See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/320150546

Advanced control of a continuous oscillatory flow crystalliser

Article  in  Control Engineering Practice · October 2017

DOI: 10.1016/j.conengprac.2017.07.008

CITATIONS READS

7 287

9 authors, including:

Furqan Tahir Karolina Krzemieniewska-Nandwani

Perceptive Engineering Limited Manchester Metropolitan University
30 PUBLICATIONS   184 CITATIONS    5 PUBLICATIONS   16 CITATIONS   

SEE PROFILE SEE PROFILE

Humera Siddique Alastair J Florence

University of Strathclyde University of Strathclyde
10 PUBLICATIONS   248 CITATIONS    245 PUBLICATIONS   3,605 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

GSK 104 View project

Continuous isolation View project

All content following this page was uploaded by Humera Siddique on 25 March 2019.

The user has requested enhancement of the downloaded file.

 Control Engineering Practice 67 (2017) 64–75

Contents lists available at ScienceDirect

Control Engineering Practice

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

Advanced control of a continuous oscillatory flow crystalliser

F. Tahir a, *, K. Krzemieniewska-Nandwani b,1 , J. Mack a , D. Lovett a , H. Siddique c , F. Mabbott c ,
V. Raval c , I. Houson c , A. Florence c
a
Perceptive Engineering Ltd, Vanguard House, Sci Tech Daresbury, Cheshire WA4 4AB, UK
b
Manchester Metropolitan University, All Saints Building, Manchester M15 6BH, UK
c
EPSRC Centre for Innovative Manufacturing in Continuous Manufacturing and Crystallisation, University of Strathclyde, Glasgow G1 1RD, UK

a r t i c l e i n f o a b s t r a c t

Keywords: This paper presents the application and challenges of achieving Model Predictive Control (MPC) on two
Continuous crystallisation continuous oscillatory baffled crystallisation reactors, delivering precise product quality control in the face of
Super-saturation control raw material fluctuations. A key advantage of MPC is that it effectively deals with multivariable interactions
Model predictive control and constraints that appear within the continuous crystallisation process. Using a flexible real-time software
Calibration models
package, a control scheme is proposed that incorporates three MPC blocks for controlling reactor cooling profile,
Crystal PSD
API (Active Pharmaceutical Ingredient) concentration and crystal size distribution, respectively. Furthermore,
the solution is customisable and transferable to different crystallisation reactors as well as various APIs.
© 2017 Elsevier Ltd. All rights reserved.

1. Introduction to continuous crystallisation
Crystallisation is a critical production step in the Pharmaceutical
Industry. Whether it is for purification of the final Active Pharmaceutical
Ingredient (API) or synthesis of compounds, it is necessary to maintain
precise process conditions to achieve a product which meets quality
specifications.
Batch processing, mainly using large continuously stirred tank re-
actors (CSTRs), is the most common approach to crystallisation for
commercial manufacturing. The crystallisation process for a given
compound is typically developed at a small scale in a research labo-
ratory. For simple cool-down crystallisation, PAT (Process Analytical
Technology) enabled experimental techniques can be used to establish
the compound’s Meta-Stable Zone (MSZ) and super-saturation control
strategies (Fig. 1). During these experiments the mixing rates, seed
loading and cooling profiles are optimised to achieve desired properties
of crystals, such as their morphology or particle size distribution. Once
these conditions are established, the process can then be scaled up to
manufacturing volumes.
As the process is scaled up, accurate control of the reactor becomes
more challenging. The Meta-Stable Zone is not only compound specific,
but also reactor specific. Process parameters such as impeller design,
mixing shear force and heat-transfer characteristics can shift the meta-
stable zone significantly from the small-scale parameters. In addition,
closed loop super-saturation control is seldom used during routine man-
ufacturing. This often results in inconsistent particle size distribution
from one batch to another. Moreover, batches containing higher levels
of fines severely limit the effectiveness of downstream processes such as
filtration and drying.
These problems have led the Pharmaceutical Industry to seek contin-
uous crystallisation methods as an alternative for batch manufacturing.
The aim is to discard the need for reactor scale up, which would mean
that the process designed in the laboratory could be readily reproduced
on a manufacturing plant (Zhao et al., 2014). Moving the crystallisation
process from batch to continuous would also mean better concentration
control, more consistent product and hence lower manufacturing costs.
Tighter control of the processes would also facilitate implementation of
the Quality-by-Design (QbD) techniques. Attaining more precise control
of product attributes using a combination of PAT and Advanced Control
techniques fulfils the QBD objectives; doing so for a continuous process
enhances the versatility and also addresses the need for increased
manufacturing flexibility.
Continuous Oscillatory Baffled Reactors (COBRs) provide an alterna-
tive to CSTRs, where mixing is achieved through oscillatory movement
of the fluid though narrower parts of the reactor. Here, a time dependent
cooling profile is transformed into a length dependent one, where
residence time can be adjusted by varying flow of feed solution. A single

* Corresponding author.
E-mail addresses: ftahir@perceptiveapc.com, ftahir49@hotmail.com (F. Tahir).
1 Co-first author.

http://dx.doi.org/10.1016/j.conengprac.2017.07.008
Received 1 November 2016; Received in revised form 18 July 2017; Accepted 21 July 2017
0967-0661/© 2017 Elsevier Ltd. All rights reserved.
F. Tahir et al.
Fig. 1. Crystallisation Meta-stable zone graph.
reactor of up to several metres in length can be assembled by combining
multiple straights. Since it is possible to achieve plug flow through
the reactor, a COBR is a natural choice for continuous crystallisation.
Moreover, a COBR allows for uniform and efficient mixing, rapid heat
transfer and reduced sheer. Literature describes several crystallisation
transfers from batch OBRs to COBRs and advantages of continuous
crystallisation (McGlone et al., 2015; Siddique, Brown, Houson, &
Florence, 2015).
A number of control/monitoring schemes have been proposed in the
literature for the crystallisation process. In Orkoulas, Kwon, Nayhouse,
and Christofides (2014), the authors propose a comprehensive dynamic
model for a continuous plug flow crystallisation unit based on the Mass,
Energy and Population balances. This model is then used to compute a
set of optimal jacket temperatures for the crystalliser as well as flow ve-
locity which minimises the crystal shape and size deviations from their
desired values. The scheme also takes account of disturbances to the
inflow solute concentration by computing new jacket temperature. In Li,
Mhaskar, Shi, El-Farra, and Christofides (2006), the authors present a
population balance model as well as a hybrid predictive control scheme
to stabilise the oscillatory crystallisation dynamics (that occur due to
complex interplay of nucleation and growth) as well as control particle
size. In Chiu and Christofides (1999), a population model and nonlinear
control scheme is developed to minimise the oscillations associated with
a homogeneous (CSTR-type) continuous crystallisation unit, obtain a
desired PSD as well as guarantee closed-loop stability. Kwon, Nayhouse,
Christofides, and Orkoulas (2014) used both population balance mod-
elling and MPC to control shape and size. The model, with one actuator
(jacket temperature) was simulated and showed that MPC can reduce
the impact of disturbances on the crystallisation process. This approach
was developed further by Mack (2013), where MPC was applied to the
temperature cooling profile, while ensuring that desired concentration
was achieved. Furthermore, Forgione et al. (2015) applied, on a test
plant, an iterative algorithm which adjusted the cooling profile at
each batch, while the product temperature was controlled with a PI
controller.
In this work, a novel advanced control scheme is proposed for the
plug flow continuous crystallisation process. The main objectives of this
solution are to control the reactor cooling profile, API concentration
and crystal particle size. Due to plug flow dynamics, some of the
aforementioned issues around oscillations are avoided. Furthermore,
due to real-time feedback from the inline PAT, the proposed control
scheme is able to deal with various disturbances such as reactor fouling
and feed impurity. The scheme considers Lactose as the API and is tested
on two commercially available continuous reactors: Cambridge Reactor
Design’s Rattlesnake COBR and Nitech’s DN15 COBR (see Sections 1.1
and 1.2, respectively).
65
Control Engineering Practice 67 (2017) 64–75
The main contribution of this paper is the development of a flexible
advanced process control scheme for continuous crystallisation which
can be applied to different APIs as well as different reactors. The
proposed scheme consists of three MPC blocks to control, in real-
time, the reactor temperature profile, API concentration and crystal size
distribution, respectively. The controllers make use of real-time particle
size information as well as concentration prediction from the designed
calibration models (based on the PAT data). This helps to overcome any
process disturbance that may occur during the run.
This paper is organised as follows: The remainder of Section 1
describes the Rattlesnake and DN15 reactors as well as the experimental
setup. A brief discussion of the model structure used for crystallisation
process is presented in Section 2. For completeness, a brief formulation
of the MPC algorithm is provided in Section 3. An overview of the pro-
posed APC solution is presented in Section 4. The reactor temperature
control scheme is formulated in Section 5. API concentration and crystal
size control schemes are respectively developed in Section 6. Finally,
conclusion is presented in Section 7. A list of acronyms is given in
the Appendix.
1.1. Rattlesnake crystallisation reactor
Rattlesnake is a COBR unit developed by Cambridge Reactor De-
sign (Cambridge Reactor Design, 0000). A lactose crystallisation study,
reported in (Siddique et al., submitted for publication), showed that
this continuous crystallisation unit is able to deliver 26% higher yield
and significant reduction in particle span as compared to its batch
counterpart.
The unit consists of four cylindrical modules, as shown in Fig. 2.
Each module has a length of 740 mm and an internal diameter of 69 mm
(giving an approximate volume of 2760 ml). To control the temperature,
each module is fitted with a double water filled jacket: a primary
jacket with co-current configuration and a secondary coil with counter-
current flow. As shown in Fig. 2, these are connected to the hot and
cold circulators, respectively. Such a design enables precise temperature
control which is important for maintaining an accurate cooling profile
during the crystallisation. The residence time in the reactor is generally
a function of the throughput and varies from approximately 60 min
upwards. Further information about the unit is available in Cambridge
Reactor Design (0000).
With regards to the experimental setup, product is fed through
two peristaltic pumps—for feed solution and seed slurry, respectively.
Additionally, oscillatory feed-flow is achieved by a push–pull hydraulic
oscillator, as indicated in Fig. 2. Different product properties in the reac-
tor (e.g. temperature, and particle size) are monitored (and controlled)
through a set of thermocouples and PAT devices. In particular, we use
Focused Beam Reflectance Measurement (FBRM) for particle size and
Fourier Transform Infrared (FTIR) spectrometer to predict the real-time
solute concentration (through the development of a calibration model—
see the sections below).
1.2. DN15 crystallisation reactor
The DN15 is another type of COBR developed by Nitech and
CMAC (Nitech Solutions, 0000). As shown in Fig. 3, it consists of
31 straights (tubes), each containing of a series of multiple baffle
and inter-baffle zones. The reactor internal diameter is 15 mm and
a length of 14 m giving a total volume of approximately 2500 ml.
The tubes are jacketed, providing liquid flow in order to control the
product temperatures throughout the reactor. The temperature range
is from 0 to 100 ◦ C and the reactor is fitted with 39 thermocouples
and 13 circulators to maintain the temperature profile. PAT equipment,
including FTIR spectrometer and FBRM, is integrated into the reactor
at certain positions. Similar to the Rattlesnake unit, DN15 is also able
to achieve plug-flow with residence times from approximately 60 min
upwards (depending on the flow). A schematic of the DN15 is shown in
Fig. 3. Further information on the unit is available at (Nitech Solutions,
0000).
F. Tahir et al. Control Engineering Practice 67 (2017) 64–75

Fig. 2. Rattlesnake COBR by Cambridge Reactor Design.

Fig. 3. DN-15 COBR by NiTech.

2. Model structure for COBR dynamics
Data driven models for the COBR unit have been identified using
the RLS (Recursive Least Square) algorithm. This has the advantage of
relatively fast calculations and good convergence properties (Goodwin
& Sin, 2009). In particular, three Output Error (OE) models are identified
for predicting cooling temperature, concentration of Lactose and crystal
size distribution, respectively (see Remark 1). The structure of the OE
prediction model considered is of the form:
𝑠𝑢
∑ unmeasured drift within the process (see e.g. Maciejowski, 2002 and
𝐲𝐤+𝟏 − 𝐲𝐤 = 𝐀𝟏 (𝐲𝐤−𝟏 − 𝐲𝐤 ) + 𝐁𝐢 𝚫𝐮𝐤−𝐢 (1)
𝑖=0
where 𝐲𝐤 is an 𝑛-dimensional vector of Controlled Variables (CVs),
and 𝐮𝐤 is an 𝑛𝑢 -dimensional vector of Manipulated Variables (MVs),
all at sample time 𝑘. The symbol 𝑠𝑢 denotes the delay spread for the
MVs, and 𝐀𝟏 , 𝐁𝐢 , ∀𝑖, are appropriately-dimensioned matrices of model-
coefficients. Finally, the 𝚫 in Eq. (1) denotes incremental signals such
that:
𝚫𝐮𝐤 = 𝐮𝐤 − 𝐮𝐤−𝟏 .
Note that the considered model (1) has an incremental structure (as
opposed to absolute). Such a model format has the advantage of
robust identification of model-coefficients as well as rejection of any
the references therein).

66
F. Tahir et al. Control Engineering Practice 67 (2017) 64–75

3. Model predictive control algorithm

Model Predictive Control (MPC) is a receding horizon technique in

which the current control action is obtained by solving on-line at each
sampling instant, a finite horizon optimised problem. The optimisation
yields an optimal control sequence and the first control in this sequence
is applied to the plant (Maciejowski, 2002). A general structure of the
MPC algorithm is given in Fig. 4.
MPC offers several advantages over traditional control schemes,
including the ability to explicitly take account of process constraints,
as well as to handle multivariable, non-minimum phase and unstable
processes. Due to these reasons, MPC has been widely implemented for
advanced process control within various industries (Qin & Badgwell,
Fig. 4. MPC structure. 2003). Furthermore, it has also been the subject of extensive academic
research – to improve robustness whilst reducing computational load
and conservatism – over the past few decades (see e.g. Mayne, Rawlings,
Rao, & Scokaert, 2000; Tahir & Jaimoukha, 2013 and the references
Remark 1. Note that (nonlinear) mechanistic or hybrid models for therein).
the crystallisation process can also be derived by taking account Given the dynamics in (1), the MPC algorithm computes at sampling
of nucleation, growth and agglomeration stages and obtaining ap-
instance 𝑘, the control move which minimises a cost function of the
propriate mass/energy/population balance equations (for example,
form:
see Georgieva, Meireles, & Feyo de Azevedo, 2003, as well as the
discussion on Chiu & Christofides, 1999; Li et al., 2006 and Orkoulas et ∑
𝑁
𝐽= (𝐲𝐤+𝐢 − 𝐲𝐬𝐩 )𝐓 𝐐(𝐲𝐤+𝐢 − 𝐲𝐬𝐩 ) + 𝚫𝐮𝐓
𝐤+𝐢−𝟏
𝐑𝚫𝐮𝐤+𝐢−𝟏 (2)
al., 2014 in Section 1). In the proposed scheme, the main emphasis is on
𝑖=1
linear, data-driven models identified by carrying out process response
tests on both DN15 and Rattlesnake. This data-driven approach for subject to the control and move constraints:
linear models has the advantage of simplicity as well as a relatively
𝐮 ≤ 𝐮 ≤ 𝐮, (3a)
quick development cycle. This is complemented by the fact that the
scheme relies upon real-time feedback from the PAT to ensure that the 𝚫𝐮 ≤ 𝚫𝐮 ≤ 𝚫𝐮 (3b)
APC keeps the particle size, concentration and temperature profiles at
the desired values (in spite of any process disturbances). Although such where 𝑁 is the prediction horizon, 𝐲𝐬𝐩 is the CV setpoint, and 𝐐, 𝐑 are
models achieve good control performance (results reported in Sections 5 the setpoint- and move-weights, respectively. Furthermore
and 6), however they do have a limited range of applicability (e.g. the
⎡ 𝚫𝐮𝐤 ⎤ ⎡ 𝚫𝐮𝐤 ⎤ ⎡ 𝚫𝐮 ⎤
startup phase). In this regard, hybrid models can be incorporated to ⎢ ⎥ ⎢ ⎥ ⎢ 𝚫𝐮 𝐤 ⎥
𝚫𝐮𝐤+𝟏 ⎥ 𝚫𝐮𝐤+𝟏 ⎥
broaden the crystalliser’s operating space as well as reduce the overall 𝚫𝐮 = ⎢ , 𝚫𝐮 ∶= ⎢ , 𝚫𝐮 ∶= ⎢ 𝐤+𝟏 ⎥
process response testing time and this forms part of future work. ⎢ ⋮ ⎥ ⎢ ⋮ ⎥ ⎢ ⋮ ⎥
⎢𝚫𝐮 ⎥ ⎢ ⎥ ⎢𝚫𝐮 ⎥
⎣ 𝐤+𝐍−𝟏 ⎦ ⎣𝚫𝐮𝐤+𝐍−𝟏 ⎦ ⎣ 𝐤+𝐍−𝟏 ⎦

Fig. 5. Transferable APC solution.

67
F. Tahir et al. Control Engineering Practice 67 (2017) 64–75

Fig. 6. Cooling profile design in a batch reactor.

with analogous definitions applying to the control vectors 𝐮, 𝐮 and 𝐮,

respectively.
A manipulation on Eq. (1) shows that the setpoint error 𝐞𝐤+𝟏 ∈ R𝑛
can be written as:
𝑠𝑢
∑
𝐞𝐤+𝟏 = 𝐲𝐤+𝟏 − 𝐲𝐬𝐩 = (𝐈 − 𝐀𝟏 )𝐞𝐤 + 𝐀𝟏 𝐞𝐤−𝟏 + 𝐁𝐢 𝚫𝐮𝐤−𝐢 (4)
𝑖=0
where 𝐈 is an identity matrix of appropriate dimensions.
It follows that the overall system can now be written in the form:

𝐒𝐤+𝟏 = 𝐆𝐒𝐤 + 𝐌𝚫𝐮𝐤 (5)

where
⎡𝐈⎤
⎡ 𝚫𝐮𝐤−𝟏 ⎤ ⎢𝟎⎥
⎢ 𝚫𝐮𝐤−𝟐 ⎥ ⎢ ⎥
⎢ ⋮ ⎥ ⎢𝟎⎥
𝐒𝐤 ∶= ⎢ ⎥, 𝐌 ∶= ⎢ ⋮ ⎥ ,
⎢𝚫𝐮𝐤−𝑠𝑢 ⎥ ⎢ ⎥
⎢ 𝐞 ⎥ ⎢𝟎⎥
⎢ 𝐤 ⎥ ⎢𝐁𝟎 ⎥ Fig. 7. APC tracking of optimised cooling profile for DN15. (For interpretation of the
⎣ 𝐞𝐤−𝟏 ⎦ ⎢ ⎥ references to color in this figure legend, the reader is referred to the web version of this
⎣𝟎⎦ article.)
⎡𝟎 𝟎 ⋯ 𝟎 𝟎 𝟎 𝟎⎤
⎢𝐈 𝟎 ⋯ 𝟎 𝟎 𝟎 𝟎⎥
⎢ ⎥
⎢𝟎 𝐈 ⋯ 𝟎 𝟎 𝟎 𝟎⎥ ̂ is a diagonal matrix of 𝐐, 𝐑
where 𝐐 ̂ is a diagonal matrix of 𝐑, and
𝐆 ∶= ⎢ ⋮ ⋮ ⋱ ⋮ ⋮ ⋮ ⋮⎥
⎢ ⎥ ⎡𝐈 𝟎 𝟎 ⋯ 𝟎⎤
⎢𝟎 𝟎 ⋯ 𝐈 𝟎 𝟎 𝟎⎥
⎢ ⎥
⎢𝐁𝟏 𝐀𝟏 ⎥ 𝐈 𝐈 𝟎 ⋯ 𝟎⎥
⎢
𝐁𝟐 ⋯ 𝐁𝑠𝑢 −𝟏 𝐁𝑠𝑢 𝐈 − 𝐀𝟏
⎥ 𝐄 ∶= ⎢ , 𝐟 ∶= 𝐮 − 𝐮𝐤−𝟏 , 𝐟 ∶= 𝐮 − 𝐮𝐤−𝟏 .
⎣𝟎 𝟎 ⋯ 𝟎 𝟎 𝟎 𝟎⎦ ⎢⋮ ⋮ ⋮ ⋱ ⋮⎥
⎢𝐈 𝐈 𝐈 ⋯ ⎥
𝐈⎦
with 𝟎 representing an appropriately-dimensioned matrix of zeros. ⎣
Now consider a matrix 𝐂 which extracts – from (5) – rows corre- As mentioned above, the first move (𝚫𝐮𝐤 ) of the optimal computed
sponding to the prediction error (i.e. 𝐞𝐤+𝟏 ). Then, iterating Eq. (5) yields: sequence 𝚫𝐮 is applied to the process and the optimisation problem (7)
is solved again at the next time step.
𝐏𝐤 = 𝐅𝐒𝐤 + 𝐇𝚫𝐮 (6) The convex quadratic program in (7) is solved within the Phar-
where maMV real-time system (Perceptive Engineering Limited, 0000), at
each sampling instance, using the dual active-set method proposed
⎡ 𝐞𝐤+𝐍 ⎤ ⎡ 𝐂𝐆𝐍 ⎤
⎢ ⎥ ⎢ ⎥ by Goldfarb and Idnani (1983) through the implementation routine
𝐞 𝐂𝐆𝐍−𝟏 ⎥
𝐏𝐤 = ⎢ 𝐤+𝐍−𝟏 ⎥ , 𝐅=⎢ , given in Powell (1983). This algorithm can be summarised as follows:
⎢ ⋮ ⎥ ⎢ ⋮ ⎥ By solving the unconstrained minimum of (7), a dual feasible point
⎢ 𝐞 ⎥ ⎢ 𝐂𝐆 ⎥
⎣ 𝐤+𝟏 ⎦ ⎣ ⎦ is obtained which is a primal optimal for a sub-problem of (7). The
⎡𝐂𝐆𝐍−𝟏 𝐌 𝐂𝐆𝐍−𝟐 𝐌 ⋯ 𝐂𝐆𝐌 𝐂𝐌⎤ algorithm then iterates to achieve dual optimality (primal feasibility).
⎢ ⎥ The iterations rely on Cholesky and QR factorisations within the update
𝐂𝐆𝐍−𝟐 𝐌 𝐂𝐆𝐍−𝟑 𝐌 ⋯ 𝐂𝐌 𝟎 ⎥
𝐇=⎢ . procedure. As mentioned in Goldfarb and Idnani (1983), it is important
⎢ ⋮ ⋮ ⋯ ⋮ ⋮ ⎥
⎢ 𝐂𝐌 𝟎 ⋯ 𝟎 ⎥
𝟎 ⎦ to note that the origin in the space of dual variables is always dual
⎣
feasible so phase 1 of active-set methods (for computing a feasible point)
By inserting (6) into (2) and re-arranging, the constrained MPC is not needed.
problem – at each sampling instance 𝑘 – can be summarised as:
𝐽 ⋆ = minimise𝚫𝐮 𝚫𝐮𝐓 (𝐇𝐓 𝐐𝐇
̂ + 𝐑)𝚫𝐮
̂ + 𝟐𝐒𝐓 𝐓̂ 𝐓 𝐓̂
𝐤 𝐅 𝐐𝐇𝚫𝐮 + 𝐒𝐤 𝐅 𝐐𝐅𝐒𝐤
Remark 2. It is worth noting here that the continuous crystallisation
(7) reactors have a large residence time as well as settling time. For
subject to ∶ 𝐟 ≤ 𝐄𝚫𝐮 ≤ 𝐟, 𝚫𝐮 ≤ 𝚫𝐮 ≤ 𝚫𝐮

68
F. Tahir et al. Control Engineering Practice 67 (2017) 64–75

Fig. 8. MPC control structure for reactor temperature profiles.

Fig. 9. Temperature PRBS step tests on DN15.

example, the dead time (pure delay) for the ’seed flow to Concentration’
loop is approximately 1 h with settling time of around 1.5 h. All this
serves to increase the size of the optimisation problem. Due to this
reason, horizon compression is used within the MPC scheme to ensure
that the optimisation problem is solved in a timely manner with regards
to the control interval (see Section 4 for details).
4. Transferable APC solution
As stated above, MPC is particularly well suited to continuous
crystallisation reactors as they often exhibit multivariable interactions,
constraints, significant time delays and slow process responses. The
MPC employs dynamic process model predictions and feedback to
compensate for both measured and un-measured process disturbances.
However, development of such control systems can be time consuming.
Therefore, the focus of this paper is to develop a flexible solution that
can be transferred between products and reactors thus reducing time
and effort. Fig. 5 shows the proposed approach to the development of a
transferable APC solution for continuous crystallisation. It is designed to
be fully adaptable to different crystallisation reactors with temperature-
based concentration control.
The proposed APC solution is implemented in PharmaMV real-
time system (Perceptive Engineering Limited, 0000) and consists of
three Model Predictive Controllers (with two of them in a cascade
arrangement—Fig. 5). Since crystallisation has relatively slow dynam-
ics, the entire process and PAT data is sampled every 5 s. The first MPC
provides control of the reactor temperature profile and has an interval
of 60 s for both DN15 and Rattlesnake reactor (further details are given
in Section 5). The particle size MPC controller, with an interval of 480 s,
uses FBRM measurements in real-time. This controller achieves the

69
F. Tahir et al. Control Engineering Practice 67 (2017) 64–75

(a) Structural overview of the temperature model coefficients.

(b) Zoomed-in version of the temperature model coefficients.

Fig. 10. Temperature model for the DN15 reactor.

desired D50 crystal size by adjusting the concentration (supersaturation)
setpoint. This setpoint is then written down to the concentration MPC
controller which has an interval of 60 s (for DN15) and manipulates
the feed and seed pumps. As shown in Fig. 5, the controller receives
real-time concentration predictions from a PLS calibration model based
on the FTIR spectral data as well as temperature.
It is worth mentioning that the control structure in Fig. 5 is gener-
alised for continuous crystallisers. This structure and process model can
be programmatically reconfigured to support transfer between reactors.
Furthermore, both cascade controllers are transferable between reactors
by model switching/adaption techniques and deliver crystals of the
desired size.
5. Temperature profile development and control
To produce crystals of the required final product quality, the cooling
profile should be optimised to ensure growth occurs within the Meta-
Stable Zone throughout the evolution of the reaction. In order to develop
an optimal cooling profile for Lactose in solution, a set of offline trials
were run in a batch stirred tank reactor (results shown in Fig. 6). The
desirable profiles were then tested on both the continuous reactors and
tuned further to optimise yield, mean particle size and span. Note that
span is a measure of the width of the particle size distribution and
is calculated by subtracting the lower decile from upper decile and
dividing the result by the median particle size.
Fig. 7 shows the optimised temperature cooling profile, as applied
to each of the thermocouples on the DN15 as well as the APC tracking
results. The red bars represent the desired cooling profile (which is the
optimal temperature value with a margin of ±2 ◦ C as the target for
the controller). The blue vertical bars represent the actual temperature
values achieved by the APC during a particular run along with the
measurement uncertainty. Finally, the blue horizontal bars are simply to
demarcate/clarify different thermocouple temperature values on Fig. 7
(they do not represent any error margins/uncertainty).
Note that a similar cooling profile was also applied to the Rat-
tlesnake.
5.1. MPC control for reactor temperature profile
In the context of reactor temperature profile control, multivariate
control schemes such as MPC, offer several benefits over manual/PI

70
F. Tahir et al. Control Engineering Practice 67 (2017) 64–75

Fig. 11. MPC controller response to temperature setpoint changes in Rattlesnake.

control. For instance, in the DN15 reactor, each circulator affects

product temperature across several straights. Therefore, adjustments
to only one of the circulators may correct temperature at the desired
straight but, at the same time, create disruption to the cooling profile
within the adjacent straights. As a result, modifications to temperature
within a single straight may require temperature changes for several
circulators. Similar arguments can be applied to the Rattlesnake reactor
(see Remark 3 below). Such a process with multivariable interactions is (a) Lactose concentration.
ideally suited to MPC as compared to many other traditional control
schemes. In this regard, as part of the offline development, MPC-
based control scheme also enables accurate comparison of the crystal
properties at different cooling profiles by delivering precise temperature
tracking control in a timely manner (see for example Fig. 11).
The proposed MPC control structures for the DN15 and Rattlesnake
reactor temperature profiles is given in Fig. 8.
To develop the temperature model in Fig. 8, a series of Pseudo
Random Binary Sequence (PRBS) tests were applied on the manipulated
variables (circulator temperatures and pump speeds). PRBS allows
statistically rich data to be obtained automatically and when used on
multiple actuators, the experimentation time can be reduced. Fig. 9
shows such a training dataset where PRBS was applied simultaneously
to multiple manipulated variables. The identified data-driven DN15 (b) Lactose spectra (900–12 000 cm−1 ).

temperature model, in terms of the step response from MVs to the CVs, is
shown in Fig. 10. In particular, Fig. 10(a) gives an overview of the model
structure. Then, Fig. 10(b) shows the close-up of the step responses
between the circulator temperatures (MVs) and thermocouple temper-
atures (CVs). This shows the aforementioned multivariable interactions
at play within the process (for example, adjusting circulator 2 affects
the temperature at thermocouples 3–6).
Remark 3. In comparison to other control algorithms such as PI
control, an MPC based advanced control scheme is more suited to the
considered continuous crystallisation process. This is due to the fact
both DN15 and Rattlesnake introduce plug flow with large residence
times (1 h+) . Furthermore, as discussed above, the process involves
actuator constraints and multivariable interactions. For instance, as
can been seen in Fig. 10, adjusting a single circulator setpoint in
Fig. 12. Training data for calibration model.
DN15 has an impact on the temperature in the adjoining straights.
Similarly, for Rattlesnake, the cold water flow in each of the four tubes is
manipulated from a single header and is therefore coupled to the others
(as shown in Fig. 2). These considerations make the implementation
of PI control scheme quite challenging whereas the MPC algorithm
is able to effectively deal with constraints and interactions within its
formulation (see Fig. 11).
The temperature control interval for both the Rattlesnake and DN15
was chosen to be 60 s. Fig. 11 shows the MPC based reactor temperature
control for the Rattlesnake unit. Notice that the MPC makes gentle

71
F. Tahir et al.
Fig. 13. Calibration model predictions for concentration.
Fig. 14. MPC control structure for concentration.
moves across several actuators to track the product temperature setpoint
changes.
6. Control of final product quality (CQA)
In a batch system, the crystallisation reaction is typically initiated
by introducing a seed slurry or by rapid cooling to induce nucleation.
Seeding is a one-time event that is significant in determining the final
particle size distribution. High seed loading results in the rapid onset
of nucleation and generation of fines. On the other hand, a low seed
loading in the solution may not be enough to induce crystal growth. With
continuous crystallisation, the opportunity exists to adjust seed-loading
to control final product quality. Similarly, with very short residence
times, the solution may not have enough time to crystallise, whilst
long residence time may affect the plug flow. Therefore, to achieve
the end product with multiple desired properties, it is advantageous to
manipulate different available parameters (e.g. seed and feed pumps).
The APC solution uses online PAT to obtain and infer in-line
measurements of crystal properties. Analytical instruments used in
this application include FTIR spectroscopy and FBRM. Such inline
methods allow Real-Time Release and/or reduce the requirement for
off-line product quality measurement. Moreover, inline measurements
play a fundamental role in product quality improvement through the
application of closed-loop control.
6.1. Real-time concentration prediction and control
In the APC solution, Super-saturation (concentration) is controlled
by adjusting seed and feed loading. To this end, both DN15 and
Rattlesnake were fitted with an FTIR spectrometer at an intermediate
position along the reactor. Chemometric techniques were then used
to develop a calibration model to predict the Lactose concentration in
solution.
A programmable experimental method was developed to allow the
generation of Meta-Stable Zone boundary information and spectral data
for calibration modelling. To minimise API consumption and devel-
opment time, the experimental method was implemented in a batch
version of the OBR. Solutions of concentration from 36% to 49% (grams
of Lactose/grams of water) were prepared for the experimental method
72
Control Engineering Practice 67 (2017) 64–75
(this is shown in Fig. 12(a)). The FBRM was applied in conjunction with
temperature cycling to determine the onset of nucleation and dissolution
at each concentration.
Secondly, the wavelength range specific to Lactose (900–1200 cm−1 )
was identified and selected for calibration modelling. Fig. 12 shows the
theoretical concentrations of Lactose (36%–49%) used in the experiment
(a) and recorded spectra (b).
Thirdly, spectra were pre-processed using Savitzky–Golay filter. The
filter with polynomial of 2nd degree and first derivative used ±5 samples
window. Finally, using the above shown data, a calibration model was
trained to predict the Lactose concentration using Partial Least Squares
(PLS) on the pre-processed spectra and reactor temperature.
PLS technique has been widely used in chemometrics to build
calibration models. The method offers reduction of variables whilst
capturing all the significant information within the entire signal range.
PLS parameter estimation was used to obtain a calibration model.
The model predicts the concentration of Lactose in a solution using
the FTIR spectrum as well as solution temperature. Information was
extracted from all wavelengths of a spectrum and solution temperature,
and stored in three latent variables. These variables explained 99%
of variability within the training data and, at the same time, were
correlated to Lactose concentration. Fig. 13 shows that there is a good
match between the designed calibration model predictions and the
theoretical concentrations used in the experiment.
Once the calibration model was commissioned in the APC platform,
the concentration control model (structure shown in Fig. 14) was identi-
fied by applying process response tests on Feed and seed pumps speeds,
respectively, as shown in Fig. 15. It can be seen that as the seed volume
is increased, the concentration of Lactose in the solution decreases.
On the other hand, there is a positive correlation between the feed
pump speed and concentration. These interactions were quantitatively
captured (using RLS) in the identified concentration control model
shown in Fig. 16. In particular, it shows the concentration response to
step changes in the seed pump and feed pump speeds, respectively.
Here, it is worth mentioning that the developed linear models were
adapted online in terms of the dead time (pure delay). This is because
when the MPC adjusts the seed and feed rates (in order to control to the
concentration setpoint), the process residence time changes. Therefore,
the APC scheme takes account of this by dynamically adapting the
time-delay within the concentration models for both the DN15 and
Rattlesnake.
6.2. Crystal size control
Offline product analysis gives an advantage of being able to obtain
more reliable Particle Size Distribution (PSD) estimates. However, lack
of instant results and inability to apply remedial changes to the process
is its biggest disadvantage. To this end, the proposed scheme uses
FBRM to estimate, online, the size and count of particles in a solution
F. Tahir et al.
Fig. 15. Feed and seed process response tests on the Rattlesnake.
Fig. 16. Rattlesnake concentration model coefficients.
Fig. 17. MPC control structure for particle size.
using scattered light. The measurements collected from FBRM can be
transformed into a Chord Length Distribution (CLD), which provides
information on size of crystals. Research shows that use of 10th,
50th and 90th percentile (D10, D50, D90) of that distribution can be
considered as relevant metrics for defining size of crystals. Furthermore,
sufficient representation of crystal size can be obtained from median
chord length (D50) alone. Therefore, this measurement was chosen to
be controlled by the third APC controller, which used concentration
of Lactose within solution as an actuator (as shown in the controller
structure in Fig. 17).
The FBRM was installed close to the outlet of each crystalliser. This
position meant that the solution was analysed near the end of the
cooling profile, where crystallisation should be at its final stage. Fig. 18
shows the control results with the cascade strategy, where a change in
setpoint of the median chord length (D50) prompts steady reduction of
73
Control Engineering Practice 67 (2017) 64–75
the Lactose concentration setpoint by the particle size controller. That
in turn causes the concentration controller to respond by increasing
the seeding, thereby achieving the lower Lactose concentration. The
APC scheme’s performance was evaluated by applying three changes
to the D50 setpoint and observing the CLD at steady-state (Fig. 19). It
is clear that the CLDs and their medians differ from each other at the
corresponding D50 setpoints. Furthermore, despite long delays in the
process (for example Fig. 18 shows the time range of over 5.5 h for two
setpoint changes), the APC solution is able to drive the crystallisation
reactor to achieve the required properties of crystals.
7. Conclusions
The paper described a transferable advanced control solution for
continuous crystallisation reactors. The solution was developed and
tested with both Cambridge Reactor Design’s Rattlesnake COBR as well
as Nitech’s DN15 COBR. The APC control was tested using Lactose as an
API, though it can easily be adapted to use other materials. Similarly,
the solution is transferable across continuous crystallisation reactors.
The proposed solution consisted of three MPC controllers. The first
one was a generic block, which is applied to control reactor temperature
in order to maintain the desired cooling profile. The other two con-
trollers were used, in a cascade manner, to drive the API concentration
and particle size (D50) towards the desired setpoints. Both of the
controllers incorporate PAT data, which allow real-time monitoring
and control of the product’s quality as well as rejection of any process
disturbances such as reactor fouling and API impurities. In particular,
spectral data collected from FTIR was used to build a calibration model,
which predicted the concentration of API in the solution. Concentration
of the API was then controlled by APC through manipulation of seed and
feed flows. Moreover, a successful control of the median Chord Length
F. Tahir et al. Control Engineering Practice 67 (2017) 64–75

Fig. 18. D50 control results on the DN15.

Fig. 19. CLDs from particle size control on DN15.

Distribution (D50), obtained from FBRM, was achieved by manipulating ∙ CLD: Chord Length Distribution
concentration of the API in real-time. ∙ COBR: Continuous Oscillatory Baffled Reactor
As mentioned above, the control scheme can easily be extended ∙ CQA: Critical Quality Attribute
to accommodate other APIs. It would then be possible to interchange ∙ CSTR: Continuous Stirred-tank Reactor
the API specific calibration and particle size models, according to ∙ CV: Controlled Variable
the manufacturing needs and production schedule. Furthermore, the ∙ FBRM: Focused Beam Reflectance Measurement
structural differences between various reactors can be compensated by
∙ FTIR: Fourier Transform Infrared
modifying/adapting time delays within the models in real time.
∙ MPC: Model Predictive Control
∙ MSZ: Meta-Stable Zone
Acknowledgement
∙ MV: Manipulated Variable
∙ OBR: Oscillatory Baffled Reactor
The authors gratefully acknowledge funding support from Innovate
UK under the project ‘MOPPS: Made To Order Processing Plants’ (grant ∙ OE: Output Error
number: 101334). ∙ PAT: Process Analytical Technology
∙ PLS: Partial Least Squares
Appendix. List of acronyms ∙ PRBS: Pseudo Random Binary Sequence
∙ PSD: Particle Size Distribution
∙ APC: Advanced Process Control ∙ QbD: Quality-by-Design
∙ API: Active Pharmaceutical Ingredient ∙ RLS: Recursive Least Square.

74
F. Tahir et al.
References
Cambridge reactor design. http://www.cambridgereactordesign.com/continuous-crystall
isation.html#.
Chiu, T., & Christofides, P. D. (1999). Nonlinear control of particulate processes. AIChE
Journal, 45(6), 1279–1297.
Forgione, M., Birpoutsoukis, G., Bombois, X., Mesbah, A., Daudey, P. J., & Van den Hof, P.
M. J. (2015). Batch-to-batch model improvement for cooling crystallization. Control
Engineering Practice, 41, 72–82.
Georgieva, P., Meireles, M. J., & Feyo de Azevedo, S. (2003). Knowledge-based hybrid
modelling of a batch crystallisation when accounting for nucleation, growth and
agglomeration phenomena. Chemical Engineering Science, 58(16), 3699–3713.
Goldfarb, D., & Idnani, A. (1983). A numerically stable dual method for solving strictly
convex quadratic programs. Mathematical Programming , 27 (1), 1–33.
Goodwin, G. C., & Sin, K. S. (2009). Adaptive filtering prediction and control. New York, NY,
USA: Dover Publications, Inc..
Kwon, J. S., Nayhouse, M., Christofides, P. D., & Orkoulas, G. (2014). Modeling and control
of crystal shape in continuous protein crystallization. Chemical Engineering Science,
107 , 47–57.
Li, M., Mhaskar, P., Shi, D., El-Farra, N. H., & Christofides, P. D. (2006). Predictive control
of particle size distribution in particulate processes. Chemical Engineering Science,
61(1), 268–281.
Maciejowski, J. M. (2002). Predictive control with constraints. Essex, England: Prentice Hall.
Mack, J. (2013). Pat based advanced process control and monitoring. In ISPE annual
meeting .
75
View publication stats
Control Engineering Practice 67 (2017) 64–75
Mayne, D. Q., Rawlings, J. B., Rao, C. V., & Scokaert, P. O. M. (2000). Constrained model
predictive control: Stability and optimality. Automatica, 36(6), 789–814.
McGlone, T., Briggs, N. E. B., Clark, C. A., Brown, C. J., Sefcik, J., & Florence, A. J. (2015).
Oscillatory flow reactors (OFRs) for continuous manufacturing and crystallization.
Organic Process Research and Development , 19(9), 1186–1202.
Nitech solutions. http://www.nitechsolutions.co.uk/products/dn15-standard/.
Orkoulas, G., Kwon, J. S., Nayhouse, M., & Christofides, P. D. (2014). Crystal shape
and size control using a plug flow crystallization configuration. Chemical Engineering
Science, 119, 30–39.
Perceptive engineering limited. http://www.perceptiveapc.com.
Powell, M. J. D. (1983). ZQPCVX a FORTRAN subroutine for convex quadratic programming .
Department of Applied Mathematics and Theoretical Physics, University.
Qin, S. J., & Badgwell, T. A. (2003). A survey of industrial model predictive control
technology. Control Engineering Practice, 11(7), 733–764.
Siddique, H., Brown, C. J., Houson, I., & Florence, A. J. (2015). Establishment of a
continuous sonocrystallization process for lactose in an oscillatory baffled crystallizer.
Organic Process Research and Development , 19(12), 1871–1881.
Siddique, H., Raval, V., Mabbott, F., Mack, J., Tahir, F., Krzemieniewska, K., Houson, I.,
Mercer, E., Florence, A. (2017). A direct design approach for ‘‘Dial a Particle size
capability’’ in continuous crystallisation of Lactose, submitted for publication.
Tahir, F., & Jaimoukha, I. M. (2013). Robust feedback model predictive control of
constrained uncertain systems. Journal of Process Control, 23, 189–200.
Zhao, L., Raval, V., Briggs, N. E. B., Bhardwaj, R. M., McGlone, T., Oswald, I. D. H., et
al. (2014). From discovery to scale-up: 𝛼-lipoic acid : nicotinamide co-crystals in
a continuous oscillatory baffled crystalliser. Crystal Engineering Communications, 16,
5769–5780.