Received 10 July 2015; revised 19 September 2015; accepted 21 December 2016.

Date of publication 31 January 2017; date of current version 23 May 2017.

Digital Object Identifier 10.1109/LLS.2017.2661981

Feedback Control of Plug-Flow

ON–OFF
Crystallization: A Case of Quality-by-Control
in Continuous Manufacturing
ANDY KOSWARA AND ZOLTAN K. NAGY, (Member, IEEE)
School of Chemical Engineering, Purdue University, West Lafayette, IN 47907 USA
CORRESPONDING AUTHOR: Z. K. NAGY (zknagy@purdue.edu)
This work was supported in part by the FDA Grant to NIPTE titled ‘‘The Critical Path Manufacturing Sector Research Initiative (U01)’’ under Grant
5U01FD004275 and in part by the Department of Defense: Defense Advanced Research Projects Agency (Award no. W911NF-16-2-0020).

ABSTRACT Plug-flow crystallization (PFC) is a promising candidate to realizing the paradigm shift
from batch-to-continuous pharmaceutical manufacturing. While PFC has been recently touted as the ideal
continuous crystallizer due to its compact design, proper mixing, and flexible cooling and antisolvent control,
it is prone to surface fouling or encrustation. In this letter, a model of encrustation growth and dissolution
dynamics coupled with PFC is discussed and a novel method of ON–OFF feedback control of PFC with
antifouling control is proposed. The study illustrates a quintessential example of quality-by-control concept,
which is complementary to the quality-by-design and essential in ensuring desired control performance and
product quality.

INDEX TERMS Dynamic optimization, feedback control, fouling, nonlinear systems, plug-flow
crystallization (PFC), spatio-temporal control.

I. INTRODUCTION
During plug-flow crystallization (PFC) of an active pharma-
ceutical ingredient (API), supersaturated API mixture nucle-
ates and sediments onto the reactor’s surface [1], [2]. Over
time, this encrust layer grows, consequently increasing the
thermal resistance between the tube and reactor wall as well
as reducing the product residence time (RT). This in turn
leads to diminishing crystal yield and quality and, in the
extreme case, stops the manufacturing process by blockage.
FIGURE 1. Domains of encrustation-PBE-PFC dynamical model.
Furthermore, encrustation constrains the attainable temper-
ature profile within the tube leading to variability in the
output crystal size distribution (CSD). This is the most sig-
nificant bottleneck in an encrustation-dominated PFC, since and population balances. The effect of encrustation in an
product consistency is an essential requirement of a contin- optimized PFC without implementation of AFC is described
uous pharmaceutical process. In this letter, an antifouling in Section III. This serves as the motivation for the utiliza-
control (AFC) design based on heating and cooling cycle tion of ON–OFF feedback controllers based on CSD and the
is presented, whereby crystal and encrust growth are peri- presence of encrust coupled with AFC implementation. This
odically alternated with crystal and encrust dissolution. The letter is concluded with a summary and discussion on future
design leads to a concept of quality-by-control (QbC), in directions in Section IV.
which ON–OFF feedback controllers are utilized to ensure
the consistency of the process output despite the unpre- II. ENCRUSTATION DYNAMICS
dictability of the response to fouling. This letter is divided A model for encrustation in a PFC inspired from the fouling
as follows. Section II describes a dynamic model for the studies of CaSO4 crystals commonly found in heat exchang-
encrustation-crystallization process, including mass, energy, ers has been described (Fig. 1) [3]. The encrustation dynamics

2332-7685
2017 IEEE. Translations and content mining are permitted for academic research only.
Personal use is also permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
VOLUME 3, NO. 1, MARCH 2017 1
 Koswara and Nagy: ON–OFF Feedback Control of PFC

is summarized in the following: set of coupled differential equations within the different
dδ dχ 1 dm domains:
= kE = ρW Cp,W ∂TW
dt dt ρE dt 1 ∂TW ∂ 2 TW ∂ 2 TW
Wall : = + +
∂t r ∂r ∂r 2 ∂z2

km 1 km f
kW
= + Cb − Csat ∂TE
 
−1 ∂TE

ρE 2 kR Encrust : =
kE 1
!1/2  ∂t ρE Cp,E Ri − r̃δ δ ∂ r̃
1 km2 km f

− + Cb − Csat 1 ∂ 2 TE ∂ 2 TE
 
4 kR2 kR + 2 +
δ ∂ r̃ 2 ∂z2
ρE
1/3 2
∂ ∂
− (1 + α1T )dp ρL2 ηg w δ Tube : (Af T ) = − (uAf T ) +
k
+
83.2w0.54 ∂t ∂z ρCp,L
where :
∂ ∂T
 

1Ef
 2π Rf h
kR = kR0 exp − Af + (TE |Rf − T ) (4)
RTf ∂z ∂z ρCp,L
Tf = T + 0.55(TR |r=Rf − T ) (1) where r̃ = (Ri − r/δ) is a dimensionless radial coordinate,
which ranges between 0 (r = Ri ) and 1 (r = Rf ) irrespective
where δ is the encrust thickness on the reactor’s wall, kE
of the encrust thickness. Axial symmetry is assumed with h
is the thermal conductivity (W · m−1 K −1 ), χ is the ther-
defined as the overall heat transfer coefficient and Cp,L the
mal resistance (m2 K · W ), ρE is the encrust density, m is
specific heat capacity of the liquid slurry. The B.C.s for the
the encrust mass per unit area (kg · m−2 ), km is the mass
different domains can be summarized as follows:
transfer coefficient of solute from the bulk solution to the
1) the first-order continuity and identical temperature at the
encrust film, kR is the adsorption rate of solute to encrust
interface between the inner reactor wall and the encrust; 2) the
and assumed to be of Arrhenius type (m4 kg−1 s−1 ), Cb is the
f first-order continuity between the rate of heat conduction and
solute concentration in the bulk liquid, Csat is the solubility heat convection at the interface between the encrust and the
within the film layer, and w is the oscillatory mixing velocity tube; and 3) the inlet temperature along the reactor is defined
[3]. α is the linear expansion coefficient (K −1 ) [5], 1T is as the feed temperature. The mass balance in turn is given as
the temperature difference between the reactor wall and the
∂ ∂ ρc ∂
encrust surface, dp is the mean particle diameter within the (Af C) = − (uz Af C) − φν (Af µ3 )
encrust, η is the film viscosity (Pa·s), and g is the gravitational ∂t ∂z ρL ∂t
acceleration. In addition, km is semiempirically correlated ρE ∂δ
− 2π (Ri − δ)
with the Sherwood number ρL ∂t
ρE = (1 − )ρc + ρL (5)
Sh = 0.034Re0.875 Sc1/3 R∞ 3
2Rf km 2Rf wρL η where µ3 = 0 L n(t, L, z) dL is the third moment of the
Sh = , Re = , Sc = . (2)
D η ρL D CSD and αν is the volumetric shape factor. ρE is the encrust
The encrustation dynamics is coupled with the crystallization density and can be calculated from the encrust void fraction .
dynamics given by the population balance equation (PBE) as The resulting encrustation-PBE dynamics is a nonlinear and
follows: highly stiff PDE, which is solved using the high-resolution
finite volume method [4]. The mass and energy balances are
∂ ∂ ∂
(Af n) + (uz Af n) + (GAf n) = 0 (3) solved via finite differences.
∂t ∂z ∂L
and with the following boundary conditions (B.C.s): III. FOULING IN OPTIMIZED PFC
G(S)n(t, L, z)|L=0 = B(S), n(t, L, z)|z=0 = nseed (L). Here, The dynamical model suggests that as PFC is optimized for
n is the CSD, uz is the mean flow velocity, G is the crys- crystal growth, the encrust formation is inevitably enhanced
tal growth rate, B is the nucleation rate, nseed is the seed due to a higher supersaturation in the boundary layer
CSD, z is the reactor axis, and L is the crystal length axis. (Fig. 2). A simulation study is performed to illustrate this
Af (t, z) = πR2f (t, z) is the flow area within the tube, which phenomenon. With the crystal volume mean size (L43 ), i.e.,
changes with time and along the reactor length due to encrus- the ratio of the fourth and third moments of the CSD, used as
tation. Rf is the PFC radius, D is the solute diffusivity, ρL is the objective of the optimization, the optimization problem is
the bulk liquid density, Re is the Reynold’s number, and Sc formulated as follows:
is the Schmidt number. The encrustation and crystallization Z
dynamics are also coupled with mass and energy transfer max J (T ) = L43 (T ), i ∈ [1, 4], T = 12τ, τ =
Ti uz
equations. The energy balance is divided into three domains,
s.t. : (1) to (5) and 20 ≤ Ti ≤ 40. (6)
namely, conduction across the reactor wall (W : r ∈
[Ri , R0 ]), conduction across the encrust (E : r ∈ [Rf , Ri ]), Note that the reactor is divided into four equal segments, each
and convection within the tube (T : r ∈ [0, Rf ]). Both associated with its own heat jacket, which are also the deci-
the conduction and convection dynamics yield the following sion variables, i.e., Ti , i ∈ [1, 4]. Potash-alum system is used

2 VOLUME 3, NO. 1, MARCH 2017

Koswara and Nagy: ON–OFF Feedback Control of PFC

FIGURE 2. Solubility phase diagram in the presence of encrustation.

During cooling (heating), the temperature of the encrust layer is lower
(higher) than that of the tube resulting in higher (lower) solubility in the
film layer.
FIGURE 3. Comparison between PFC with and without fouling. RT and %
blockage with fouling (top) and CSD (middle) and supersaturation
(bottom) with and without fouling.
as the model compound for this letter as its crystal kinetics
are well known. However, its encrust kinetics are not known
and thus, when possible, the parameter values are either taken IV. ROBUST ON–OFF FEEDBACK CONTROL
from previously verified fouling experiments [5] or assumed The AFC design divides the multisegment PFC into two
as in [3]. The compound-specific nucleation, growth, and symmetric parts, which periodically cycle between cooling
dissolution kernels are summarized in the following: and heating regions. Here, the control objective is also to
maximize crystal growth but with the additional domain-
 
−1EG
G = kG exp (1 − exp(−γ (L + β)))(σ − 1)g specific constraints, namely: during cooling, the temperature
RT

jb
 is optimized for maximization of crystal growth, while during
j
B(S) = ja exp + kb MT S b heating, complete encrust dissolution is enforced. The opti-
T 3 (ln σ )2
mization problem can be formulated as follows:
C
MT = ρ c α ν µ 3 , σ = , S = C − Csat
Csat max J (T ) = L43 (T ) s.t. : (1) to (5) and (7), and
Ti
Csat = AT + BT 2 + CT 3
in cooling segment : 20 ≤ Ti ≤ 35
kD
for S < 0 (dissolution) : G = − q S d . (7) in heating segment : 39 ≤ Ti ≤ 45, δ(T ) = 0. (8)
L
The seed is modeled as a normally
√ distributed crystal size The bounds of the different temperature segments are chosen
density, i.e., nseed = (κ/σseed 2π )exp(−((L −µseed )2 /σseed
2 )), such that only growth and dissolution take place in the desig-
where κ is a scaling factor, which is determined accord- nated cooling and heating regions, respectively. In addition,
ing to the desired seed mass fraction. Genetic algorithm the optimization is performed for two cycles with the set
is the optimization routine of choice, with the initial of I.C.s obtained from the first run. In the first cycle, the
conditions (I.C.s) set according to the following settings: I.C.s are the same as the optimized PFC case but with an
1) the tube temperature equals the in-flow temperature; initial encrust thickness on the second half—heating region—
2) the tube concentration starts at supersaturation; and 3) of the reactor (Fig. 4, bottom-right). It is important to note
there is no initial crystal or encrust. As seen in Fig. 3, while that since the temperature profile within the tube cannot be
the optimization leads to formation of large crystals (middle), manipulated arbitrarily such that supersaturation is main-
it also results in significant blockage (top), rendering it a tained while encrust is dissolving, undersaturation takes place
batch process. In addition, when the result is compared with in time in the leading part of the dissolution segment causing
crystallization without fouling by removing the encrustation the crystal to dissolve (Fig. 4, bottom-left). Consequently,
kinetics in the dynamical model, not only is blockage pre- optimal spatial and temporal cooling and heating temperature
vented, but the heat transfer dynamics and supersaturation profile (Fig. 4, top-right) must be obtained such that overall
level are reduced such that both the crystal mean size and crystal growth takes place along the reactor (Fig. 4, top-left)
yield are higher (Fig. 3, bottom). and all the while encrust formation and dissolution are peri-

VOLUME 3, NO. 1, MARCH 2017 3


FIGURE 4. Simulation of AFC via heating and cooling cycle. The figure
includes the snapshots of (top-left) the CSD evolution and (top-right) the
temperature profile at the end of one cycle, (bottom-left) spatial and
temporal supersaturation dynamics, and (bottom-right) encrustation
dynamics.
FIGURE 5. Response of PFC with ON–OFF CSD feedback controller in
terms of collected product output (top: dotted solid line) and
waste (top: dotted faint line) CSD, and L43 (bottom).
odically cycled between the segments so as to maintain
continuous operations. An open-loop AFC implementation,
however, means that optimal temperature profile is calculated
iteratively at the end of each cycle in order to maintain
complete encrust dissolution, maximize yield, and ensure
that the product lot is consistently within specifications. In
order to avoid this stringent requirement, the periodic tem-
4 VOLUME 3, NO. 1, MARCH 2017
Koswara and Nagy: ON–OFF Feedback Control of PFC
FIGURE 6. Response of encrust feedback controller in terms of encrust
evolution from cycle-to-cycle in one segment. The controller increases the
heating duration in the first cycle as it senses that encrust persists. It
subsequently increases the heating temperature in the following odd
cycle by a preset constant and reduces the duration of the heating cycle
to prevent overdissolution. The next odd cycle uses a temperature profile,
which is the midpoint of the previous two heating temperatures. This
three-step procedure is repeated.
perature control obtained in the optimization can be cou-
pled with a feedback sensor on the CSD (e.g., a Focused
Beam Reflectance Microscopy probe), which determines the
period of output stream collection for as long as the product
is within the upper (150 µm) and lower bound specifica-
tions (100 µm) (Fig. 5). In addition, an onset encrust sensor
(e.g., a gray-scale camera) can be utilized to actively modify
the duration and magnitude of the heating action, such that
complete dissolution of the encrust occurs without crystal
overdissolution (Fig. 6). Hence, the process can continue
indefinitely and with the confidence that the product would
always be within specifications.
V. SUMMARY AND DISCUSSION
A novel ON–OFF feedback control of a PFC with AFC imple-
mentation is described herein. The feedback controllers,
which are based on CSD measurements and encrust sensing,
overcome the limitations of the PFC process due to fouling.
As a result, encrust is periodically negated and without signif-
icant reduction in product quality and consistency. This letter
demonstrates an important example of QbC, an essential
concept for realizing a truly continuous process.
REFERENCES
[1] Z. K. Nagy and R. D. Braatz, ‘‘Advances and new directions in
crystallization control,’’ Annu. Rev. Chem. Biomol. Eng., vol. 3,
pp. 55–75, Jul. 2012.
[2] M. Jiang et al., ‘‘Continuous-flow tubular crystallization in slugs sponta-
neously induced by hydrodynamics,’’ Cryst. Growth Design, vol. 14, no. 2,
pp. 851–860, 2014.
[3] A. Majumder and Z. K. Nagy, ‘‘Dynamic modeling of encrust formation
and mitigation strategy in a continuous plug flow crystallizer,’’ Cryst.
Growth Design, vol. 15, no. 3, pp. 1129–1140, 2015.
[4] R. Gunawan, I. Fusman, and R. D. Braatz, ‘‘High resolution algorithms for
multidimensional population balance equations,’’ AIChE J., vol. 50, no. 11,
pp. 2738–2749, 2004.
[5] M. Bohnet, ‘‘Fouling of heat transfer surfaces,’’ Chem. Eng. Technol.,
vol. 10, no. 1, pp. 113–125, 1987.