Chemical Engineering Science 127 (2015) 362–373

Contents lists available at ScienceDirect

Chemical Engineering Science

journal homepage: www.elsevier.com/locate/ces

Advanced control approaches for combined cooling/antisolvent

crystallization in continuous mixed suspension mixed product
removal cascade crystallizers
Yang Yang, Zoltan K. Nagy n
School of Chemical Engineering, Purdue University, West Lafayette, IN 47907, USA

H I G H L I G H T S

Systematic analysis of a MIMO control system for continuous crystallization.

Attainable regions defined for size and yield using different control approaches.

Feasibility of decentralized PID control approach is analyzed.

Different nonlinear model predictive control approaches are proposed and analyzed.

art ic l e i nf o a b s t r a c t

Article history: The control approaches for a continuous two stage mixed suspension mixed product removal (MSMPR)
Received 16 November 2014 cascade crystallizer are studied in this work. Both cooling and antisolvent addition are applied at both
Received in revised form stages to manipulate the process. Considering both crystal size and yield as controlled variables, the
14 January 2015
attainable region of crystal size and yield is obtained. Two advanced control schemes are discussed:
Accepted 26 January 2015
(1) decentralized proportional-integral-derivative (PID) control; and (2) nonlinear model predictive
Available online 2 February 2015
control (NMPC). While decentralized PID control framework is proved to require change of control
Keywords: structure when a relatively large operating region is essential, nonlinear model predictive control
Combined cooling/antisolvent scheme shows superior performance for fast target product property change-over and disturbance
crystallization
rejection.
Mixed suspension mixed product removal
& 2015 Elsevier Ltd. All rights reserved.
crystallizer
Nonlinear model predictive control
Optimal control
Crystal size
Yield

1. Introduction efficiency of downstream operations (e.g. filtration, washing and

Crystallization is an efficient and economical unit operation that
is extensively used in pharmaceutical industry to purify active
pharmaceutical ingredients (APIs) (Qamar et al., 2010). Typically a
batch crystallizer is utilized, with either cooling, antisolvent addi-
tion or evaporation applied to generate supersaturation, which is
the thermodynamic driving force for a crystallization process. In
addition, for high yield purpose, combined cooling/antisolvent
crystallization (CCAC) is also widely used (Sheikhzadeh et al.,
2008). Additionally, CCAC has great advantages to achieve desired
crystal size distribution (CSD) (Nagy et al., 2008; Yang and Nagy,
2014; Lindenberg et al., 2009), which not only determines the
n
Corresponding author. Tel.: þ 1 7654940734.
E-mail address: zknagy@purdue.edu (Z.K. Nagy).
drying), but also influences bioavailability and dissolution behavior
of an API (Nagy et al., 2011; Vetter et al., 2014). Thus it is essential to
design suitable control strategies for crystallization processes.
In general, the crystallization process can be controlled via either
model-free or model-based approaches. The applications of both
approaches in batch crystallization have been studied in depth in
literature. Model-free approaches, such as direct nucleation control
(DNC) (Abu Bakar et al., 2009; Saleemi et al., 2012), or super-
saturation control (SSC) (Gron et al., 2003; Nagy and Aamir, 2012)
are methodologies that maintain the operating curve within the
metastable zone to avoid or control nucleation or generate con-
trolled dissolution. For model-based approaches, typically a popula-
tion balance model (PBM) is used to describe the evolution of the
CSD in the crystallization process and to obtain open-loop optimal
temperature or/and antisolvent addition profiles that can produce
desired CSD (Acevedo and Nagy, 2014; Rawlings et al., 1993; Xie et

http://dx.doi.org/10.1016/j.ces.2015.01.060
0009-2509/& 2015 Elsevier Ltd. All rights reserved.
 Y. Yang, Z.K. Nagy / Chemical Engineering Science 127 (2015) 362–373
al., 2001; Zhang and Rohani, 2003). In addition, more advanced
model-based approaches that solve the open loop optimization
repeatedly, such as model predictive control (MPC) has been
applied in batch crystallization process (Hermanto et al., 2009;
Kalbasenka et al., 2007; Nagy and Braatz, 2003). MPC uses the
mathematical model and real time measurements to optimize the
current operating curve, based on the predicted future behavior of
the system. For systems of high complexity and nonlinearity,
nonlinear model predictive control (NMPC) is used instead of linear
model predictive control (LMPC). Hermanto et al. (2009) utilized
NMPC strategy to control the polymorphic transformation in a
batch crystallization process. It was proved to be more robust than
other existing control strategies like temperature control or con-
centration control. A robust extended Kalman filter-based NMPC
framework was developed by Nagy and Braatz (2003) for batch
crystallization processes. The proposed NMPC approach consider-
ably enhanced robust performance as compared with open-loop
optimal control. Although in general these model-based approaches
are more complicated to design compared to model-free
approaches, they can enhance process understanding, provide
theoretically optimal recipe, and require much smaller number of
experiments than statistical experimental design (Nagy et al., 2008;
Nagy and Aamir, 2012).
Due to great operating flexibility and short development time
required, the widespread industrial use of batch crystallizer is very
common (Randolph and Larson, 1988). However, the intrinsic dis-
advantages of batch crystallization, such as low process and equip-
ment efficiency and batch to batch inconsistency, have strongly
motivated the development of continuous crystallization systems.
Theoretically, continuous crystallization could have much better
quality consistency, process and equipment efficiency, and productiv-
ity as compared with batch crystallization (Quon et al., 2012). Several
types of continuous crystallizers have been developed, including
mixed suspension mixed product removal (MSMPR) crystallizer
(Alvarez et al., 2011; Quon et al., 2012; Wong et al., 2012; Zhang et
al., 2012), plug flow crystallizer (PFC) (Alvarez and Myerson, 2010;
Eder et al., 2010; Vetter et al., 2014) and oscillatory baffled crystallizer
(OBC) (Lawton et al., 2009) to name only a few. Compared to the other
two, MSMPR crystallization processes are more frequently applied in
industrial crystallization since they provide a smaller technology
change from batch and are simpler to operate (Vetter et al., 2014).
MSMPR crystallizer is an ideally well-mixed vessel that has feed
solution continuously entering and product slurry continuously with-
drawn. It can be used in various configurations, including single stage,
multistage, or with recycle loops (Alvarez et al., 2011; Quon et al.,
2012; Su et al., 2015; Vetter et al., 2014; Wong et al., 2012; Zhang et
al., 2012). Cooling, or antisolvent addition, or both could be applied to
generate supersaturation at a certain MSMPR crystallization stage. It
can be operated at steady-state to produce crystals of consistent
purity, yield, CSD and polymorphic form. Vetter et al. (2014) modeled
an MSMPR cascade crystallization process using PBM. CCAC of aspirin
was used as a case study, in which the particle size attainable regions
of MSMPR cascade crystallizer of different number of stages were
obtained and compared with batch crystallizer and PFC. Quon et al.
(2012) implemented and optimized a two stage MSMPR cascade
crystallizer for continuous reactive cooling crystallization of aliskiren
hemifumarate. Crystals of both high yield and purity were obtained.
Alvarez et al. (2011) implemented a three stage continuous MSMPR
cascade crystallizer to produce cyclosporine crystals using cooling to
generate supersaturation. A PBM was developed in the same work,
and used to optimize crystal size, yield and purity, as well as to obtain
crystallization kinetic parameters. Wong et al. (2012) constructed two
continuous single stage MSMPR crystallizers with recycling loop for
cooling crystallization and CCAC. They experimentally demonstrated
the feasibility and potential advantages (e.g. high yield) of having
recycle loop in a single stage MSMPR crystallizer. A continuous CCAC
363
NMPC
FC FC
Feed solution Antisolvent Antisolvent
FBRM UV/Vis
LC LC
TC TC
Fig. 1. Schematic representation of a hierarchical NMPC implementation structure.
FC, TC, LC are flow rate controller, temperature controller and level controller,
respectively.
process using two stage MSMPR cascade crystallizer was implemen-
ted by Zhang et al. (2012), and was experimentally proved to be able
to well control CSD, yield and purity. In these continuous MSMPR
crystallization systems, process analytical technology (PAT) tools like
focused beam reflectance measurement (FBRM) and high perfor-
mance liquid chromatography (HPLC) were successfully implemen-
ted. In addition, novel PAT implementations like video camera have
also been applied in some other continuous processes (Simon and
Myerson, 2011). The development of novel PAT tools has enabled the
progress of both model-free and model-based control approaches
(e.g. NMPC) (Nagy et al., 2013; Simone et al., 2014a, 2014b).
While model-based optimization, design and experimental
investigation of continuous single stage and multistage MSMPR
systems is becoming more common in the literature, to the author's
best knowledge, currently there are no published papers working
on model-based advanced control approaches for CCAC in MSMPR
cascade crystallizers, with both crystal size and yield controlled at
the same time. To achieve this goal, this work systematically studied
the feasibility of two different control schemes: decentralized PID
control and nonlinear model predictive control (NMPC). For decen-
tralized PID control approach, local linearization method and
relative gain array (RGA) analysis were used to examine its
feasibility. Whereas for NMPC approach, its performance was
evaluated under both servo control and regulatory control scenar-
ios. In this work, not only nucleation and growth kinetics, but also
controlled dissolution kinetics were considered. In the case study,
aspirin (acetylsalicylic acid), ethanol and water were used as model
compound, solvent and antisolvent, respectively.
The schematic representation for a closed-loop hierarchical NMPC
implementation structure in continuous two stage MSMPR cascade
crystallizer is presented in Fig. 1. In principle, PAT tools like FBRM and
UV/Vis spectroscopy can be used for online measurements of CSD
and solute concentration, from which mean crystal size and yield can
be inferred in real time. As a result, the process model can be
updated during the process using those measurements via closed-
loop optimal control. Sensor noise and error were not considered in
this work. In addition, in practice typically low level flow rate and
temperature controllers would be used in a hierarchical control
structure with the NMPC, as shown in Fig. 1. In this work, in order
to simplify the simulation study, antisolvent addition rates and
temperatures were manipulated directly from NMPC scheme.
2. Population balance model of the continuous two-stage
CCAC-MSMPR system
In this work, a continuous two stage MSMPR cascade crystal-
lizer is modeled using PBM. Both the temperature and antisolvent
364 Y. Yang, Z.K. Nagy / Chemical Engineering Science 127 (2015) 362–373

dμjð2Þ 1 1
¼ μð1Þ ð2Þ ð2Þ
j  τ μj þ jG2 μj  1 ; j Z 1; ð8Þ
Feed solution Antisolvent F1 Antisolvent F2 dt α1 τ 1 2

dμ0ð1Þ F ð0Þ 1
¼ out μ0ð0Þ  μð1Þ þ B1 ; ð9Þ
dt V1 τ1 0
Temperature T1 Temperature T2
dμ0ð2Þ 1 ð1Þ 1 ð2Þ
¼ μ  μ þ B2 ; ð10Þ
dt α1 τ 1 0 τ 2 0
Crystal mean size Ln
Yield Y where Bi is the nucleation rate [ ml  1 min  1] of stage i.
Assume ci is solute concentration [g/ml] of stage i, kv is crystal
Fig. 2. Schematic representation of the continuous CCAC process in a two stage
MSMPR cascade crystallizer.
volume shape factor [dimensionless], ci is crystal density [g/cm3].
The mass balance equation of solute of stage i can be written as

addition rate can be manipulated at both stages, in order to control dc1 F ð0Þ c1
¼ out c0   3kv ρc G1 μð1Þ
2 ; ð11Þ
supersaturation levels and residence times (Fig. 2). The crystal dt V1 τ1
mean size and yield of the second stage are the two controlled
dc2 c1 c2
variables. The following assumptions are used in this work: ¼   3kv ρc G2 μð2Þ
2 : ð12Þ
(1) vessel is ideally well-mixed (2) the outlet flow contains exactly dt α1 τ 1 τ 2
the same compositions as inside the vessel (3) only nucleation, The values of kv and ρc are chosen as π =6 and 1.4 g/cm3,
growth and dissolution are considered, whereas breakage or respectively (Lindenberg et al., 2009). c0 is the concentration of
agglomeration are negligible and (4) mixture of aspirin, ethanol feed solution, which is taken to be 0.40 g/ml.
and water is ideal solution. Define vi as the volumetric fraction of solvent in solvent/
The PBM of stage i can be expressed as (Ramkrishna, 2000; antisolvent mixture of stage i, and v0 as the solvent fraction of
Randolph and Larson, 1988; Vetter et al., 2014): feed solution (v0 is chosen as 0.792). Then the following mass
∂V i ni ðL; tÞ ði  1Þ ∂V i Gi ni ðL; tÞ balance equations of solvent can be derived:
¼ F out ni  1 ðL; tÞ  F ðiÞ
out ni ðL; tÞ  ; i ¼ 1; 2; ð1Þ
∂t ∂L
dv1 F ð0Þ v1
where L is the characteristic crystal size [cm], ni ðL; tÞ is the ¼ out v0  ; ð13Þ
dt V1 τ1
volumetric number density [/ml] of stage i at time t [min], Gi is
the growth rate [cm/min] of stage i, F ðiÞ out is the outlet volumetric dv2 v1 v2
flow rate [ml/ min] of stage i, F ð0Þ ¼  : ð14Þ
out is the volumetric flow rate dt α1 τ 1 τ 2
[ml/min] of feed solution. F ð0Þ
out is fixed at 5 ml/min. V i [ml] is the
solution volume of stage i. In this work, V 1 and V 2 are assumed to The above equations can be easily expanded if more than two
be constant. This assumption is easy to implement by using level stages are being used.
controllers. V 1 and V 2 are fixed at 65 and 167 ml, respectively. As a The crystallization kinetics are described by the following
result, Eq. (2) can be obtained: empirical equations (Lindenberg et al., 2009):
 
X
i kG2
G ¼ kG1 exp  ðcn ðS  1ÞÞkG3 ; if S Z 1; ð15Þ
F ðiÞ ð0Þ
out ¼ F out þ F j; i ¼ 1; 2; ð2Þ RT
j¼1
   
where F 1 , F 2 are antisolvent addition rates [ml/min] at the first kB2 kB3
B ¼ kB1 exp  exp  2 ; if S Z 1; ð16Þ
and the second stage. RT ln S
Eq. (1) can be simplified using the method of moments  
kG2
(Randolph and Larson, 1988). The j th moment of stage i can be DG ¼  kG1 exp  ðcn ð1  SÞÞkG3 ; if S o 1; ð17Þ
defined as RT
Z 1 !
 
μðiÞ
j ðtÞ ¼ Lj ni ðL; tÞdL; i ¼ 1; 2; j ¼ 0; 1; 2; 3; 4: ð3Þ kB2 kB3
0 DB ¼  kB1 exp  exp  2 ; if S o1; ð18Þ
RT ln ð1=SÞ
We assume growth rate is size-independent, then Eq. (4) can be
derived from Eqs. (1) and (3): S ¼ c=cn ; ð19Þ
ðiÞ
dμ ði  1Þ ði  1Þ where S is the supersaturation ratio [dimensionless], cn is the
μj  F ðiÞ ðiÞ ðiÞ
out μj þ V i Gi jμj  1 ;
j
Vi ¼ F out j Z 1: ð4Þ
dt equilibrium solubility [g/ml], T is temperature [K]. Assume T 1 , T 2
Residence time and dilution factor of stage i are expressed as τi are temperatures [1C] at the first and the second stage. DG and DB
and αi , respectively: are dissolution rates. When S o 1, the negative values of nucleation
rate and growth rate are used to approximate dissolution rates
Vi
τi ¼ ; i ¼ 1; 2; ð5Þ (Qamar et al., 2010; Nagy et al., 2011). kG1 , kG2 , kG3 , kB1 , kB2 , kB3 are
F ðiÞ
out empirical parameters from literature (Lindenberg et al., 2009).
Solubility data of aspirin in ethanol and water mixture at different
Viþ1
αi ¼ ; i ¼ 1: ð6Þ temperatures is also available in literature (Lindenberg et al.,
Vi
2009).
Then Eqs. (4)–(6) can lead to a group of ordinary differential Assume Ln and Y are number based mean crystal size [μm] and
equations (ODEs) for a two stage cascade: yield [dimensionless] of the second stage, respectively, calculated
with the following equations:
dμð1Þ F ð0Þ
out ð0Þ 1
j
¼ μ  μð1Þ þ jG1 μjð1Þ
1 ; j Z1; ð7Þ
Ln ¼ μ1ð2Þ =μð2Þ
dt V1 j τ1 j 0 ; ð20Þ
 Y. Yang, Z.K. Nagy / Chemical Engineering Science 127 (2015) 362–373
F1
Two stage Ln
F2 MSMPR cascade
T1 crystallizer Y 6
T2
Fig. 3. Schematic representation of investigated multi-input multi-output system.
Table 1
Five possible different control approaches for two-stage MSMPR cascade crystal-
lization systems.
Control method Controlled output variable Manipulated input variable
Nucleation control Ln F1, T1
Growth control Y F2, T2
Antisolvent control F1, F2
Temperature control T1, T2
Global control F1, F2, T1, T2
c2 F ð2Þ
out
Y ¼ 1 : ð21Þ
c0 F 0
In this study, heat transfer and model mismatch are not
considered.
3. Results and discussion
3.1. Attainable region of crystal size and yield
According to Fig. 2, the proposed control structure can be
considered as a multi-input multi-output (MIMO) system, as shown
in Fig. 3. Since in this process more potential manipulating inputs
exist than desired controlled variables, it is interesting to evaluate
different control architectures. Five control methods are investigated
based on the inputs that are manipulated, as presented in Table 1.
The basic operating scenario was defined as: F 1 , F 2 , T 1 , T 2 equal
1.05 ml/min, 2.60 ml/min, 25 1C, 25 1C, respectively. It was found
that crystal mean size Ln and yield Y would reach constant values of
369 mm and 0.685 after the system reaching steady-state, by
simulating the process based on Eqs. (2), (5)–(21). For the manipu-
lated input variables, the following operating limits were used:
0 r F 1 ; F 2 r 10 ml=min;
10rT 1 ; T 2 r40o C:
The attainable regions of crystal size and yield of the proposed five
control methods can be obtained by simulating the process considering
model Eqs. (2), (5)–(21) and operating limits (22) and (23). Since the
process is highly nonlinear and dissolution is also considered, it is
possible that the maximum and minimum crystal sizes corresponding
to a certain yield may not guarantee that the crystal sizes in between
are also achievable. Therefore an exhaustive search method is used
based on a Monte Carlo type simulation study. The process is simulated,
and the crystal size and yield are recorded repeatedly, with each
variable taking a value within the operating limits and changing with
a small step in the next simulation. The hierarchical for loops were
implemented in Matlab. Finally, the boundary of all recorded crystal
sizes and yields were plotted, as presented in Fig. 4. It should be
acknowledged that the attainable regions (except global control)
depend on the values of the unchanged variables, however the follo-
wing phenomena are always observed. Due to the largest degree of
freedom, the non-square global control method has the largest attain-
able region. Antisolvent and temperature control methods also have
relatively large attainable regions, because they can manipulate super-
saturation at both stages. Since nucleation and growth control methods
can only well control one of the two stages, they have relatively small
attainable regions. In addition, the Pareto front of crystal size and yield
365
1
Nucleation Control
Growth Control
0.9
Antisolvent Control
0.8 Temperature Control
3 Global Control
4
0.7 2
8
1
0.6
9
Yield [-]
0.5 5
7
0.4
0.3
0.2
0.1
0
100 200 300 400 500 600 700 800
Mean size [µm]
Fig. 4. Attainable regions of crystal size and yield in the case of the proposed five
control methods. Operating points 1–7 were used in decentralized PID control
analysis. Operating points 1, 8, 9 were used in NMPC analysis.
is clear in global control method, indicating that generally there is a
compromise between crystal size and yield. Within the operating limits,
large crystal size and high yield are difficult to achieve at the same time.
Different control approaches enable to operate at different combina-
tions of size and yield, and unless the global control is used changing
the product (size) and process (yield) requirements may require the
change of control architecture. These attainable regions are used in the
development of advanced control frameworks discussed later.
3.2. Analysis of system dynamics
Before proposing the control approach, it is beneficial to investi-
gate the basic dynamic features of the studied MIMO system (e.g.
whether the system is linear or nonlinear, or how inputs and outputs
interact). In this work, the response of each output variable is
analyzed when step changes of each input variable are performed.
As shown in Fig. 5, starting from basic operating scenario at 0 min, F1,
F2, T1, T2 are increased by a step change of 0.5 ml/min or 3 1C at 100
ð22Þ
and 400 min, respectively. According to the response curves for
crystal size and yield, it is clear that there is strong interaction
ð23Þ
among the four inputs and the two outputs. In addition, inverse
responses and oscillations are present, and the system is nonlinear
and complex. For example, the crystal mean size will finally increase
if T2 is increased by a small step, however it will decrease if T2 is
increased by a large step (Fig. 5(d)). The size increase is caused by the
controlled dissolution and fines removal of mild heating at the
second stage, whereas large particles are dissolved and crystal mean
size is reduced if the second stage is heated too much.
3.3. Feasibility of decentralized proportional-integral-derivative
(PID) control
Decentralized PID control is a general control approach for
MIMO control systems. It is relatively easy to implement since only
standard PID controllers are required. Using decentralized control
the interacting MIMO system is controlled using two single-input
single-output (SISO) systems. The design of the decentralized
control requires pairing each controlled variable with manipulated
variable. The feasibility of this approach was evaluated using local
linearization method and relative gain array (RGA) analysis. Global
control method is not discussed here since RGA analysis requires
equal number of inputs and outputs. For the proposed other four
control methods, firstly two or three operating points from their
366 Y. Yang, Z.K. Nagy / Chemical Engineering Science 127 (2015) 362–373

Yield [-] Mean size [µm] F1 [ml/min]

Mean size [µm] F [ml/min]

3 4
2
1 3
2
0

2
0 100 200 300 400 500 600 700 0 100 200 300 400 500 600 700

400 400
350 350
300 300
250
250
0 100 200 300 400 500 600 700 0 100 200 300 400 500 600 700

0.72

Yield [-]
0.69
0.7 0.685
0.68 0.68
0 100 200 300 400 500 600 700 0 100 200 300 400 500 600 700
Time [min] Time [min]
T [ °C]

30 30

Mean size [µm] T [ °C]

1

25 25

2
0 100 200 300 400 500 600 700 0 100 200 300 400 500 600 700
Mean size [µm]

400 380
300 370
360
200 350
0 100 200 300 400 500 600 700 0 100 200 300 400 500 600 700

0.7
Yield [-]

Yield [-]

0.68
0.65
0.66 0.6
0 100 200 300 400 500 600 700 0 100 200 300 400 500 600 700
Time [min] Time [min]
Fig. 5. Response of crystal mean size and yield with respect to step change in (a) antisolvent addition rate at stage 1 (F1), (b) antisolvent addition rate at stage 2 (F2),
(c) temperature at stage 1 (T1), and (d) temperature at stage 2 (T2).

Table 2
Selected operating points for local linearization and RGA analysis.

Control method Point F 1 [ml/min] F 2 [ml/min] T 1 [1C] T 2 [1C] Ln [mm] Y[dimensionless]

Nucleation control 1 1.05 2.60 25.0 25.0 369 0.685

2 3.00 2.60 15.0 25.0 214 0.740

Growth control 1 1.05 2.60 25.0 25.0 369 0.685

3 1.05 2.50 25.0 20.0 204 0.761

Antisolvent control 1 1.05 2.60 25.0 25.0 369 0.685

4 3.00 3.00 25.0 25.0 205 0.741
5 0.50 0.50 25.0 25.0 417 0.498

Temperature control 1 1.05 2.60 25.0 25.0 369 0.685

6 1.05 2.60 15.0 15.0 217 0.836
7 1.05 2.60 30.0 35.0 487 0.479

corresponding attainable regions were chosen. The value, corre- y ¼ Cx þ Du; ð25Þ
sponding operating condition and exact position of these points where x is the model state vector, u is inputs and y is outputs. In
are indicated in Table 2 and Fig. 4. It should be noticed that point this work, the process model (2), (5)–(21) was linearized at
1 is the basic operating scenario. different operating points mentioned in Table 2 using the linear-
Assume the standard mathematical representation of a linear ization toolbox in Simulink, in which matrices A, B, C, D can be
time invariant (LTI) state-space model is: obtained. These matrices were then used to calculate the RGAs at
those operating points using the following equation:

dx
¼ Ax þ Bu; ð24Þ RGA ¼ ð  CA  1 B þ DÞðð  CA  1 B þ DÞ  1 ÞT ð26Þ
dt
 Y. Yang, Z.K. Nagy / Chemical Engineering Science 127 (2015) 362–373 367

Table 3 use of more advanced control approaches such as nonlinear model

RGAs and input–output pairings for selected operating points of the four proposed predictive control (NMPC) may be required.
control methods.

Control method Point RGA Pairing 3.4. Feasibility of nonlinear model predictive control (NMPC)
" #  
Nucleation control 1 λ11 λ13 0:733 0:267 y 1 3 u1 3.4.1. NMPC problem formulation
λ21 λ23 0:267 0:733 y 2 3 u3
Compared to decentralized PID control, NMPC is a more
  advanced control approach since it can handle complex and highly
2  0:028 1:028 y 1 3 u3
y 2 3 u1
nonlinear systems. NMPC approach uses the current measure-
1:028  0:028
ments and nonlinear model to calculate the predicted output
variables in a future time period named as prediction horizon.
" #   The future input variables in the prediction horizon are optimized
Growth control 1 λ12 λ14  0:265 1:265 y 1 3 u4
y 2 3 u2
to hold the output variables close to the set-point. Typically only
λ22 λ24 1:265  0:265
the first point of each optimal input variable is implemented. This
  optimization is then repeated when the next measurement is
3 2:734  1:734 y 1 3 u2
 1:734 2:734 y 2 3 u4 available. In this work, sampling time ðδÞ was chosen as 2 min.
Prediction horizon (P) or number of prediction intervals was
chosen as 25. Control horizon (M) or number of control moves
" #   was chosen as 5. In this work, open-loop optimal control solution
Antisolvent control 1 λ11 λ12  0:532 1:532 y 1 3 u2
λ21 λ22 1:532  0:532 y 2 3 u1 approach was used, and model mismatch was not considered. The
optimization problem can be formulated as:
 
4 0:865 0:135 y 1 3 u1
0:135 0:865 y 2 3 u2 min J ¼ 0:5ð1  wÞJ 1 þ 0:5ð1  wÞJ 2 þwJ 3 ; i ¼ 1; 2; 3; 4;
ui ðt k Þ;ui ðt k þ δÞ;:::;ui ðt k þ T p Þ

  ð29Þ
5 0:708 0:292 y 1 3 u1
0:292 0:708 y 2 3 u2 subject to:
model Eqs. (2), (5)–(21),
Z tk þ T p
" #  
Temperature control 1 λ13 λ14 0:026 0:974 y 1 3 u4 J1 ¼ ðy1  y1set Þ2 dt; ð30Þ
λ23 λ24 0:974 0:026 y 2 3 u3 tk

    Z
6 0:643 0:357 y 1 3 u3 y1set 2 t k þ T p
J2 ¼ ðy2  y2set Þ2 dt ð31Þ
0:357 0:643 y 2 3 u4 y2set tk

   
7 0:920 0:080 y 1 3 u3
y1set 2
0:080 0:920 y 2 3 u4 J3 ¼ traceðQ ΔuΔu0 Þ; ð32Þ
F0
2 3
u1 ðt k þ δÞ  u1 ðt k Þ; :::; u1 ðt k þM δÞ  u1 ðt k þ ðM 1ÞδÞ
Δu ¼ 6
4 ::: 7
5; ð33Þ
where A, B, C, D are matrices from LTI state-space model (24) and u4 ðt k þ δÞ  u4 ðt k Þ; :::; u4 ðt k þM δÞ  u4 ðt k þ ðM 1ÞδÞ
(25). For convenience, we write F 1 , F 2 , T 1 , T 2 , Ln , Y as u1 , u2 , u3 , u4 ,
2 3
y1 , y2 , respectively. Since there are only two inputs and two outputs, 1 0 0 0
all the RGAs are two by two matrices. Define the elements of RGA as: 60 1 0 0 7
6 7
Q ¼6 7; ð34Þ
" # 40 0 0:2 0 5
λmi λmj 0 0 0 0:2
RGA ¼
λni λnj ; ð27Þ

ui ðtÞ ¼ ui ðt k þ M δÞ; t ¼ t k þ ðM þ1Þδ; …; t k þ T p ; i ¼ 1; 2; 3; 4;

ð∂ym =∂ui Þu ð35Þ
λmi ¼ : ð28Þ
ð∂ym =∂ui Þy
0 r u1 ðtÞ; u2 ðtÞ r10; k ¼ t k ; t k þ δ; …; t k þ T p ; ð36Þ
RGA can be used as a tool for control pairing. The pairing which
has positive λij closest to one is the most desired. The final RGAs 10 r u3 ðtÞ; u4 ðtÞ r 40; k ¼ t k ; t k þδ; …; t k þ T p ; ð37Þ
and pairings obtained at those operating points are summarized in
Table 3. The results indicate that for a certain control method, the  δ r ui ðt þ δÞ  ui ðtÞ r δ; k ¼ t k ; t k þ δ; …; t k þ T p ; i ¼ 3; 4; ð38Þ
feasible pairing between inputs and outputs depends on the
position of the specific operating point. A change-over of the  δ r ui ðt k Þ  ui ðt k  1 Þ r δ; i ¼ 3; 4; ð39Þ
control pairing is necessary when the operating point is changed. where t k is the first time point of the k th optimization, T p is the
Thus decentralized PID control approach is not convenient to be time length of prediction horizon (50 min in our case), y1set and y2set
implemented when a relatively large operating region is required, are the set-points for crystal size and yield, respectively. Eqs. (30)
since suitable control would require change in input–output and (31) are cumulative errors of crystal size and yield compared to
topology and use a variable control structure. This is caused by their corresponding set-points. Eqs. (32)–(34) are penalty term for
the significant interactions among inputs and outputs and strong the inputs, which can help avoid sharp change in inputs and
nonlinearity of the process. These results demonstrate that con- promote smooth input responses and closed loop stability, Δu is a
ventional control of a cascade MSMPR may not be practically matrix that contains variations of each input, Q is a diagonal
feasible in none of the configurations when changing operating weighting matrix, as shown in Eq. (34) for the global control
conditions or product quality specifications are required and the method. In the case of the other four control methods, if ui is not
368 Y. Yang, Z.K. Nagy / Chemical Engineering Science 127 (2015) 362–373

Table 4
Selected set-points to evaluate NMPC feasibility.

Time [min] Set-point

Point Belonged attainable region Mean size [mm] Yield [dimensionless]

0–50 1 Nucleation control, growth control, antisolvent control, temperature control, global control 369 0.685
50–200 8 Nucleation control, growth control, antisolvent control, temperature control, global control 200 0.720
200–350 1 Nucleation control, growth control, antisolvent control, temperature control, global control 369 0.685
350–500 9 Antisolvent control, temperature control, global control 450 0.550

addition rate (F2) [ml/min] addition rate (F1) [ml/min]

500
Size [µm]

10

Stage 1 antisolvent
300

100 5
0 100 200 300 400 500

1 0
Yield [-]

0 100 200 300 400 500

0.75

0.5 Stage 2 antisolvent

0 100 200 300 400 500 10

1.5
5
C.V. [-]

0.5 0
0 100 200 300 400 500
Time [min] 0 100 200 300 400 500
Time [min]

5
Concentration [g/ml]

0.4 stage 1
4.5
stage 2
0.3
4
0.2
0.1 3.5
CPU time [min]

0 3
0 100 200 300 400 500
2.5
Supersaturation ratio[-]

1.8 2
stage 1
1.6 1.5
stage 2
1.4
1
1.2
0.5
1
0
0 100 200 300 400 500 0 100 200 300 400 500
Time [min] Time [min]
Fig. 6. NMPC performance of the antisolvent control method: (a) crystal size, yield and coefficient of variation (C.V.) of the CSD (red dotted lines are set-points);
(b) antisolvent addition rates at stage 1 and 2; (c) concentration and supersaturation ratio; and (d) CPU time. (For interpretation of the references to color in this figure
legend, the reader is referred to the web version of this article.)

used as manipulated variable, then Q ii becomes zero. To reduce the 3.4.2. Servo control
number of optimization variables while maintaining long enough A servo control scenario was designed to evaluate the feasibility
prediction horizon for stability, inputs after the control horizon (M) of NMPC using the five proposed control methods. As presented in
were set equal to the last values of inputs in the control horizon, as Table 4, three set-points (points 1, 8 and 9) are selected from the
shown in Eq. (35). Eqs. (36) and (37) are operating limits for the attainable region (Fig. 4) in order to evaluate NMPC performance
inputs. Eqs. (38) and (39) are the temperature rate limits that are set when fast change-over among these set-points are required. The
as 71 1C/min. w is the weight for penalty term, which is chosen as operating points 1, 8 and 9 correspond to three typical desired
0.05. As shown in the objective function (29), crystal size and yield product properties: medium size and medium yield, high yield and
were set of equal weights. Sequential quadratic programming (SQP) small size, and large size with low yield. It should be noticed that
technique was used to solve the optimization problem. points 1 and 8 belong to the attainable regions of all five control
 Y. Yang, Z.K. Nagy / Chemical Engineering Science 127 (2015) 362–373 369

500
Size [µm]
10

F [ml/min]
300

100 5
0 100 200 300 400 500

1
1 0
Yield [-]

0 100 200 300 400 500

0.75

0.5
0 100 200 300 400 500 40

30

T [°C]
1.5
C.V. [-]

1
1 20

0.5 10
0 100 200 300 400 500
Time [min] 0 100 200 300 400 500
Time [min]

500
Size [µm]

10

F [ml/min]
300

100 5
0 100 200 300 400 500
2

1 0
Yield [-]

0 100 200 300 400 500

0.75

0.5
0 100 200 300 400 500 40

30
T [°C]

1.5
C.V. [-]

1 20

0.5 10
0 100 200 300 400 500
Time [min] 0 100 200 300 400 500
Time [min]
Fig. 7. NMPC implementations: (a) controlled variables of nucleation control; (b) manipulated variables of nucleation control; (c) controlled variables of growth control; and
(d) manipulated variables of growth control (red dotted lines are set-points). C.V. is coefficient of variation of the CSD. (For interpretation of the references to color in this
figure legend, the reader is referred to the web version of this article.)

methods, whereas point 9 is outside of the attainable regions of
nucleation and growth control methods (Table 4). Points 1, 8 and
9 have similar positions as the operating points selected in the
analysis of decentralized PID control, making the evaluations of
decentralized PID control and NMPC comparable. The results indicate
that the NMPC performs well even in the case when decentralized
PID control would require change in the control architecture. During
time 0 to 50 min, the system was assumed to work at steady-state
with operating point 1 used as set-point. Then the set-point was
changed to point 8, point 1, point 9 at 50 min, 200 min, 350 min,
respectively, as indicated in Table 4.
The NMPC performance of the antisolvent control method is used
as an example for detailed analysis. The controlled variables, manipu-
lated variables, concentration and supersaturation ratio profiles, and
CPU time are shown in Fig. 6. For the antisolvent control method, the
NMPC scheme responds fast to the set-point change among points 1,
8, and 9, and finally maintained the system at steady-state at the new
set-points (Fig. 6(a)). Although coefficient of variation is not a
controlled variable, it is monitored here in order to guarantee that
after the crystal mean size reaches a steady-state value, the width of
size distribution is also in steady-state. Additionally, the manipulated
input responses are generally smooth (Fig. 6(b)). Fig. 6(c) indicates
that high supersaturation level is desired at the first stage when high
yield and small crystal size are the control objectives, whereas low
supersaturation levels at both stages are preferred when large crystal
size and low yield are desired. In general, for most optimizations the
CPU time was less than 2 min, which was smaller than the sampling
time (Fig. 6(d)). This indicates that the proposed framework is
promising for practical implementation. However it has to be men-
tioned that the solution approach for the NMPC optimization was
based on simple sequential solution technique implemented in
Matlab. The computational time can be reduced significantly using
more efficient optimization (e.g. multiple shooting) and/or other
software implementation such as OptCon (Mesbah et al., 2011, 2012;
Nagy et al., 2007; Simon et al., 2009).
The NMPC performances of the other four control methods are
shown in Figs. 7 and 8. Generally the NMPC works well for the
operating points that are within the attainable region of a particular
control approach. Since operating point 9 is outside of the attainable
region of the growth and nucleation control approaches these are
unable to bring the system to this operating point. The extra degrees
of freedom in the case of global control method provides faster
response however requires longer CPU time due to larger number of
optimization variables. The feasibilities of NMPC of all five proposed
methods are summarized in Table 5. According to Tables 4 and 5, the
necessary condition for a certain control method and NMPC scheme
to be feasible is that the set-point must be within the attainable
region of that control approach. The NMPC optimization problem will
370 Y. Yang, Z.K. Nagy / Chemical Engineering Science 127 (2015) 362–373

500
Size [µm]

40

T [ml/min]
300
30
100
0 100 200 300 400 500 20

1
1 10
Yield [-]

0 100 200 300 400 500

0.75

0.5
0 100 200 300 400 500 40

30

T [ °C]
1.5
C.V. [-]

2
1 20

0.5 10
0 100 200 300 400 500
Time [min] 0 100 200 300 400 500
Time [min]

F [ml/min] F1 [ml/min]
500 10
Size [µm]

5
300 0
0 100 200 300 400 500
100
0 100 200 300 400 500
10
5
1 0
Yield [-]

0 100 200 300 400 500

2

0.75
40
T [°C]

0.5
0 100 200 300 400 500 20
1

0 100 200 300 400 500

1.5 Time [min]
C.V. [-]

T [°C]

1 40
20
2

0.5 0 100 200 300 400 500

0 100 200 300 400 500
Time [min] Time [min]

Fig. 8. NMPC implementations: (a) controlled variables of temperature control; (b) manipulated variables of temperature control; (c) controlled variables of global control;
and (d) manipulated variables of global control (red dotted lines are set-points). C.V. is coefficient of variation of the CSD. (For interpretation of the references to color in this
figure legend, the reader is referred to the web version of this article.)

Table 5 1
Nucleation Control
NMPC performance using the five proposed control methods.
0.9 Growth Control
Time Set- NMPC is feasible (Y) or not (N) Antisolvent Control
0.8 Temperature Control
[min] point Control method
1 Global Control
0.7
Antisolvent Temperature Nucleation Growth Global
0.6
Yield [-]

0–50 1 Y Y Y Y Y
50–200 8 Y Y Y Y Y 0.5
200–350 1 Y Y Y Y Y
350–500 9 Y Y N N Y 0.4

0.3

0.2

be unfeasible if the set-point is out of the attainable region. Thus 0.1

knowing the attainable region of crystal size and yield is important
0
for choosing and designing a certain control approach. 100 200 300 400 500 600 700 800
Mean size [µm]

3.4.3. Regulatory control
Disturbances are unavoidable in actual processes. Therefore it is
important to know how disturbances can influence the attainable
regions of crystal size and yield. For example, when feed flow rate
changes from 5 ml/min to 6 ml/min, it is found that the basic
operating set-point (point 1) is no longer within the attainable region
of nucleation, antisolvent and temperature control methods, as shown
Fig. 9. Attainable region of crystal size and yield when feed flow rate increases
from 5 ml/min to 6 ml/min. Operating point 1 is the basic operating set-point.
in Fig. 9. This is because the attainable regions shift due to this
disturbance. As discussed in previous section, the necessary condition
for a certain control method and NMPC scheme to be feasible is that
the set-point must be within the attainable region of that control
 Y. Yang, Z.K. Nagy / Chemical Engineering Science 127 (2015) 362–373 371

approach. Therefore in principal these three methods can no longer
produce on-spec products if such a disturbance exists. Since the
attainable region of global control method is the largest, and also it
shifts the least, the global control method should be much more
robust than the other four methods in terms of disturbance rejection.
Thus regulatory control was considered in this section to evaluate
the disturbance rejection of the NMPC in the case of the global control
method. Two scenarios were investigated for disturbance rejection, one
with disturbance in the feed solution flow rate F 0 and another in the
concentration c0 , respectively (Table 6). Step disturbances of magnitude
between 10% and 20% were applied to the process after time 50 min.
The set-point was maintained at operating point 1 (Ln ¼
369 μm; Y ¼ 0:685) all the time for both scenarios. The disturbance
profile, controlled variables and manipulated variables for scenario
1 and 2 are presented in Fig. 10. It is clear that the crystal size and
yield can be well controlled at around the desired set-point, even
though large disturbances from feed solution are present. The coeffi-
cient of variation (or width of size distribution) can also be kept almost
constant under the simulated large disturbance. Thus the proposed
NMPC approach is robust in terms of disturbance rejection, if the set-
point is still within the attainable region when disturbances are present.
4. Conclusions

A continuous two stage mixed suspension mixed product removal

Table 6
(MSMPR) cascade crystallizer was modeled in this work using the
Two scenarios for disturbance rejection evaluation. F0, c0 are feed solution flow rate population balance model, with both cooling and antisolvent addi-
and solute concentration. tion applied to generate supersaturation. A multi-input multi-output
(MIMO) control system was investigated, with crystal size and yield
Time [min] Scenario 1 Scenario 2
targeted as controlled variables, and antisolvent addition rates and
F 0 [ml/min] c0 [g/ml] F 0 [ml/min] c0 [g/ml] temperatures at both stages used as manipulated variables. In
addition, based on different selections of manipulated variables, five
0–50 5.0 0.40 5.0 0.40 control methods (nucleation control, growth control, antisolvent
50–200 6.0 0.44 control, temperature control, and global control) were discussed.
200–350 5.0 0.40
350–500 4.0 0.36
The attainable regions of crystal size and yield of the proposed five
control methods were obtained. The feasibility of two advanced

b
F2 [ml/min] F1 [ml/min]

5
Size [µm] F0 [ml/min]

6
5
4 0
0 100 200 300 400 500 0 100 200 300 400 500

450 5
375
300 0
0 100 200 300 400 500 0 100 200 300 400 500
Yield [-]

0.75 35
T [°C]

0.675 25
1

0.6 15
0 100 200 300 400 500 0 100 200 300 400 500
Time [min]
C.V. [-]

1.5 35
T [°C]

1 25
2

0.5 15
0 100 200 300 400 500 0 100 200 300 400 500
Time [min] Time [min]
F2 [ml/min] F1 [ml/min]
C0 [g/ml]

0.45 5
0.4
0.35 0
0 100 200 300 400 500 0 100 200 300 400 500
Size [µm]

450 5
375
300 0
0 100 200 300 400 500 0 100 200 300 400 500
Yield [-]

0.75 35
T [°C]

0.675 25
1

0.6 15
0 100 200 300 400 500 0 100 200 300 400 500
Time [min]
C.V. [-]

1.5 35
T [°C]

1 25
2

0.5 15
0 100 200 300 400 500 0 100 200 300 400 500
Time [min] Time [min]
Fig. 10. Performance of the global control NMPC for disturbance rejection: (a) scenario 1, disturbance profile in feed flow rate (F0), crystal size, yield, and coefficient of
variation (C.V.) of CSD (red dotted lines are set-points); (b) scenario 1, manipulated variables: F1, F2, T1, T2 are antisolvent addition rates and temperatures at stages 1 and 2;
(c) scenario 2, disturbance profile in feed concentration (c0), crystal size, yield, and coefficient of variation (C.V.) of the CSD (red dotted lines are set-points); and (d) scenario
2 manipulated variables. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
372 Y. Yang, Z.K. Nagy / Chemical Engineering Science 127 (2015) 362–373

control approaches, decentralized PID control and nonlinear model ρc Crystal density [g/cm3]
predictive control (NMPC) were evaluated. Decentralized PID control τi Residence time at stage i [min]
is proved to require change of control structure when a relatively References
large operating region is essential based on local linearization
method and relative gain array (RGA) analysis, indicating that this Abu Bakar, M.R., Nagy, Z.K., Saleemi, A.N., Rielly, C.D., 2009. The impact of direct
process requires more advanced control approaches, such as NMPC, nucleation control on crystal size distribution in pharmaceutical crystallization
processes. Cryst. Growth Des. 9 (3), 1378–1384.
for smooth operation. It is found that the necessary condition for Acevedo, D.,, Nagy, Z.K., 2014. Systematic classification of unseeded batch crystal-
NMPC to be feasible is that the set-point must be within the lization systems for achievable shape and size analysis. J. Cryst. Growth 394,
corresponding attainable region of that particular control method. 97–105.
Alvarez, A.J., Singh, A., Myerson, A.S., 2011. Crystallization of cyclosporine in a
In general when this condition is met, NMPC scheme may be able to multistage continuous MSMPR crystallizer. Cryst. Growth Des. 11, 4392–4400.
show good control performance for fast target product qualities (size) Alvarez, A.J., Myerson, A.S., 2010. Continuous plug flow crystallization of pharma-
or process requirement (yield) change-over, as well as process ceutical compounds. Cryst. Growth Des. 10, 2219–2228.
Eder, R.J., Radl, S., Schmitt, E., Innerhofer, S., Maier, M., Gruber-Woelfler, H., Khinast,
disturbances rejection. It is also shown that the non-square global
J.G., 2010. Continuously seeded, continuously operated tubular crystallizer for
control approach with four manipulated inputs (temperatures and the production of active pharmaceutical ingredients. Cryst. Growth Des. 10,
flows in both stages) provides faster closed loop response and the 2247–2257.
largest attainable region. However, it should be noted that the Gron, H., Borissova, A., Roberts, K.J., 2003. ATR-FTIR spectroscopy for closed-loop
supersaturation control of a batch crystallizer producing mono- sodium
practical performance of the NMPC highly depends on the accuracy glutamate crystals of define size. Ind. Eng. Chem. Res. 42, 198–206.
of the model used in the controller. Another limitation of NMPC is Hermanto, M.W., Chiu, M., Braatz, R.D., 2009. Nonlinear model predictive control
related to sensor noise and error, which are not considered in this for the polymorphic transformation of L-Glutamic acid crystals. AICHE J. 55 (10),
2631–2645.
study either. But they can be quantified using experimental data in Kalbasenka, A.N., Spierings, L.C.P., Huesman, A.E.M., Kramer, H.J.M., 2007. Application of
future works. These uncertainties would lead to variations in the seeding as a process actuator in a model predictive control framework for fed-batch
boundaries of the attainable regions for control and could be crystallization of ammonium sulphate. Part. Part. Syst. Charact. 24, 40–48.
Lawton, S., Steele, G., Shering, P., Zhao, L., Laird, I., Ni, X.W., 2009. Continuous
incorporated in formulating the robust NMPC counterpart of the crystallization of pharmaceuticals using a continuous oscillatory baffled crystal-
control framework presented in this work. lizer. Org. Process Res. Dev. 13, 1357–1363.
Lindenberg, C., Krattli, M., Cornel, J., Mazzotti, M., 2009. Design and optimization of
a combined cooling/antisolvent crystallization process. Cryst. Growth Des. 9 (2),
1124–1136.
Nomenclature Mesbah, A., Huesman, A.E.M., Kramer, H.J.M., Nagy, Z.K., van den Hof, P.M.J., 2011.
Real-time control of seeded batch crystallization processes. AIChE J. 57 (6),
1557–1569.
m(i)
j jth moment at stage i [varies]
Mesbah, A., Nagy, Z.K., Huesman, A.E.M., Kramer, H.J.M., Van den Hof, P.M.J., 2012.
A, B, C, D Matrices in linear time invariant state-space model [varies] Real-time control of industrial batch crystallization processes using a popula-
Bi Nucleation rate at stage i [/min ml] tion balance modeling framework. IEEE Trans. Control Syst. Technol. 20 (5),
cn Solubility [g/ml] 1188–1201.
Nagy, Z.K., Aamir, E., 2012. Systematic design of supersaturation controlled crystal-
ci Concentration at stage i [g/ml] lization processes for shaping the crystal size distribution using an analytical
co Concentration of feed solution [g/ml] estimator. Chem. Eng. Sci. 84, 656–670.
DB, DG Dissolution rate [varies] Nagy, Z.K., Aamir, E., Rielly, C.D., 2011. Internal fines removal using population
balance model based control of crystal size distribution under dissolution,
F(0)
out Feed solution flow rate [ml/min] growth and nucleation mechanisms. Cryst. Growth Des. 11, 2205–2219.
Fi Antisolvent addition rate at stage i [ml/min] Nagy, Z.K., Fevotte, G., Kramer, H., Simon, L.L., 2013. Recent advances in the
F(i)
out Outlet flow rate at stage i [ml/min] monitoring, modelling and control of crystallization systems. Chem. Eng. Res.
Des. 91 (10), 1903–1922.
Gi Growth rate at stage i [cm/min] Nagy, Z.K., Mahn, B., Franke, R., Allgower, F., 2007. Efficient output feedback
J Objective [varies] nonlinear model predictive control for temperature control of industrial batch
ki Kinetics parameters [varies] reactors. Control Eng. Pract. 15, 839–859.
Nagy, Z.K., Braatz, R.D., 2003. Robust nonlinear model predictive control of batch
kv Shape factor [dimensionless] processes. AIChE J. 49 (7), 1776–1786.
L Crystal size [cm] Nagy, Z.K., Fujiwara, M., Braatz, R.D., 2008. Modelling and control of combined
Ln Number based mean crystal size [mm] cooling and antisolvent crystallization processes. J. Process Control 18,
856–864.
M Control horizon [dimensionless]
Qamar, S., Mukhtar, S., Seidel-Morgenstern, A., 2010. Efficient solution of a batch
ni Population number density at stage i [/ml] crystallization model with fines dissolution. J. Cryst. Growth 312 (20),
P Prediction horizon [dimensionless] 2936–2945.
Q Weighting matrix in penalty term [dimensionless] Quon, J.L., Zhang, H., Alvarez, A., Evans, J., Myerson, A.S., Trout, B.L., 2012.
Continuous crystallization of aliskiren hemifumarate. Cryst. Growth Des. 12,
S Supersaturation ratio [dimensionless] 3036–3044.
t Time [min] Ramkrishna, D., 2000. Population Balances: Theory and Applications to Particulate
Ti Temperature at stage i [1C] Systems in Engineering. Academic Press, San Diego.
Randolph, A.D., Larson, M.A., 1988. Theory of Particulate Process, 2nd edition
TP Length of prediction horizon [min] Academic Press, New York.
u Input variable [varies] Rawlings, J.B., Miller, S.M., Witkowski, W.R., 1993. Model identification and control
Vi Solution volume at stage i [ml] of solution crystallization processes: a review. Ind. Eng. Chem. Res. 32,
1275–1296.
w Weight for penalty term [dimensionless] Saleemi, A.N., Rielly, C.D., Nagy, Z.K., 2012. Automated direct nucleation control for
x State vector in state-space model [varies] in situ dynamic fines removal in batch cooling crystallization. CrystEngComm
Y Yield [dimensionless] 14, 2196–2203.
Sheikhzadeh, M., Trifkovic, M., Rohani, S., 2008. Adaptive MIMO neuro-fuzzy logic
y Output variable [varies] control of a seeded and an unseeded anti-solvent semi-batch crystallizer.
yset Set-point of output variable [varies] Chem. Eng. Sci. 63, 1261–1272.
αi Dilution factor at stage i [dimensionless] Simon, L.L., Nagy, Z.K., Hungerbuhler, K., 2009. Model based control of a liquid
δ Sampling time [min] swelling constrained batch reactor subject to recipe uncertainties. Chem. Eng. J.
153, 151–158.
λij i, jth element of relative gain array [dimensionless] Simon, L.L., Myerson, A.S., 2011. Continuous antisolvent plug-flow crystallization of
ν0 Solvent volumetric fraction in solvent mixture of feed a fast growing API. In: Proceedings of the 18th International Symposium on
solution [dimensionless] Industrial Crystallization (ISIC 18).
Simone, E., Saleemi, A.N., Nagy, Z.K., 2014a. Raman, UV, NIR, and mid-IR spectro-
νi Solvent volumetric fraction in solvent mixture at stage i scopy with focused beam reflectance measurement (FBRM) in monitoring
[dimensionless] polymorphic transformations. Chem. Eng. Technol. 37, 1305–1313.
 Y. Yang, Z.K. Nagy / Chemical Engineering Science 127 (2015) 362–373
Simone, E., Saleemi, A.N., Tonnon, N., Nagy, Z.K., 2014b. Active polymorphic
feedback control of crystallization processes using a combined Raman and
ATR-UV/Vis spectroscopy approach. Cryst. Growth Des. 14, 1839–1850.
Su, Q., Nagy, Z.K., Rielly, C.D., 2015. Pharmaceutical crystallisation processes from
batch to continuous operation using MSMPR stages: modelling, design and
control. Chem. Eng. Process. 89, 41–53.
Vetter, T., Burcham, C.L., Doherty, M.F., 2014. Regions of attainable particle size in
continuous and batch crystallization processes. Chem. Eng. Sci. 106, 167–180.
Wong, S.Y., Tatusko, A.P., Trout, B.L., Myerson, A.S., 2012. Development of contin-
uous crystallization processes using a single-stage mixed-suspension, mixed-
product removal crystallizer with recycle. Cryst. Growth Des. 12, 5701–5707.
373
Xie, W., Rohani, S., Phoenix, A., 2001. Dynamic modelling and operation of a seeded
batch cooling crystalliser. Chem. Eng. Commun. 187, 229–249.
Yang, Y., Nagy, Z.K., 2014. Model-based systematic design and analysis approach for
unseeded combined cooling and antisolvent crystallization (CCAC) systems.
Cryst. Growth Des. 14, 687–698.
Zhang, G.P., Rohani, S., 2003. On-line optimal control of a seeded batch cooling
crystallizer. Chem. Eng. Sci. 58, 1887–1896.
Zhang, H., Quon, J., Alvarez, A.J., Evans, J., Myerson, A.S., Trout, B., 2012. Develop-
ment of continuous anti-solvent/cooling crystallization process using cascaded
mixed suspension, mixed product removal crystallizers. Org. Process Res. Dev.
16, 915–924.