims to obtain the optimal
temperature profiles required to achieve desired target CSDs. The technique can serve as a soft sensor for predicting
the CSD, or as a computationally efficient algorithm for off-line design or online adaptation of operating policies
based on knowledge of the full CSD data.
1. Introduction
Crystallization from solution is an industrially important unit
operation due to its ability to provide high purity separation. It
is widely used in industries such as petrochemicals, pharmaceuticals, microelectronics, and food processing. Batch cooling
crystallization provides the advantages of being simple, flexible,
and generally requiring less process development and investment
than many other separation/purification techniques. However,
despite the long history and widespread application of batch
crystallization, there are a disproportionate number of problems
associated with its control. The crystal size distribution (CSD)
is a key factor in the production of a high quality product and
determines the efficiency of the downstream processes.1,2
Therefore many problems in downstream processes can be
attributed to poor particle characteristics established in the
crystallization step. The control objectives for batch crystallization processes can be defined in terms of product purity,
crystal habit, morphology, average particle size, crystal size
distribution, bulk density, product filterability, and dry solid flow
properties, which all depend on the CSD. The main difficulty
in batch crystallization is to accomplish a uniform and reproducible CSD.3 Online control during the process allows for
improved crystalline product quality, shorter process times, and
reduction or elimination of compromised batches. Even if some
of these objectives can be expressed in terms of the moments
of the distribution, knowing and predicting the entire shape of
the distribution allows the design and adaptation of operating
policies to achieve improved product quality.
A potential way to enhance the control of CSD is to use
supersaturation control,4-10 or direct nucleation control,11,12
which drives the process within the metastable zone to avoid
nucleation or to generate controlled nucleation/dissolution
* To whom correspondence should be addressed. E-mail: z.k.nagy@
lboro.ac.uk. Tel.: +44 1509 222516. Fax: +44 1509 223923.
Loughborough University.
BASF SE.
8576
fn(L) Lj dL,
j ) 1, 2, ...,
(2)
fn(L, t)
w (t) (L (t), L)
i
(3)
i)1
where Nq is the number of quadrature points and the corresponding weights, wi, and abscissas, Li, can be determined
through the product-difference (PD) algorithm41 or via direct
solution of a differential-algebraic (DAE) system,42 based on
the idea of minimizing the error committed by replacing the
integral from the moment definition with its quadrature
approximation,
j )
Nq
fn(L) Lj dL
wL
j
i i
(4)
i)1
q
dj
)j
wiLj-1
i G(S, Li ;g) +
dt
i)1
B(S;b)rj0,
j ) 1, 2, 3,...
(5)
fn(L, t)
(G(S, L;g)fn(L, t))
+
) B(S;b)(r0, L)
t
L
dfn
dL fn
dt fn
)
+
dZ
i
dZ
i L
dZ
i t
(1)
(6)
(7)
(8)
dG(S, L;g)
dfn(L, t)
) -fn(L, t)
+ B(S;b) (r0, L)
dt
dL
(9)
1 if L [0, r0]
0 if L [0, r0]
8577
(10)
(11)
8578
fn(L) ) fn,0(L0) 1 +
kgSg(1 - p)t
(1 + L0)1-p
p/(p-1)
(16)
(12)
where g )[kg,g,,p] is the growth parameter vector, for growthdominated systems (B ) 0) the model eqs 8-9, reduce to the
form:
dL
) kgSg(1 + L)p
dt
(13)
dfn(L, t)
) -kgSgp(1 + L)p-1fn(L, t)
dt
(14)
(1-p)
- 1)/
(15)
8579
value
8.5708
0.0050
1.5777
1.0000
0.0380
3.4174
min{wf
Nd
(f
exp
2
V,k(Ll)-fV,k (Ll)) + wC
k)1 l)1
(C
2
- Cexp
k ) }
k)1
(17)
subject to
(18)
min e e max
(f
(19)
fV,l ) fn,lLl /
3
n,lLl
Ll)
l)1
8580
Figure 8. Comparison between measured and simulated CSD using the analytical CSD estimator (experiment B).
8581
Figure 10. Simulated dynamic evolution of CSD with optimal temperature profile throughout the batch.
reaches its constant set point value, the predicted CSD will
converge to the correct value.
5. Model-Based Dynamic Optimization of Temperature
Trajectories for Shaping the CSD
The combined QMOM-MOCH approach described in section 2 was used to solve the PBM in a model-based dynamic
optimization scheme, with the aim to determine optimal
temperature trajectories which yield desired target CSDs at the
end of the batch. The final CSD is dependent on the supersaturation profile created over the batch time, hence the cooling
trajectory is of critical importance for the final CSD. In the
optimization both the temperature trajectory and the batch time
were optimized. For the solution of the dynamic optimization
problem the batch time horizon [0, tf] is divided into Nb equally
spaced time intervals of t (stages), with discrete time steps tk
) jt, j ) 0, 1, ..., Nb, and the temperature trajectory is
approximated by a piece-wise linear function determined by the
fixed initial temperature at t ) 0, T(0), and the slopes RT(j) in
each discretized period t. Note, that since the batch time is
also optimized the duration of the time interval t changes
during the optimization, but the number of disretizations Nb is
fixed. This formulation allows an easy incorporation of the
temperature rate constraints (as bounds on the decision variables
RT(j)), which are very important to obtain a practically implementable temperature trajectory. The optimization problem is
formulated as follows:
K
min
RT(j),tf
Nd
(f
target
2
v,k(Ll)-fV,k (Ll))
k)1 l)1
(20)
subject to
RT,min e RT( j) e RT,max,
j ) 0, 1, ..., Nb
(21)
0 e tf e tf,max
(22)
C(tf) e Cf,max
(23)
with RT(j) the elements of the vector containing the slopes, (dT/
dt) for the temperature trajectories and tf is the total batch time,
C(tf) is the solute concentration at the end of the batch and Cf,max
is the maximum acceptable concentration at the end of the batch
target
are the values of the
to achieve the required yield, fV,k and fV,k
simulated and the target volume probability distribution functions at the discrete time steps k ) 1, ..., K, where measurement
data were available, corresponding to the discretized sizes Ll, l
) 1, ..., Nd, with Nd being the number of experimental size bins.
The optimization problem is solved using a sequential quadratic
programming (SQP) approach implemented in the Matlab
function fmincon. The kinetic parameters used are described in
Table 1.
Figures 10 and 11 show the results of the distribution shaping
optimal control. The seed and target distributions are shown
on the first plot (at t ) 0) in Figure 10. The target CSD is a
fictitious bimodal distribution. For simulation purposes a monomodal seed was selected, represented by a Gaussian distribution
with mean of 54 m and standard deviation of 15 m, chosen
based on CSD measurement of the seed used in the experiments.
Figure 10 shows the dynamic evolution of the CSD toward the
target bimodal distribution, when the optimal temperature
trajectory shown in Figure 11a is implemented in the simulation.
The final CSD is in good agreement with the target distribution.
8582
Figure 11. Optimal control results of the simulations for bimodal target distribution with pronounced secondary peak: (a) optimized temperature profile with
50 discretization points, (b) phase diagram showing solubility and optimal operating curve, (c) concentration profile during the batch (d) supersaturation
profile (kg/kg slurry) during the batch, (e) nucleation rate profile during the batch, (f) growth rate profile during the batch.
Figure 12. Optimal control results for bimodal target distribution with minimized fines due to secondary nucleation: (a) comparison of model and target
distribution, (b) optimized temperature profile with 30 discretization intervals, (c) concentration profile during the batch, (d) supersaturation profile during
the batch.
The first 10-15 min of the batch are devoted mainly to the
growth of the seed crystals, which approach rapidly the larger
size mode of the target distribution, due to the high driving force
in this periodslarge supersaturation as can be seen on Figure
11b,d. The supersaturation achieves its peak value at about 10
min (see Figure 11d) yielding the appearance of a second mode
due to secondary nucleation, which develops clearly by t ) 17
min. The optimal temperature profile, shown in Figure 11a, was
in Figure 11c. The temperature profile shows several distinguishing slope changes, which correspond to secondary nucleation events generated to achieve the desired shape of the target
distribution. After the initial peak at around 10 min, the
supersaturation raises again between 25-40 min and subsequently toward the end of the batch between 50-55 min, as
can be seen in Figure 11d. These increases in the supersaturation
lead to additional nucleation events and growth, required to
achieve the target distribution, as shown in Figure 11e,f. Figure
10 shows that the first mode of the bimodal target distribution
is skewed requiring the subsequent nucleation events to improve
the shape of the obtained CSD. The third region of increased
supersaturation has significantly smaller values and is mainly
to facilitate growth of the larger particles to give a better fit to
the second mode of the target CSD; Figure 10 shows no
significant change in the first mode of the CSD after t ) 41
min, indicating mainly the growth of the second mode.
The previous target CSD showed a pronounced secondary
peak to illustrate the ability of the approach to achieve a bimodal
distribution. In industrial scenarios, the major emphasis is to
suppress the nucleation and minimize the formation of small
particles during the crystallization process. Therefore the
temperature profile was optimized for another fictitious target
distribution, with a significantly smaller fraction of fine particles.
Figure 12a shows that the CSD predicted by the model is very
close to the target distribution. The slight deviation in the fitting
of the first peak of the target distribution indicates that it is not
possible to achieve any arbitrary target distribution by designing
the cooling profile only. There are limitations on the potentially
attainable CSD shapes given by the particular nucleation and
growth kinetics, as well as the shape of the seed distribution.
The optimized temperature profile ranges from 40 to 7 C in
Figure 12b and shows a sequence of gradually increasing cooling
rate until t ) 60 min, when the nucleation of the particles
required for the second small peak occurs. Figure 12d shows
that this nucleation event leads to a decrease in the supersaturation, to a level which is maintained for the next 20 min of
the batch, initially by reducing the cooling rate. The cooling
rate then is gradually increased again to produce a slightly higher
supersaturation level to complete the growth of the particles to
the required size and to achieve the required yield of 60%; see
Figure 11c. The batch time for this case has increased
significantly compared to the previous case to approximately
95 min. The number of discretizations used is 30 since the
previous case showed that no significant improvement was
obtained by increasing the discretization to 50. Although the
secondary peak is significantly reduced in the second case, note
that with the desired shape of the primary mode of the target
distribution and required yield it is not possible to eliminate
completely the secondary peak from the resulting CSD. However
the two case-studies illustrate that the proposed method can be
used to obtain optimal temperature trajectories for designing
different target distributions. The combined QMOM-MOCH,
model can predict the CSD efficiently and can be used with
off-line or online optimization approaches for designing crystallization systems to produce consistent target CSDs.
6. Conclusions
The paper describes a new methodology of solving population
balance equations. The approach combines the quadrature
method of moments (QMOM) with the method of characteristics
(MOCH), and provides a computationally efficient method of
reconstructing the crystal size distribution (CSD). The approach
can also be used to design optimal temperature profiles for
8583
8584