You are on page 1of 9

Article

pubs.acs.org/IECR

Batch-Wise Nonlinear Model Predictive Control of a Gas Antisolvent


Recrystallization Process for the Uniform Production of Micronized
HMX with Carbon Dioxide as the Antisolvent
Shin Je Lee, Sungho Kim, Bumjoon Seo, Youn-Woo Lee, and Jong Min Lee*
School of Chemical and Biological Engineering, Institute of Chemical Processes, Seoul National University, Seoul 08826, Republic of
Korea

ABSTRACT: Novel crystallization processes that use supercritical fluids have recently attracted considerable attention because
See https://pubs.acs.org/sharingguidelines for options on how to legitimately share published articles.
Downloaded via INDIAN INST OF TECH KHARAGPUR on August 28, 2021 at 17:37:22 (UTC).

they can overcome the problems associated with conventional crystallization processes. The gas antisolvent (GAS) process is one
of the promising techniques and has been applied to many applications. However, control of the GAS process is a quite
challenging problem and has not yet been studied due to the complex liquid−vapor equilibrium and particle formation kinetics.
This work proposes a batch-wise nonlinear model predictive control (BNMPC) approach to the GAS process to obtain the
desired particle size distribution (PSD) of HMX, a widely used explosive, which should be small and uniform for stability.
Although a dynamic model of the GAS crystallization is required for BNMPC, the previously developed model is too complex for
real-time applications. We propose a model simplification strategy for the conventional model using experimental data. We also
employ a high-resolution method (HRM) to solve effectively a partial differential equation (PDE). The simulation results show
that BNMPC can produce more uniform and smaller HMX particles.

1. INTRODUCTION
Crystallization plays a considerable role in producing various
chemical products, such as polymers, dyes, pharmaceuticals,
and explosives in fine-chemical industries. The industrial
applications of crystallization include semiconductor manufac-
turing, drug delivery, and the food industry. Recently, a number
of new applications have also involved crystallization processes,
such as the production of nano and amorphous materials.1
Over the past decades, crystallization techniques have been well
established in nearly all process industries as a method for
producing, purifying, and recovering solid materials, and these
techniques have witnessed major advancements in both
academic and industrial areas. Figure 1. (a) Schematic representation of the GAS process and (b)
However, there are several practical problems associated with experimental equipment.
the process, which are mostly attributed to the use of large
amounts of toxic solvents. The produced substances can be process starts. The antisolvent CO2 is pumped to high pressure
contaminated with the solvent, and waste solvent streams are over its critical point and injected into the vessel from the
inevitably generated in most crystallization processes. Using bottom to achieve better mixing of the solvent and antisolvent.
supercritical fluids as the solvent can overcome the drawbacks After the high-pressure CO2 is injected, the solution volume is
of the conventional processes because this approach uses expanded; however, the precipitator volume is maintained
environmentally benign solvents such as CO2. Moreover, the constant for the entire process. Acetone does not vaporize, and
unique physical properties of supercritical fluids make the thus, it does not affect the supersaturation or particle formation.
process easy to adjust and operable under mild conditions.2,3 In When the CO2 dissolves in acetone, the solvent strength on the
particular, the gas antisolvent (GAS) recrystallization process solute is reduced, the system becomes supersaturated and
has attracted considerable interest because of its ability to particles are produced. After a holding time, the solution is
produce fine particles. Two-way mass transfers of antisolvent drained under isobaric conditions to wash and collect the
and organic solvent facilitate uniform nucleation and almost particles.
instantaneous crystallization to yield ultrafine particles with a With increasing interest, many studies on the GAS process
narrow particle size distribution (PSD).4 A conceptual have been reported. Thermodynamic studies of phase behavior
representation of the GAS process is presented in Figure 1a,
and lab-scale GAS equipment is shown in Figure 1b. A Received: May 6, 2015
semibatch precipitator with constant volume, V, has one inlet to Revised: November 2, 2015
which the compressed CO2 gas is injected. A solution dissolved Accepted: November 5, 2015
with the solute is initially loaded in a precipitator before the Published: November 6, 2015

© 2015 American Chemical Society 11894 DOI: 10.1021/acs.iecr.5b01690


Ind. Eng. Chem. Res. 2015, 54, 11894−11902
Industrial & Engineering Chemistry Research Article

in the GAS process and their experimental results have been algebraic equations, and it is very complex to solve explicitly at
explored.5−7 Some studies have examined the accurate every sample time. Therefore, we propose a model
determination of the volume expansion in the GAS process simplification approach by simple algebraic manipulations and
because it is related to the product quality.4,8,9 Many using experimental data. Then, we use a high-resolution
experimental studies on selecting the optimal operating method (HRM) as the method for solving the PDE in the
conditions have been performed, and key parameters include GAS model because it is more stable and accurate for a steep
the antisolvent addition rate, temperature, organic solvent type, function than FDM or FEM.
and initial solution concentration.10−14 Systematic mathemat- The remainder of this paper is organized as follows. In
ical modeling of the GAS process was developed by Muhrer et Section 2, a fundamental model of the GAS process and its
al. 15 using a population balance model (PBM) with simplification approach using the experimental data are
thermodynamics and particle formation kinetics. The devel- presented, and the HRM discretization is also described. The
oped model has been validated with experimental data in the BNMPC algorithm is presented in Section 3, and the
literature,16,17 and a novel discretization method for solving the simulation results of HMX recrystallization are shown in
model has also been reported.17 Section 4, followed by concluding remarks in Section 5.
This work addresses the recrystallization of cyclotetramethy-
lenetetranitramine (HMX), a widely used explosive, using the 2. DYNAMIC MODEL OF THE GAS PROCESS AND
GAS process with acetone and CO2 as a solvent and MODEL SIMPLIFICATION
antisolvent, respectively. It has already been shown that A mathematical model for the GAS process was completely
HMX can be recrystallized into small and uniform particles developed by Muhrer et al.15 using a population balance model
using the GAS process in the literature.18,19 High-energy (PBM) combined with particle formation dynamics and
materials such as explosives must be small with a uniform PSD thermodynamics. After introducing this model, we propose a
for performance and stability. This emphasizes the need for simplification strategy of the model to be effectively applied to
controlling the GAS process. control schemes. Then, a high-resolution method (HRM) is
Although obtaining fine crystals with a uniform distribution is proposed to solve the PBM because the HRM exhibits superior
a critical design problem in the GAS process, studies on numerical accuracy compared to other well-known discretiza-
controlling the GAS process have hardly progressed compared tion methods such as FDM or FEM.
with the extensive studies for model development and various 2.1. Fundamental Model of the GAS Process. The PBM
applications. This is because controlling the PSD of the GAS is used to express how particle sizes are distributed as a partial
process is a challenging problem due to its inclusion of complex differential equation (PDE). The one independent variable is
vapor−liquid equilibrium and particle kinetics. Although there the time, t; the other variable is a property coordinate, such as
are a few studies on the control of liquid antisolvent the particle size, L. General crystallization dynamics are
crystallization,20−22 the liquid antisolvent process is quite described by28
different from the GAS process in that it does not include
vapor−liquid equilibrium in the model. ∂n(L , t ) ∂{G(L , t )n(L , t )}
+ = q(L , t , f )
This work proposes a nonlinear model predictive control ∂t ∂L (1)
(NMPC) approach for controlling the PSD of the GAS process where n(L,t) is the population density function representing a
because the system model is highly nonlinear. MPC particle size distribution (PSD); t and L are the time and the
applications for conventional crystallization have been widely size, respectively; G(L,t) is the growth rate of particles; and
studied and implemented in industries.23−25 In industrial q(L,t,f) is the creation/depletion rate, which accounts for
crystallization processes, the control problem has always been nucleation, aggregation, breakage, and material leaving or
an issue and has significantly advanced due to the importance entering the system.
of manufacturing crystals with consistent quality. These In the dynamic GAS model based on eq 1, several
advances have been enabled by the progress in in situ real- assumptions are made while retaining the basic dynamic
time sensor technologies.26 Robust performance can be behavior of the system:15 the temperature in the vessel is
achieved by repeating the optimization online on the basis of maintained constant such that no energy balance is needed; the
real-time measurements and state estimation.27 The nonlinear growth rate, G, is size independent; and aggregation and
GAS model is linearized every sample time in the control breakage contributions of particles are also neglected. These
algorithm, and the conventional MPC is implemented batch- assumptions simplify eq 1 to
wise because the GAS recrystallization is a batch process.
Simulation results show that the final product quality of HMX ∂n
+G
∂n
=0
is improved at the end of the final batch by a batch-wise ∂t ∂L (2)
nonlinear model predictive control (BNMPC) strategy. To the
The change of the PSD according to the volumetric expansion
authors’ best knowledge, there are no published results
of the liquid phase should be added to eq 2 because it
regarding the implementation of MPC to the GAS recrystalliza-
determines the supersaturation of the solute, which is the
tion, which is more challenging because of the complexity
driving force for particle formation
raised by the phase equilibrium, crystallization kinetics, and
calculation of a partial differential equation (PDE) describing ∂n ∂n n d(NLvL)
population balances in the model. +G + =0
∂t ∂L NLvL dt (3)
In the MPC algorithm, a dynamic process model of the
system is used to predict the effects of future actions on the where NL [mol] is the molar hold-up in the liquid phase and vL
outputs. Therefore, the process model is the essential element [m3/mol] is the molar volume of the liquid phase. Therefore,
of the MPC controller. However, the previously developed NLvL is the liquid volume, and d(NLvL)/dt represents the liquid
model of the GAS process consists of a PDE, ODEs, and volume change with time. The last term of eq 3 explains the
11895 DOI: 10.1021/acs.iecr.5b01690
Ind. Eng. Chem. Res. 2015, 54, 11894−11902
Industrial & Engineering Chemistry Research Article

effect of the liquid volume expansion on the PSD in the GAS well describes the vapor−liquid equilibrium of the solvent and
process. Finally, the PBM for the GAS process can be obtained antisolvent with a moderate complexity in the GAS process.
as the PDE. Twelve unknown variables (n, NL, NV, NP, vL, vV, xA, xS, xP,
The material balances on the antisolvent (A), solvent (S), yA, yS, and P) can be obtained by solving a system of partial
and particle (P) are differential, ordinary differential, and algebraic equations of eqs
3−11 and 13−15 with the following initial conditions
d(NLxA + NVyA )
= QA P = Patm (17)
dt (4)

d(NLxS + NVyS ) n(0, L) = 0 (18)


=0
dt (5)
xsNL + yS NV = NS0 (19)
d(NLx P + NP)
=0 x PNL = NP0
dt (6) (20)
where NV [mol] and NP [mol] are the molar hold-ups in the where N0S and N0P are the initial molar amounts of solvent and
vapor and solid phases, respectively; xi and yi are the mole solute, respectively.
fractions of component i in the liquid and vapor phases, The boundary condition for the PDE is given as
respectively (i = A, S, P); and QA [mol/s] is the molar flow rate
B
of the antisolvent and is the input variable. It is found that the n(t , 0) =
antisolvent addition rate, QA, has the strongest influence on the G (21)
final product,15 and it can be manipulated to control particle where B and G are the nucleation and growth rates,
size, PSD, and morphology.29 respectively, defined as
The molar hold-up in the solid phase, NP, is given by
⎧ B′ + B ′′ S > 1
NLvLk vm3 B=⎨
NP = ⎩0 S≤1 (22)
vP (7)
where kv, vP, and m3 are the volume shape factor, molar volume ⎧ kg(S − 1)g S > 1

of the solid phase, and third moment of the population density G=⎨ ⎪

function, respectively. The mole fractions of the three ⎩0 S≤1 (23)


components are subject to where B′ and B″ are the primary and secondary nucleation
xA + xS + x P = 1 (8) rates, respectively; S is the supersaturation level; and kg and g
are the rate constants. Constitutive equations of the primary
yA + yS = 1 (9) and secondary nucleation rates can be found in the
literature.15,30 The nucleation and growth rates are functions
and liquid−vapor equilibrium leads to the isofugacity relation-
of S, and the control of crystallization is to balance the
ships
nucleation and growth rates to adjust the particle size and
fA,L = fA,V (10) achieve the desired PSD.
The supersaturation, S, is defined by the ratio of fugacities of
fS,L = fS,V (11) the solid in the liquid and solid phases,

The fugacities in the liquid and vapor phases are calculated as fP,L
S=
fi , α = zi , αϕi , αP (i = A, S, α = L, V) fP,P (24)
(12)
where ϕ is the fugacity coefficient. The total volume, V, of the The fugacity of the solid at the solid phase, f P,P, is calculated
crystallizer is constant and satisfies using the Poynting correction factor
NLvL + NVv V = V (13) ⎡ v (P − P0) ⎤
fP,P = fP,L (P0 , T , x0) × exp⎢ P ⎥
where the liquid and vapor molar volumes, vL and vV [m3/mol], ⎣ RT ⎦ (25)
respectively, are calculated from an equation of state as where P0 and x0 indicate a reference pressure and composition
P = P(vL , T , x) (14) at the reference pressure, respectively.
2.2. Control-Relevant GAS Model. The comprehensive
P = P(v V , T , y) (15) GAS model is difficult to solve explicitly because complex
particle dynamics and thermodynamics are involved in the
The Peng−Robinson equation of state (PR-EOS) with system. We present a simplified model with some algebraic
quadratic mixing rules is used manipulations by using experimental data. The model should
RT aα be simplified to be solved within a sample time in the model-
P= − 2 (α = L, V) based control algorithm because solving the entire mathemat-
vα − bα vα + 2vαbα − bα2 (16)
ical model of GAS is very time-consuming.
where R is the gas constant, T is the temperature, vα is the After the model is simplified, it reduces to six equations and
molar volume of the α phase, and aα and bα are the parameters six unknown variables when the input is fixed.
of the PR-EOS of the α phase. The PR-EOS is used because it Summation of eqs 4−6 using eqs 8−9 provides
11896 DOI: 10.1021/acs.iecr.5b01690
Ind. Eng. Chem. Res. 2015, 54, 11894−11902
Industrial & Engineering Chemistry Research Article

dNL d NV dNP the PBM is particularly challenging because the population


+ + = QA density function, n(L,t), extends over orders of magnitude, and
dt dt dt (26)
thus, the distribution is very sharp. We use the HRM because it
and we specify three derivatives as f, g, and h as follows provides high-order accuracy with a lower computational load
dNL than other finite difference or finite element methods (FDM or
=f FEM) for sharp distributions.31,32 It does not create numerical
dt (27)
diffusion or dispersion as in first-order and second-order
d NV methods.31,32
=g The HRM with second-order accuracy discretizes eq 2 to32
dt (28)

dNP dni G
=h + (n L + − n Li−) = 0, i = 1, ..., N
(29) dt ΔL i (37)
dt
Then, eq 26 can be represented as where
f + g + h = QA (30) 1
n Li+ = ni + ϕ(θi+)(ni + 1 − ni)
2 (38)
Additionally, the derivatives of vL and vV are assigned to p and
q, respectively
1
dvL n Li− = ni − 1 + ϕ(θi−)(ni − ni − 1)
2 (39)
=p
dt (31)
and ΔL is the mesh element width; N is the divided mesh
dv V number; and ϕ(r) is the flux limiter function. It depends on the
=q degree of smoothness of the distribution, which is defined by
dt (32)
We can calculate p and q from the experimental results and the ni − ni − 1 + ϵ
θi+ =
fact that the liquid and vapor mole fractions, xi and yi, depend ni + 1 − ni + ϵ (40)
on the antisolvent addition rate, QA.5 It is assumed that the
pressure change is linear with the antisolvent addition rate, and ni − 1 − ni − 2 + ϵ
this assumption is found to be reasonable by solving the θi− =
ni − ni − 1 + ϵ (41)
equations describing the relationship between the pressure
change and addition rate. Then, the molar volumes of the liquid where ϵ = 10−10. Here, we choose the Van Leer flux limiter
and vapor phases, vL and vV, can be calculated from eq 16 with because it does not exhibit any numerical dispersion for one-
mole fractions, and the pressure and derivatives are also dimensional problems and provides full second-order accu-
obtained. racy32
The differentiation of eq 13 leads to
|r | + r
dvL dN dv dN ϕ(r ) =
NL + v L L + NV V + v V V = 0 1 + |r | (42)
dt dt dt dt (33)
and substitution of eqs 27, 28, 31, and 32 into eq 33 and Thus, eq 36 is changed into the discretized form of
rearrangement for f provides ∂ni G n
1 =− (n Li+ − n Li−) − (vLf + NLp)
f=− (NLp + NVq + v V (Q A − h)) ∂t ΔL NLvL (43)
vL − v V (34)
The dimension of the equation is dependent on the mesh size,
where h can be formulated from eq 7. The liquid phase change i.
part of eq 3 can be separately expressed as The HRM is compared with the first-order FDM for eq 2
∂n ∂n n ⎛ dNL dv ⎞ with the size-independent growth rate (G = 1.0 μm/s), and the
+G + ⎜v L + NL L ⎟ = 0 boundary condition is given as n(0,t) = 0. The initial
∂t ∂L NLvL ⎝ dt dt ⎠ (35) distribution is33
and substitution of eqs 27 and 31 into eq 35 provides ⎧1 × 1010 if 10 μm < L < 20 μm

∂n ∂n n n(L , 0) = n0(L) = ⎨
⎩0

= −G − (vLf + NLp) otherwise
∂t ∂L NLvL (36)
(44)
Finally, the simplified model consists of eqs 27−29, 31, 32, and
The analytical solution of this problem with an initial profile is
36 with six unknown variables of NL, NV, NP, vL, vV, and n. The
the initial profile translated by a length Gt, that is,
initial and boundary conditions are the same as those of the
fundamental model. Then, the partial derivative of n with n(L , t ) = n0(L − Gt ) (45)
respect to L in eq 36 should be discretized to solve the PDE,
and we apply a high-resolution method (HRM) to the PDE in The population densities for three solution approaches are
the next subsection. compared in Figure 2. The HRM exhibits better accuracy than
2.3. HRM Solution Strategy. A specific numerical solution the FDM, and it does not considerably change as the simulation
scheme is required to solve numerically the GAS model because time increases, whereas the accuracy of FDM decreases as time
it includes the PDE in the PBM. The numerical simulation of increases due to the numerical dispersion.
11897 DOI: 10.1021/acs.iecr.5b01690
Ind. Eng. Chem. Res. 2015, 54, 11894−11902
Industrial & Engineering Chemistry Research Article

which can be used to compute the control actions as in the


usual linear MPC formulations, is developed using the
following equations.34 For a prediction horizon p,
Ykj + 1 | k = Y k0,+j 1 | k + Skj ΔUkj , j = 1, ..., p (54)
where
⎡yj ⎤ ⎡ Δu j = u j − u j ⎤
⎢ k+1|k ⎥ ⎢ k|k k|k k−1|k ⎥
⎢ j ⎥ ⎢ Δu j
y = ukj+ 1 | k − ukj| k ⎥
Ykj + 1 | k = ⎢ k + 2 | k ⎥, ΔUkj = ⎢ k + 1 | k ⎥
⎢ ⋮ ⎥ ⎢ ⋮ ⎥
Figure 2. Comparison of HRM and FDM for the solution of eq 45 at ⎢ ⎥ ⎢ ⎥
(a) 30 s and (b) 60 s. ⎢yj ⎥ ⎢⎣ Δukj+ m − 1 | k ⎥⎦
⎣ k+p|k ⎦
(55)
3. BATCH-WISE NONLINEAR MODEL PREDICTIVE
and Sjk is the matrix of step response coefficients for the linear
CONTROL (BNMPC) system, which may be calculated as described in the literature.35
3.1. Nonlinear Process Model and Linearization. The ykj + l|k is the predicted value of y at time k + l based on the
nonlinear process model has the following representation information available at time k. The vector Y0,jk+1|k provides the
xk + 1 = f (xk , uk) + wk (46)
output trajectories if no control actions were taken.
3.3. Calculation of Control Action. At time k, the control
yk = g (xk , uk) + vk (47) action to be solved online at every sampling instance in the
NMPC algorithm is
where xk, uk, and yk are the vectors of nx system states, nu inputs,
and ny measured variables at the kth sampling instance and wk minΔU{||Q (R kj + 1 | k − Ykj + 1 | k)||22 + ||R ΔUkj||22 } (56)
and vk are the zero-mean white noises on the system states and subject to
measured variables. The system dynamics and the measure-
ments are described by the nonlinear functions f and g. The j
umin ≤ ukj+ l | k ≤ umax
j
, l = 0, ..., m − 1
states, inputs, and outputs of the GAS recrystallization process
are Δukj+ l | k ≤ Δumax
j
, l = 0, ..., m − 1
x = [ni , NL , NV , NP , vL , v V ]T , i = 1, ..., M (48) j
ymin ≤ ykj+ n | k ≤ ymax
j
, n = 1, ..., p (57)
u = QA (49) where Rjk+1|kis the vector of future set points corresponding to
M Yjk+1|k in the jth batch, Q and R are the weighting matrices, and
umin, umax , ymin, and ymax are the minimum and maximum values
y= ∑ niΔL
i=1 (50) of the input and output variables, respectively. User-selected
tuning parameters are the prediction horizon (p), control
where M is the mesh size of the particle density function, n. horizon (m)(m ≤ p), and weighting matrices in the objective
Therefore, the dimension of the system is (M+5) . We set M as function (Q and R). The above constrained optimization
100 and ΔL as 1 μm. The measured output, y, represents the problem can be expressed as a quadratic program (QP), which
total number of particles with particle sizes from 0 to 100 μm. is solved online at every sampling time. The first element of
The calculation of the total number of particles is from the ΔU is implemented on the plant, and the optimization problem
moment of a distribution. is reformulated and solved at the next time step iteratively
At time k, the nonlinear functions f and g in xk and uk are utilizing the most current measurements.
linearized with respect to xk and uk−1 using a Jacobian matrix at 3.4. Batch Input Update. After the jth batch process is
every sample time. finished, the final input values become the initial input values in
xk̅ + 1 = Ak xk̅ + Bk uk̅ (51) the next (j+1)th batch as follows
yk̅ = Ckxk̅ (52) u0j + 1 = uNj , j = 1, ..., J (58)

where where J is the total batch number. Therefore, the PSD quality
approaches the set point trajectory as the number of batch
∂f ∂f ∂g processes increases.
Ak = |x , u , Bk = |x , u , Ck = |x , u
∂x k k −1 ∂u k k −1 ∂x k k −1
(53) 4. RESULTS AND DISCUSSION
and x¯k = xk − xs, u¯k = uk − us, and y¯k = yk − ys. After 4.1. Open-Loop Simulation and Experimental Vali-
linearization of nonlinear functions, the conventional MPC dation. The dynamic model of the GAS process is simulated
algorithm can be applied to the linearized state-space model. for 100 s with constant CO2 addition rates of 20, 50, and 100
3.2. Linear Prediction of Future Outputs. Because the mL/min to determine the effect of the addition rate on the final
system is a semibatch process, the MPC algorithm is PSD. The temperature is assumed to be fixed at 30 °C because
implemented within each batch. This is possible by using the the GAS process is an isothermal process. The particle size, L, is
real-time measurements of PSD using ATR-FTIR or ATR-UV/ discretized from 0 to 100 μm with a mesh size of 1 μm, and
V.26 In the jth batch, a linear prediction of future outputs, thus, the total number of meshes is 101. Calculations are
11898 DOI: 10.1021/acs.iecr.5b01690
Ind. Eng. Chem. Res. 2015, 54, 11894−11902
Industrial & Engineering Chemistry Research Article

performed using MATLAB R2014a with an Intel(R) Core-


(TM) i5-3570 CPU @ 3.40 GHz (4 CPUs). Figure 3 shows the

Figure 3. Final PSDs of HMX at 20, 50, and 100 mL/min at 30 °C.

open-loop simulation results of the final PSDs when the CO2


addition rates are 20, 50, and 100 mL/min. It is observed that
the PSD becomes narrower and that the average particle size
decreases as the CO2 addition rate increases. This result is
because the fast CO2 addition rate induces a high super-
saturation level in a short time, leading to a sudden burst of
nucleation while simultaneously preventing the particle
growth.19 In other words, the fast CO2 addition rate increases
the nucleation-to-growth ratio, and this ratio varies with the
antisolvent addition rates. Therefore, with the rapid CO2
addition rate, a large amount of small particles that do not
achieve full growth are formed at once.
The mean particle sizes and variances of the PSDs are also Figure 4. (a−d) FESEM of HMX at 25, 30, 40, and 50 °C and CO2
provided in Table 1 to compare quantitatively the results. From addition rate of 10 mL/min and (e−h) FESEM of HMX at 25, 30, 40,
the variances, we can observe that the PSD becomes narrower and 50 °C and CO2 addition rate of 50 mL/min.
and has uniform particles as the CO2 addition rate increases.

Table 1. Mean Particle Sizes and Variances of the Final PSDs addition rate is 50 mL/min, the temperature has a strong
CO2 addition rate (mL/min) mean size (μm) variance influence on the particle size, as shown in Figure 5b, whereas
under 10 mL/min, the particle size does not considerably
20 44.0 501
change, as shown in Figure 5a. Moreover, when the operating
50 38.9 342
temperature is greater than 40 °C, the mean particle size does
100 28.5 106
not considerably decrease, even when the CO2 addition rate
increases from 10 to 50 mL/min.
The simulation results are also compared with the From the open-loop test and the experimental results, it is
experimental data to show that the simplification of the observed that uniform particles with improved quality can be
dynamic model of GAS is acceptable. In the experiments, the obtained by controlling the process.
CO2 addition rate is set to 10 or 50 mL/min at constant 4.2. Simulation Results of BNMPC. The proposed
temperatures of 25, 30, 40, and 50 °C and γ-butyrolactone BNMPC is applied to the GAS crystallization process to
(GBL) is used as a solvent. Figure 4 shows the FESEM (field- obtain the desired PSD of HMX. The tuning parameters and
emission scanning electron microscopy) results of HMX. From process constraints for the BNMPC are presented in Table 2.
Figure 4, we can observe that the real shape of HMX is not a The minimum input value is selected as 0.0134 mL/min such
circle, although we assume that the shape of HMX is circular in that the bimodal PSD is avoided, and the maximum is an
the model, and small and uniform particles can be obtained arbitrary realistic value. The simulation time and sampling time
under 25 °C and 50 mL/min, as shown in Figure 4e. In Figure are 100 and 0.5 s, respectively. Figure 6 shows the mean particle
5, the comparison results of mean particle sizes between the sizes and input profiles based on the results of BNMPC of the
experiment and the model are shown. The observation that the GAS process during 5 batches. In the first batch, the objective is
operating condition of 25 °C and 50 mL/min produces a to maximize the particle number as close to 109 as possible to
smaller and more uniform particle can be quantitatively make the particles small because the particle size would be
confirmed in Figure 5. Under this condition, the mean particle smaller if a large number of particles are formed. This strategy
size is 20 μm. We can also determine the effect of temperature is based on the fact that we can only measure the total particle
on the particle size from the experiment. When the CO2 number in the sample time.
11899 DOI: 10.1021/acs.iecr.5b01690
Ind. Eng. Chem. Res. 2015, 54, 11894−11902
Industrial & Engineering Chemistry Research Article

batch and so forth. The input profiles during each batch are
slightly changed, and the final mean particle size is reduced
through 5 batches.
As shown in Figure 6, the input profiles suddenly increase for
the first 3 batches. This result is because the particle becomes
larger with time and the CO2 addition rate should increase to
control the particle size to be as small as possible. Under the
initial conditions, the particle size is very small, and as the
particle grows, the mean particle size gradually increases. After
the particle grows to some extent, the input rate suddenly
increases in the middle of the process to control the particle
size as small. After 3 batches, the CO2 addition rate does not
considerably increase because several initial batches have the
greatest effect on the PSD.
The final particle sizes are quantitatively compared in Figure
7, along with the initial input values for 5 batches. The final

Figure 5. Mean particle size comparison of experimental data and


simulation results at CO2 addition rates of (a) 10 and (b) 50 mL/min.

Table 2. Tuning Parameters and Constraints for the


BNMPC
parameter value
p 5
Figure 7. Comparison of final mean particle sizes and initial input
values for 5 batches.
m 5
Q 0.3
R 0.2 mean particle sizes for 5 batches are 33.6, 31.0, 24.7, 18.8, and
umin 0.0134 mL/min 19.5 μm, and the mean size no longer decreases after the fourth
umax 1340 mL/min batch. Additionally, the initial input values are 10.01, 0.0697,
Δumax 0.267 mL/min 0.0290, 0.0211, and 0.0194 mL/min.
The final PSDs at 5 batches are shown in Figure 8 to show
the uniformity of the PSD of HMX in the fifth batch. From
The initial CO2 addition rate of the first batch is 10 mL/min, Figure 8, it is observed that the final PSD at the fifth batch is
and it decreases to 0.0697 mL/min at the end of the batch. The the most uniform, while its mean size, 19.5 μm, is somewhat
final input value is maintained as the initial value of the second larger than 18.8 μm, the mean size of the fourth batch. In

Figure 6. Mean particle sizes and input profiles obtained from BNMPC for 5 batches.

11900 DOI: 10.1021/acs.iecr.5b01690


Ind. Eng. Chem. Res. 2015, 54, 11894−11902
Industrial & Engineering Chemistry Research Article

considerable influence on the final PSD. The average particle


sizes of Seeds 1, 2, and 3 are 5, 10, and 20 μm, respectively, and
they are changed to 38.4, 46.1, and 53.9 μm at the end of the
batch through the GAS process.

5. CONCLUDING REMARKS
Control issues of the GAS process are quite challenging
because the system is highly nonlinear and includes complex
liquid−vapor equilibrium and particle kinetics, and it has not
yet been studied. In this work, we propose a batch-wise
nonlinear model predictive control (BNMPC) to control the
GAS recrystallization process to make HMX, which is widely
used as an explosive, small and uniform for stability. To
implement the MPC, a dynamic model of the GAS process is
Figure 8. Final PSDs of HMX for 5 batches. required.
The dynamic GAS model can be represented as a PBM and
conclusion, fine and uniform HMX particles can be obtained
includes a PDE that requires a discretization method to be
from the BNMPC algorithm in the GAS crystallization process.
We also observe that the PSD of the fifth batch with an input numerically solved. HRM is introduced because it provides
value of approximately 25 mL/min is considerably smaller than high-order accuracy. Additionally, in the GAS model, several
the PSD with an input value of 100 mL/min in the open-loop constitutive equations are also included: thermodynamics
simulation. This result is because the open-loop process is just a related to liquid volume expansion and nucleation and growth
one-batch process with an input value of 100 mL/min, whereas rates of particle kinetics. The full dynamic model of the GAS
the closed-loop process goes through five batch processes, each process is difficult to solve in the sampling time. Therefore, we
with a specific input profile. Repeating 5 batches should be propose a model simplification strategy using experimental data
more effective than simply one batch even though it is with the and calculation of the PR-EOS. The particle size calculated
maximum input rate, 100 mL/min. If we compare the PSD with
from the simplified model shows insignificant error with the
100 mL/min in Figure 3 and the PSD of the first batch in
Figure 8, then the 100 mL/min clearly provides substantially experimental results. Through an open-loop simulation and the
smaller particles than the input profile of the first batch in the experimental results, we can observe that the particle size of
closed-loop simulation. However, through five batch processes, HMX decreases and that its PSD is more uniform as the CO2
the PSDs are effectively controlled to provide an optimal PSD, addition rate increases.
as shown in Figure 8. After several batches, the particle sizes The BNMPC algorithm is presented to control the GAS
become smaller, and they are used as the initial particles in the process. Within a batch, the conventional MPC is implemented,
next batch. Although they are dissolved in the solvent, it affects and the final input value is transferred to the initial input of the
the initial PSD of HMX, as presented by Kim et al.18 next batch. The BNMPC results show that the final mean
In other words, the BNMPC control results are affected by
particle size of the HMX decreases after 5 batches and that the
the CO2 feed rate and by the initial seed PSD in a batch. To
determine the effect of the initial seed PSD on the final PSD, final PSD at the fifth batch is the most uniform.
three HMX seed PSDs with different shapes are used while the In this work, we apply the MPC algorithm to the GAS
CO2 addition rate remains at a constant value of 50 mL/min. crystallization process as the first step of model-based control of
The PSD variation at the end of the batch is shown in Figure 9. this system. A batch-to-batch control strategy can also be
It is observed that the shape of the initial PSD has a introduced to the GAS process in future research.

Figure 9. Final PSDs of HMX for 5 batches.

11901 DOI: 10.1021/acs.iecr.5b01690


Ind. Eng. Chem. Res. 2015, 54, 11894−11902
Industrial & Engineering Chemistry Research Article

■ AUTHOR INFORMATION
Corresponding Author
dipropionate using carbon dioxide. Ind. Eng. Chem. Res. 2007, 46,
8009−8017.
(17) Bakhbakhi, Y. A discretized population balance for particle
*J. M. Lee. Tel.: +82 2 8801878. Fax: +82 2 8887295. E-mail: formation from gas antisolvent process: The combined Lax-Wendroff
jongmin@snu.ac.kr. and Crank-Nicholson method. Comput. Chem. Eng. 2009, 33, 1132−
Notes 1140.
(18) Kim, S.-J.; Lee, B.-M.; Lee, B.-C.; Kim, H.-S.; Kim, H.; Lee, Y.-
The authors declare no competing financial interest.


W. Recrystallization of cyclotetramethylenetetranitramine (HMX)
using gas anti-solvent (GAS) process. J. Supercrit. Fluids 2011, 59,
ACKNOWLEDGMENTS 108−116.
The Energy Efficiency & Resources Core Technology Program (19) Lee, B.-M.; Kim, S.-J.; Lee, B.-C.; Kim, H.-S.; Kim, H.; Lee, Y.-
of the Korea Institute of Energy Technology Evaluation and W. Preparation of micronized β-HMX using supercritical carbon
dioxide as antisolvent. Ind. Eng. Chem. Res. 2011, 50, 9107−9115.
Planning (KETEP) granted financial resources from the (20) Sheikhzadeh, M.; Trifkovic, M.; Rohani, S. Real-time optimal
Ministry of Trade, Industry & Energy, Republic of Korea control of an anti-solvent isothermal semi-batch crystallization process.
(2012T100201687). This work is supported by the Korea Chem. Eng. Sci. 2008, 63, 829−839.
Ministry of Environment as a Project for Developing Eco- (21) Nagy, Z. K.; Fujiwara, M.; Braatz, R. D. Modelling and control
Innovation Technologies (GT-11-G-02-001-3). of combined cooling and antisolvent crystallization processes. J. Process


Control 2008, 18, 856−864.
REFERENCES (22) Nowee, S. M.; Abbas, A.; Romagnoli, J. A. Model-based optimal
strategies for controlling particle size in antisolvent crystallization
(1) Larsen, P.; Patience, D.; Rawlings, J. Industrial crystallization operations. Cryst. Growth Des. 2008, 8, 2698−2706.
process control. IEEE Control Syst. Mag. 2006, 26, 70−80. (23) Rohani, S.; Haeri, M.; Wood, H. Modeling and control of a
(2) Jung, J.; Perrut, M. Particle design using supercritical fluids: continuous crystallization process Part 2. Model predictive control.
literature and patent survey. J. Supercrit. Fluids 2001, 20, 179−219. Comput. Chem. Eng. 1999, 23, 279−286.
(3) Knez, Z.; Weidner, E. Particles formation and particle design (24) Dokucu, M. T.; Park, M.-J.; Doyle, F. J., III Multi-rate model
using supercritical fluids. Curr. Opin. Solid State Mater. Sci. 2003, 7, predictive control of particle size distribution in a semibatch emulsion
353−361. copolymerization reactor. J. Process Control 2008, 18, 105−120.
(4) Mukhopadhyay, M. Partial molar volume reduction of solvent for (25) Hermanto, M. W.; Chiu, M.-S.; Braatz, R. D. Nonlinear model
solute crystallization using carbon dioxide as antisolvent. J. Supercrit. predictive control for the polymorphic transformation of L-glutamic
Fluids 2003, 25, 213−223. acid crystals. AIChE J. 2009, 55, 2631−2645.
(5) Dixon, D. J.; Johnston, K. P. Molecular thermodynamics of (26) Nagy, Z.; Braatz, R. Advances and New Directions in
solubilities in gas antisolvent crystallization. AIChE J. 1991, 37, 1441− Crystallization Control. Annu. Rev. Chem. Biomol. Eng. 2012, 3, 55−75.
1449. (27) Larsen, P.; Rawlings, J.; Ferrier, N. An algorithm for analyzing
(6) Shariati, A.; Peters, C. Measurements and modeling of the phase noisy, in situ images of high-aspect-ratio crystals to monitor particle
behavior of ternary systems of interest for the GAS process: I. The size distribution. Chem. Eng. Sci. 2006, 61, 5236−5248.
system carbon dioxide+ 1-propanol+ salicylic acid. J. Supercrit. Fluids (28) Ramkrishna, D. Population balances: Theory and applications to
2002, 23, 195−208. particulate systems in engineering; Academic Press: London, 2000.
(7) de la Fuente, J. C.; Shariati, A.; Peters, C. J. On the selection of (29) Gallagher, P. M.; Coffey, M.; Krukonis, V.; Hillstrom, W. Gas
optimum thermodynamic conditions for the GAS process. J. Supercrit. anti-solvent recrystallization of RDX: formation of ultra-fine particles
Fluids 2004, 32, 55−61. of a difficult-to-comminute explosive. J. Supercrit. Fluids 1992, 5, 130−
(8) De la Fuente Badilla, J.; Peters, C.; de Swaan Arons, J. Volume 142.
expansion in relation to the gas-antisolvent process. J. Supercrit. Fluids (30) Mersmann, A. Crystallization technology handbook. Drying
2000, 17, 13−23. Technol. 1995, 13, 1037−1038.
(9) Elvassore, N.; Bertucco, A.; Di Noto, V. On-line monitoring of (31) Gunawan, R.; Fusman, I.; Braatz, R. High resolution algorithms
volume expansion in gas-antisolvent processes by UV-vis spectroscopy. for multidimensional population balance equations. AIChE J. 2004, 50,
J. Chem. Eng. Data 2002, 47, 223−227. 2738−2749.
(10) Müller, M.; Meier, U.; Kessler, A.; Mazzotti, M. Experimental (32) Qamar, S.; Elsner, M.; Angelov, I.; Warnecke, G.; Seidel-
study of the effect of process parameters in the recrystallization of an Morgenstern, A. A comparative study of high resolution schemes for
organic compound using compressed carbon dioxide as antisolvent. solving population balances in crystallization. Comput. Chem. Eng.
Ind. Eng. Chem. Res. 2000, 39, 2260−2268. 2006, 30, 1119−1131.
(11) Muhrer, G.; Mazzotti, M.; Müller, M. Gas antisolvent (33) Qamar, S.; Ashfaq, A.; Warnecke, G.; Angelov, I.; Elsner, M.;
recrystallization of an organic compound Tailoring product PSD and Seidel-Morgenstern, A. Adaptive high-resolution schemes for multi-
scaling-up. J. Supercrit. Fluids 2003, 27, 195−203. dimensional population balances in crystallization processes. Comput.
(12) Kim, C.-K.; Lee, B.-C.; Lee, Y.-W.; Kim, H. S. Solvent effect on Chem. Eng. 2007, 31, 1296−1311.
particle morphology in recrystallization of HMX (cyclotetramethyle- (34) Ricker, N.; Lee, J. Nonlinear model predictive control of the
netetranitramine) using supercritical carbon dioxide as antisolvent. Tennessee Eastman challenge process. Comput. Chem. Eng. 1995, 19,
Korean J. Chem. Eng. 2009, 26, 1125−1129. 961−981.
(13) Lee, B.-M.; Jeong, J.-S.; Lee, Y.-H.; Lee, B.-C.; Kim, H.-S.; Kim, (35) Garcia, C.; Prett, D.; Morari, M. Model predictive control:
H.; Lee, Y.-W. Supercritical antisolvent micronization of cyclo- theory and practice-a survey. Automatica 1989, 25, 335−348.
trimethylenetrinitramin: Influence of the organic solvent. Ind. Eng.
Chem. Res. 2009, 48, 11162−11167.
(14) Widjojokusumo, E.; Veriansyah, B.; Youn, Y.-S.; Lee, Y.-W.;
Tjandrawinata, R. R. Co-precipitation of loperamide hydrochloride
and polyethylene glycol using aerosol solvent extraction system.
Korean J. Chem. Eng. 2013, 30, 1797−1803.
(15) Muhrer, G.; Lin, C.; Mazzotti, M. Modeling the gas antisolvent
recrystallization process. Ind. Eng. Chem. Res. 2002, 41, 3566−3579.
(16) Dodds, S.; Wood, J. A.; Charpentier, P. A. Modeling of the gas-
antisolvent (GAS) process for crystallization of beclomethasone

11902 DOI: 10.1021/acs.iecr.5b01690


Ind. Eng. Chem. Res. 2015, 54, 11894−11902

You might also like