Control Engineering Practice 78 (2018) 12–23

Contents lists available at ScienceDirect

Control Engineering Practice

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

Framework design for weight-average molecular weight control in

semi-batch polymerization
S.D. Salas a , N. Ghadipasha a , W. Zhu a , T. Mcafee b , T. Zekoski b , W.F. Reed b , J.A. Romagnoli a, *
a
Department of Chemical Engineering, Louisiana State University, Baton Rouge, LA 70803, USA
b Department of Physics and Engineering Physics, Tulane University, New Orleans, LA 70118, USA

ARTICLE INFO ABSTRACT

Keywords:
Nonlinear state estimation
Discrete-time extended Kalman filter
Weight-average molecular weight control
Semi-batch free-radical polymerization
Dynamic optimization
A framework that embraces a state-of-the-art sensor, multi-objective dynamic optimization, nonlinear state
estimation and control, is designed and implemented to achieve target weight-average molecular weight
trajectories. The Automatic Continuous Online Monitoring of Polymerization reactions (ACOMP) is combined
for the first time with a nonlinear state observer for full polymer characterization and signal processing. A
hybrid variation of the discrete-time extended Kalman filter (ℎ-DEKF) is formulated based on an auto-tuning
procedure that uses a stochastic global optimization technique. A number of optimal policies are generated and
experimentally tested. Results are provided through investigations into the free-radical aqueous polymerization
of acrylamide using potassium persulfate as initiator.

1. Introduction products such as thermal stability, strength, and capability of materials

The competitive landscape of the chemical industry requires prac-
tical tools for the Smart Manufacturing of chemicals with desired final
properties. Polymers represent a large part of this industry and their
synthesis involves complex processes. End manufacturers require raw
polymeric materials with uniform characteristics for industrial applica-
tions. In the past decades, techniques and strategies for state estimation
(Hedengren & Eaton, 2015; Kozub & MacGregor, 1992; Nicholson,
López-Negrete, & Biegler, 2014; Romagnoli & Sanchez, 1999; Schuler &
Schmidt, 1992; Srinivasan, Kasthurikrishnan, Cooks, Krishnan, & Tsao,
1995; Weiss, Romagnoli, & Islam, 1996) and control (Cho, Chung, &
Lee, 2000; Eaton & Rawlings, 1990; Ghadipasha, Romagnoli, Tronci, &
Baratti, 2015; Norquay, Palazoglu, & Romagnoli, 1998; Romagnoli &
Palazoglu, 2005) have evolved for solving industrial problems in chem-
ical reactors and unit operations. Advanced manufacturing techniques
are important as they allow the efficient use of feedstock materials,
energy, and labor while also improving safety in a chemical facility.
Particularly in the polymer industry the combination of mechanistic
model complexity and inadequate real-time characterization techniques
present crucial challenges.
A major goal in polymer synthesis is the achievement of a target
molecular weight distribution (MWD) as it leads to desired physico-
chemical properties. Indeed, MWD is fundamental in the sense that
it influences physical, rheological, and thermal characteristics of final
for being processed (Gentekos, Dupuis, & Fors, 2016; Heidemeyer &
Pfeiffer, 2002; Isayev, 2011). External disturbances such as temperature,
degree of mixing, concentration and purity of reactants affect final
MWDs due to the sensitive molecular structure of polymers (McKeen,
2014). The lack of reliable real-time techniques for monitoring MWD,
weight average molecular weight (𝑀𝑤 ) and number average molecular
weight (𝑀𝑛 ) makes it difficult to directly control these properties.
A recent state-of-the-art smart sensor for polymerization processes,
known as the ACOMP system, provides real-time measurements of
different polymer properties and follows the evolution of monomer
and polymer concentration along a liquid-phase polymerization (Flo-
renzano, Strelitzki, & Reed, 1998; McAfee et al., 2016; Reed & Alb,
2013). Some examples of the ACOMP applications include the moni-
toring of polyacrylamide synthesis (Giz, Catalgil-Giz, Alb, Brousseau,
& Reed, 2001), prediction of molecular weight in semi-batch free-
radical homopolymer reactors (Kreft & Reed, 2009a), and predictive
control of average composition and MWDs in semi-batch free-radical
copolymerization reactors (Kreft & Reed, 2009b). Although state-of-the-
art sensors such as the ACOMP system can overcome many needs, the
available measurements are not enough for a complete understanding of
the dynamic evolution of the system. Therefore, the implementation of
a nonlinear observer is necessary for better comprehension of the evolu-
tion of the state variables as well as to improve available measurements.

* Correspondence to: 3315N Patrick F. Taylor, Louisiana State University, Baton Rouge, LA 70803, USA.
E-mail address: jose@lsu.edu (J.A. Romagnoli).

https://doi.org/10.1016/j.conengprac.2018.06.004
Received 2 February 2018; Received in revised form 9 May 2018; Accepted 5 June 2018
0967-0661/© 2018 Elsevier Ltd. All rights reserved.
S.D. Salas et al. Control Engineering Practice 78 (2018) 12–23

Notation Greek letters

𝐴𝑑 Pre-exponential factor for initiator decomposition, 𝜆0 Zeroth moment, [mol m−3 ]

[min−1 ] 𝜆1 First moment, [mol m−3 ]
𝐴𝑝 Pre-exponential factor for propagation 𝜆2 Second moment, [mol m−3 ]
[m3 mol−1 min−1 ] 𝛼 Probability of propagation, dimensionless
𝐴𝑡𝑑 Pre-exponential factor for termination by dispropor- 𝜌𝑚 Mass density of monomer, [kg m−3 ]
tionation [m3 mol−1 min−1 ] 𝜌𝑖 Mass density of initiator, [kg m−3 ]
𝐶𝑖 Concentration of initiator inside the reactor, [mol 𝜌𝑠 Mass density of solvent, [kg m−3 ]
m−3 ]
𝐶𝑖𝑓 Concentration of initiator in the initiator flow rate,
Even though direct measurements from the ACOMP system provide
[mol m−3 ]
an accurate assessment of the system, the complete state vector can
𝐶𝑚 Concentration of monomer inside the reactor, [mol
seldom be measured. In addition, mathematical models often become
m−3 ]
increasingly inaccurate as the reaction proceeds. Thus, a well-adjusted
𝐶𝑚𝑓 Concentration of monomer in the monomer flow rate,
combination of both experimental measurements and mathematical
[mol m−3 ]
models allows a better understanding of the system dynamics. Devel-
𝐶𝑠 Concentration of solvent inside the reactor, [mol m−3 ]
oped more than half a century ago, the Kalman filter (KF) (Kalman,
𝐶𝑠𝑖𝑓 Concentration of solvent in the initiator flow rate,
1960; Kalman & Bucy, 1961) provides an elegant basis for incorporating
[mol m−3 ]
both the available data and a mechanistic model of the process of
𝐶𝑠𝑚𝑓 Concentration of solvent in the monomer flow rate,
interest (Qin, 2014). The KF addresses the problem of estimating states
[mol m−3 ]
of a discrete-time controlled process governed by a linear stochastic
𝐸𝑑 Activation energy in initiator decomposition, [J
difference equation. The assumptions of linearity for both the measure-
mol−1 ]
ments and the state transitions are crucial for its correctness (Simon,
𝐸𝑝 Activation energy in propagation, [J mol−1 ]
2006). Nonetheless, state transitions and measurements are rarely linear
𝐸𝑡𝑑 Activation energy in termination by disproportiona-
in practice. For nonlinear systems, the extended Kalman filter (EKF)
tion, [J mol−1 ]
is one the most broadly utilized state estimators for the chemical
𝑒𝑓 Initiator efficiency, dimensionless
process industry (Hashemi, Kohlmann, & Engell, 2016). A particular
𝐹𝑖 Flow rate of initiator to the reactor, [m3 min−1 ]
configuration of the EKF is the discrete-time extended Kalman filter
𝐹𝑚 Flow rate of monomer to the reactor, [m3 min−1 ]
(DEKF) (Simon, 2006). The method combines the nonlinear model and
𝐹𝑜𝑢𝑡 Extraction flow rate out of the reactor for sampling
its linearization to compute the state estimates.
purposes, [m3 min−1 ]
In polymeric systems, many authors have studied a variety of state
𝑔 Chain length distribution, dimensionless
estimators. Kozub and MacGregor (1992) considered different EKF
𝑘𝑑 Initiator decomposition rate constant, [min−1 ]
configurations for semi-batch polymerizations. Ellis, Taylor, and Jensen
𝑘𝑓 𝑚 Chain transfer to monomer rate constant,
(1994) implemented a MWD estimator for the methyl-methacrylate
[m3 mol−1 min−1 ]
using measurements from gel permeation chromatography (GPC) for
𝑘𝑓 𝑠 Chain transfer to solvent rate constant,
experimental testing. Tatiraju and Soroush (1997) compared the perfor-
[m3 mol−1 min−1 ] mance of an EKF with a nonlinear observer for a methyl-methacrylate
𝑘𝑝 Propagation rate constant, [m3 mol−1 min−1 ] system in a continuous reactor. Gentric, Pla, and Corriou (1997) and
𝑘𝑡𝑐 Termination by combination rate constant, Gentric, Pla, Latifi, and Corriou (1999) proposed an optimal policy
[m3 mol−1 min−1 ] formulation integrated with a geometric controller and an EKF for a
𝑘𝑡𝑑 Termination by disproportionation rate constant, batch emulsion-polymerization reactor. The trajectories followed by
[m3 mol−1 min−1 ] the controller were optimal temperature profiles that lead to desired
𝑚 Upper bound for chain length interval, dimensionless characteristics of the final polymer. Li, Corripio, Henson, and Kurtz
𝑀𝑛 Number average molecular weight, [kg mol−1 ] (2004) proposed a framework for online state estimation and parameter
𝑀𝑤 Weight average molecular weight, [kg mol−1 ] estimation. Statistical methods such as the particle filter have achieved a
𝑛 Lower bound for chain length interval, dimensionless robust performance in nonlinear state estimation (Hashemi et al., 2016).
𝑁𝑖 Total amount of initiator inside the reactor, [mol] Due to the difficulty of measuring the reactor contents, other authors
𝑁𝑖0 Initial amount of initiator inside the reactor, [mol] included the management of delayed and infrequent measurements
𝑁𝑚 Total amount of monomer inside the reactor, [mol] (Galdeano, Asteasuain, & Sánchez, 2011; Gopalakrishnan, Kaisare, &
𝑁𝑚0 Initial amount of monomer inside the reactor, [mol] Narasimhan, 2011). Beyer, Grote, and Reinig (2008) presented a com-
𝑁𝑚𝑓 Total amount of monomer added to the reactor from bined structure for control and state estimation using a sigma-point
the flow rates, [mol] KF for improving the control action in a Chylla–Haase polymerization
𝑁𝑠 Total amount of solvent inside the reactor, [mol] reactor. Although results exhibit satisfactory behavior and adequate
𝑁𝑠0 Initial amount of solvent inside the reactor, [mol] state estimation capability, some of them are purely simulation results
𝑃0 Concentration of live polymer in the reactor, [mol and others lack the advantages provided by the ACOMP system such as
m−3 ] directly measuring the control objective.
𝑅𝑔𝑎𝑠 Ideal gas constant, [J mol−1 K−1 ] In this contribution, a framework that combines a state-of-the-art
𝑇 Temperature of the reactor, [K] sensor, multi-objective dynamic optimization, nonlinear state estima-
𝑡 Time, [min] tion and control is proposed for a free-radical polymerization towards
𝑉 Volume of the material inside the reactor, [m3 ] full characterization and target polymer production. The ACOMP system
𝑤𝑚 Molecular weight of monomer, [kg mol−1 ] is combined for the first time with a nonlinear state observer, taking
𝑤𝑖 Molecular weight of initiator, [kg mol−1 ] advantage of its measuring ability in order to obtain better estimates
𝑤𝑠 Molecular weight of solvent, [kg mol−1 ] of the underlying model. As a first step, optimal policies are generated
𝑋 Monomer conversion, dimensionless by solving a multi-objective dynamic optimization problem targeting
polymers with different weight-average molecular weight trajectories

13
S.D. Salas et al. Control Engineering Practice 78 (2018) 12–23

and consequently different MWDs. A hybrid variation of the discrete- Weight average and number average molecular weight:
time extended Kalman filter (ℎ-DEKF) is formulated using a novel auto- 𝜆1 𝜆
tuned procedure based on searching model errors with a stochastic 𝑀𝑛 = 𝑤 𝑚 , 𝑀𝑤 = 𝑤𝑚 2 (12)
𝜆0 𝜆1
global optimization technique. Thus, the filter free-parameters are tuned
using the Parallel Local Metric Stochastic Radial Basis Function with Chain length distribution:
Restart (ParLMSRBF-R) algorithm (Regis & Shoemaker, 2007, 2009). [( ) ( ) ]
The tuning-optimization problem is proposed in such a way that the dif- 𝑑𝑔 (𝑚, 𝑛) 𝑘𝑝 𝑁𝑚 𝑃0 𝛼 + 𝑚 (1 − 𝛼) 𝛼 + (𝑛 + 1) (1 − 𝛼)
= 𝛼 𝑚−1 − 𝛼𝑛
ference between measured and estimated values is minimized depending 𝑑𝑡 𝜆1 𝑉 𝛼 𝛼
( )
on the individual characteristic of measurements. An in-house module, 𝑔 (𝑚, 𝑛) 𝑑 𝜆1 𝑉
formulated and implemented in a Python 2.7 environment, integrates − (13)
𝜆1 𝑉 𝑑𝑡
the measurements, nonlinear model, state estimation and control. This
platform allows full functionality and connectivity to the ACOMP server In this work, 20 chain lengths delimit the chain length intervals
to update and modify the process behavior. Different experiments were displayed in Eq. (13). Here, 𝑚 and 𝑛 represent the upper and lower bound
performed to test the response of a set of trajectories obtained during the of the intervals, and they are calculated using the arbitrary formula
from Crowley and Choi (1997b). For a more detailed description of the
dynamic optimization of the model. To demonstrate the effectiveness of
nonlinear model, kinetic and thermodynamic parameters please refer to
the module and the proposed framework for the Smart Manufacturing
our previous publication (Ghadipasha et al., 2016), and for variables and
of polymers, the polymerization of acrylamide in water solution using
constants description refer to the Notation and Greek Letters section.
potassium persulfate (KPS) as initiator is studied.

2.2. Dynamic optimization

2. Background

As a first step, offline dynamic optimization searches for optimal

2.1. Mathematical model
policies that lead to polymers with desired final characteristics. Follow-
ing the criteria from previous works (Ghadipasha et al., 2016, 2017),
The underlying mechanistic model refers to the free-radical poly-
gPROMS and more specifically its dynamic optimization tool gOPT
merization of acrylamide using KPS as initiator. The model has the
computes these policies. The gOPT function is convenient for real-
capability of describing the experimental system in a range of conditions
time applications as it uses a piecewise control variable definition,
and has adaptability to various applications, e.g., dynamic optimization,
permitting the selection of a time interval in which decision variables
nonlinear state estimation. In addition, it was previously tested and
are constant, and later updated in the following interval. To ensure
experimentally validated for applications including optimal operation
optimality the method applies adaptive control vector parameterization
(Ghadipasha et al., 2016) and control (Ghadipasha et al., 2017). The
(Schlegel, Stockmann, Binder, & Marquardt, 2005), which translates
computation of the MWD in semi-batch follows an analogous approach
the problem into a nonlinear programming problem by parametrizing
as proposed by Crowley and Choi (1997a). The set of dynamic equations
the control profiles. Eqs. (14) and (15) illustrates a general dynamic
are as follows:
optimization problem suitable for a number of specific trajectories of
𝑀𝑤 time-evolution.
𝑑𝑁𝑚 ( )
= − 𝑘𝑝 + 𝑘𝑓 𝑚 𝑃0 𝑁𝑚 + 𝐹𝑚 𝐶𝑚𝑓 − 𝐹𝑜𝑢𝑡 𝐶𝑚 (1)
𝑑𝑡 ( )2 ( )2 ( )2
𝑑𝑁𝑖 𝑋𝑓 𝑀𝑤,𝑓 ∑
𝑛𝑐
𝑔𝑖,𝑓
= −𝑘𝑑 𝑁 𝑖 + 𝐹𝑖 𝐶𝑖𝑓 − 𝐹𝑜𝑢𝑡 𝐶𝑖 (2) min 𝑤1 −1 + 𝑤2 −1 + 𝑤3 −1
𝑑𝑡 𝑡𝑓 ,𝑢(𝑡),𝑠(𝑡)𝑋𝑡 𝑀𝑤,𝑡 𝑖=1
𝑔𝑖,𝑡
𝑑𝑁𝑠 ( )
= −𝑘𝑓 𝑠 𝑁𝑠 𝑃0 +𝐹 𝑖 𝐶 𝑠𝑖𝑓 + 𝐹𝑚 𝐶𝑠𝑚𝑓 − 𝐹𝑜𝑢𝑡 𝐶𝑠 (3) + 𝑤4 𝑁𝑚𝑓 ,𝑓 (14)
𝑑𝑡
( )
𝑑 𝜆0 𝑉 ( ) Constrained to:
1
= 𝑘𝑓 𝑚 𝑁𝑚 + 𝑘𝑡𝑑 𝑃0 𝑉 + 𝑘𝑓 𝑠 𝑁𝑠 𝛼𝑃0 + 𝑘𝑡𝑐 𝑃02 𝑉 (4)
( 𝑑𝑡 ) 2
𝑑 𝜆1 𝑉 [( )( ) ] 𝑃0 ⎧𝑗 (𝑡 ) − 𝑗 = 0
= 𝑘𝑓 𝑚 𝑁𝑚 + 𝑘𝑡𝑑 𝑃0 𝑉 + 𝑘𝑓 𝑠 𝑁𝑠 2𝛼 − 𝛼 2 + 𝑘𝑡𝑐 𝑃0 𝑉 (5) ⎪ 𝑚𝑖𝑛0 0

(
𝑑𝑡
)
(1 − 𝛼) ⎪𝑡𝑓 ≤ 𝑡𝑓 ≤ 𝑡𝑚𝑎𝑥
𝑓
⎨ 𝑚𝑖𝑛 (15)
𝑑 𝜆2 𝑉 [( )( 3 ) 𝑚𝑎𝑥
= 𝑘𝑓 𝑚 𝑁𝑚 + 𝑘𝑡𝑑 𝑃0 𝑉 + 𝑘𝑓 𝑠 𝑁𝑠 𝛼 − 3𝛼 2 + 4𝛼 ⎪𝑢 ≤ 𝑢 (𝑡) ≤ 𝑢
𝑑𝑡 ⎪𝑠 ≤ 𝑠 (𝑡) ≤ 𝑠
𝑚𝑖𝑛 𝑚𝑎𝑥
⎩
] 𝑃0
+ 𝑘𝑡𝑐 𝑃0 𝑉 (𝛼 + 2) (6) where subscript 𝑓 stands for final value and 𝑡 for the target value of
(1 − 𝛼)2
the optimized variable. Notice that the first three terms in the objective
where, function seek to achieve target properties while the fourth term searches
the optimal use of monomer. The weighting factors (𝑤𝑖 ) vary depending
𝑁𝑚 = 𝐶𝑚 𝑉 , 𝑁𝑖 = 𝐶𝑖 𝑉 , 𝑁𝑠 = 𝐶𝑠 𝑉 on the particular MWD of interest which is obtained by following
𝑘𝑝 𝐶𝑚 specific 𝑀𝑤 trajectories (Regis & Shoemaker, 2009). In the constraints,
𝛼= (7) 𝑗0 symbolizes the initial conditions of the system, 𝑡𝑓 represents the
𝑘𝑝 𝐶𝑚 +𝑘𝑓 𝑚 𝐶𝑚 +𝑘𝑓 𝑠 𝐶𝑠 +𝑘𝑡𝑐 𝑃0 +𝑘𝑡𝑑 𝑃0
√ timehorizon, 𝑢(𝑡) the control variables subjected to the pilot plant
2𝑒𝑓 𝐶𝑖 𝑘𝑑 physical constraints, and 𝑠(𝑡) represents the time variant parameters
𝑃0 = (8)
𝑘𝑡𝑐 + 𝑘𝑡𝑑 also constrained to the pilot plant characteristics, e.g., volume of the
contents inside the reactor. Even though temperature is an important
𝜌𝑠 = −0.0031𝑇 2 − 0.1467𝑇 + 1003 (9) control variable, which influences the MWD and subsequently the final
𝜆 𝑉 𝑤𝑚 𝑁𝑚 𝑤𝑚 𝑁𝑠 𝑤𝑠 𝑁𝑖 𝑤𝑖 polymer properties, the slow dynamics of the cooling/heating system
𝑉 = 1 + + + (10)
𝜌𝑝 𝜌𝑚 𝜌𝑠 𝜌𝑖 endorses an isothermal operation. However, the optimal operational
temperature set point is obtained simultaneously while calculating the
Conversion:
other optimal time-varying control variables.
𝑡 𝑡
Recipes for three optimal semi-batch experiments corresponding to
𝑁𝑚0 + ∫0 𝐹𝑚 𝐶𝑚𝑓 𝑑𝑡 − 𝐶𝑚 𝑉 − ∫0 𝐹𝑜𝑢𝑡 𝐶𝑚 𝑑𝑡 the increasing, constant, and decreasing 𝑀𝑤 time-evolution are gener-
𝑋= 𝑡
(11)
𝑁𝑚0 + ∫0 𝐹𝑚 𝐶𝑚𝑓 𝑑𝑡 ated to test the flexibility of the proposed framework. Fig. 1 illustrates

14
S.D. Salas et al. Control Engineering Practice 78 (2018) 12–23

Table 1
Experiment initial conditions for different trajectories.
Trajectory description 𝑁𝑖0 [mol] 𝑁𝑚0 [mol] 𝑁𝑠0 [mol] 𝐶𝑖𝑓 [mol m−3 ] 𝐶𝑚𝑓 [mol m−3 ]
Increasing 𝑀𝑤 0.008 0.05 30.0 3.69932 1406.866
Constant 𝑀𝑤 0.008 0.05 30.0 3.69932 1406.866
Decreasing 𝑀𝑤 0.008 0.10 35.0 3.69932 1406.866

∙ Derive the Jacobian matrix of partial derivatives for the model,

linearize around its previous a posteriori estimate and approxi-
mate it to a discrete-time approach (Hashemi & Engell, 2016).

( )
𝜕𝑓 ||
𝑭𝑘−1 = 𝑒𝑥𝑝 ⋅ 𝛥𝑡𝛼 (18)
𝜕𝑥 ||𝑥̂ +
𝑘−1

∙ Calculate the time update of the a priori estimates and covariance

error matrix.

𝐱̂ 𝑘− = 𝑓𝑘−1 (𝐱̂ 𝑘−1

+
, 𝒖𝑘−1 , 0) (19)

𝑷𝑘− = 𝐅𝑘−1 𝑷𝑘−1

+
𝐅𝑇𝑘−1 + 𝐐 (20)
∙ Compute the Jacobian matrix of partial derivatives for the mea-
surements and linearize around its current a priori states.
𝜕ℎ𝑘 ||
𝑯𝑘 = (21)
𝜕𝑥 ||𝑥̂ −
𝑘

∙ Calculate the measurement update of the estimates and a posteri-

Fig. 1. Final chain length distribution of optimal trajectories.
target. MWDs for the different 𝑀𝑤 trajectories. Table 1 provides details
on the initial experimental set up as well as the concentration of
monomer and initiator in the inlet flows.
2.3. Nonlinear state estimation with auto-tuned error-driven initialization
For the formulation of the nonlinear state estimation, the mecha-
nistic model described before can be represented in a discrete compact
form:
( )
𝒙𝒌 = 𝑓𝑘−1 𝒙𝒌−𝟏 , 𝒖𝒌−𝟏 , 𝒘𝒌−𝟏 𝒘𝒌 ∼ 𝑁[0, 𝑸] (16)
( )
𝒚𝑘 = ℎ𝑘 𝒙𝑘 , 𝒗𝑘 𝒗𝑘 ∼ 𝑁[0, 𝑹] (17)
[ ]𝑇
where, 𝑘 is the discrete time, 𝒙 = 𝑁𝑚 , 𝑁𝑖 , 𝑁𝑠 , 𝜆0 𝑉 , 𝜆1 𝑉 , 𝜆2 𝑉 repre-
[ ]𝑇
sents the state vector, 𝒖 = 𝐹𝑚 , 𝐹𝑖 , 𝑇 denotes the manipulated variables
[ ]𝑇
vector, 𝒚 = 𝐶𝑚 , 𝑀𝑤 , 𝑉 is the measurements vector, and 𝑓 (⋅) and ℎ (⋅)
are the nonlinear model and measurements, respectively. It is assumed
that states and measurements have uncorrelated, zero mean and white
Gaussian noise, w and v, with covariance error 𝑸 and 𝑹, respectively.
A hybrid variation of the DEKF proposed by Simon (2006) serves as the
algorithm for nonlinear state estimation. The method is purely recursive
and starts after reaching a threshold iteration (𝑘𝛼 ) determined by the
( )
filter initialization time, 𝑡𝛼 = 𝑘𝛼 + 1 𝛥𝑡𝛼 .
First step for𝑘 < 𝑘𝛼 : Model Initialization, the model from Eq. (16)
initializes with 𝒙𝟎 , and measured control variables.
Second step for𝑘 = 𝑘𝛼 : ℎ-DEKF Initialization, the filter free-parameters
are set based on model and measurement errors. Thus, matrix 𝑹 contains
the diagonal of the observations error vector (𝝈 𝑣 2 ), 𝑸 the diagonal of
the model error vector (𝝈 𝑤 2 ) which is the product between the current
states and its relative errors (𝝈 𝑤 = 𝒙𝑘𝛼 𝒆𝑤 ). The initial covariance error
matrix 𝑷0 is the product of 𝑸 and an arbitrary constant.
Third step for𝑘 > 𝑘𝛼 : ℎ-DEKF, the superscripts ‘‘+’’ and ‘‘−’’ denote
a priori and a posteriori states, respectively. A priori refers when the
calculation is done before 𝒚𝑘 is considered, and a posteriori after 𝒚𝑘 is
considered. The following occurs:
ori covariance error. K is the Kalman gain matrix and computes
the amount of correction to incorporate or take from the states on
their a priori estimation. The stability of the ℎ-DEKF is evaluated
through the eigenvalues of the estimation error dynamics (𝑭 −
𝑲𝑯) (Bonniwell, Schneider, & Yaz, 2015).
( )−1
𝑲𝑘 = 𝑷𝑘− 𝑯𝑘𝑇 𝑯𝑘 𝑷𝑘− 𝑯𝑘𝑇 + 𝑹 (22)
[ ]
𝐱̂ 𝑘+ = 𝐱̂ 𝑘− + 𝐊𝑘 𝒚𝑘 − ℎ𝑘 (𝐱̂ 𝑘− , 0) (23)
( )
𝑷𝑘+ = 𝐈 − 𝐊𝑘 𝐇𝑘 𝑷𝑘− (24)
The integration in real-time of the nonlinear model for computing
the state estimates, and its linear approximation for calculating the
covariance error and filter gain matrices provides accuracy and ro-
bustness of the nonlinear state estimation. The advantages of the ℎ-
DEKF encompass the nature of estimation and the predictive–corrective
ability. The recursive characteristic allows rapid estimation in real-time,
which is mandatory for online deployment. However, a disadvantage
is that it lacks the ability of taking into account bounds on process
variables and other constraints (Haseltine & Rawlings, 2005).
To improve the filter action, an adequate tuning of its
free-parameters is required. The tuning problem addresses the search
of the 𝒆𝑤 that minimizes the squared difference between estimated
and noise-reduced measurements (𝑦̄𝑘 ) depending on the measurement
covariance error matrix. In general, the tuning procedure follows a more
systematic methodology that uses model and measurement errors (Leu
& Baratti, 2000). The tuning cost function presented in Eq. (25) follows
the criteria from Chan, Ngan, Fung, and Rad (2001) and Kozub and
MacGregor (1992). In addition, to avoid unreal kinetics and unstable
behavior, the problem is constrained to only positive values of the a
posteriori estimates.
∑
𝑘
( ) ( )𝑇
min 𝒚̄𝑘 − 𝒉𝑘 (𝒆𝑤 ) 𝑹−1 𝒚̄𝑘 − 𝒉𝑘 (𝒆𝑤 ) (25)
𝑘=1
Constrained to:
0 < 𝐱̂ 𝑘+ (26a)
0 ≤ 𝒆𝑤 ≤ 0.08 (26b)
The algorithm to tune the ℎ-DEKF is the Parallel Local Metric
Stochastic Radial Basis Function with Restart (ParLMSRBF-R) algorithm

15
S.D. Salas et al. Control Engineering Practice 78 (2018) 12–23

Fig. 3. Framework of the integration of experimental apparatus and computa-

tional tools.

Fig. 2. Laboratory pilot plant setup.

(Regis & Shoemaker, 2007, 2009) and was implemented following the
criteria from Salas, Geraili, and Romagnoli (2017). The algorithm uses
a surrogate model, e.g. radial basis function, and evaluates multiple
points simultaneously. When no improvement is attested, the algorithm
restarts to avoid local optima. In summary, the algorithm follows a
master–worker criterion, assuming that 𝛾 processors are available and
two function evaluations consume the same computational time. A set of
initial points are generated by a space filling (or other similar) technique
to evaluate the cost function. The surrogate model is initially fitted and
then updated using the output(s) from each iteration. ParLMSRBF-R
performs exploitation of the solution domain by keeping track of the
consecutive failed 𝐶𝑓 𝑎𝑖𝑙 , and successful 𝐶𝑠𝑢𝑐𝑐𝑒𝑠𝑠 , iterations. When 𝐶𝑓 𝑎𝑖𝑙
or 𝐶𝑠𝑢𝑐𝑐𝑒𝑠𝑠 exceed a predefined tolerance value, the step size is reduced Fig. 4. Flow diagram of the functionality of the module.
by half or doubled, respectively. Later, the recorded values of 𝐶𝑓 𝑎𝑖𝑙 and
𝐶𝑠𝑢𝑐𝑐𝑒𝑠𝑠 are reset.
nm determine the monomer and polymer concentrations, and the 𝑀𝑤 is
3. Experimental design calculated from the static/multi-angle light scattering data. Air bubbles
are purged from the ACOMP sensors to guarantee uncontaminated and
3.1. Experimental equipment & setup stable measurements.
Once the experiments finish, the Gel Permeation Chromatography
Fig. 2 illustrates a diagram of the polymerization pilot plant ex- (GPC) method provides the final MWD using an in-house built system.
hibiting its main components and functionality. The experimental It consists of a Shimadzu LC-10ADVp (Columbia, MD) highpressure sol-
equipment consists of a 1.5 L reactor including auxiliary fittings and vent delivery pump that provides a continuous flow of sodium chlorine
instrumentation monitored by the ACOMP system. Pumps driven by solution through the GPC column followed by a Shimadzu RID-10A
encoded stepper motors inject monomer and initiator solutions into the differential refractometer. The sample injector is an IDEX/Rheodyne
reactor. An external jacket keeps the inner temperature of the reactor MX-II with a 50 μL sample loop volume. A series of polyethylene oxide
constant. From a high flow rate recirculation loop (∼40 mL/min), the standards ranging from 25,000 to 1 million g/mol are utilized to create
ACOMP extracts a constant sample stream. For pilot scale, a rate of 0.5 a standard column calibration. The described method is referred as GPC
mL/min is sufficient. This setup allows minimal time delay from the standard calibration.
time new materials are added into the reactor until they are detected
by the sensor. The ACOMP analyzes the sample by diluting it 80 3.2. Framework & module functionality
times with deionized water and homogenizing it in a mixing chamber.
Ultraviolet visible absorption spectroscopy, viscometry and multiangle As discussed before, the proposed framework includes optimal recipe
laser light scattering detectors measure the sample. Spectroscopy at 245 formulation, nonlinear state estimation, and feedback control using the

16
S.D. Salas et al. Control Engineering Practice 78 (2018) 12–23

Fig. 5. (a) ParLMSRBF -R convergence profile searching optimal 𝒆𝑤 for auto-tuned error-driven free-parameters (b) Eigenvalues of the estimation error dynamics for
different OL experiments, where increasing, constant and decreasing 𝑀𝑤 trajectory.

Fig. 6. Results for the increasing 𝑀𝑤 trajectory OL. (a) 𝑀𝑤 and 𝑀𝑛 time evolution, (b) final chain length distribution estimated by ℎ-DEKF and experimental results,
(c) estimated chain length distribution along the reaction.

same methodology presented by Ghadipasha et al. (2017). The final
aim is to achieve full polymer characterization and subsequent mon-
itoring while improving the weight-average molecular weight control
action in order to achieve target properties. Fig. 3 presents different
components of the framework including experimental equipment and
the online/offline computational tools. Preliminary tasks include the
generation of optimal recipes by minimizing a multi-objective dynamic
optimization problem using gPROMS, and the tuning of the ℎ-DEKF free-
parameters using the ParLMSRBF-R algorithm as described previously.
Once the experiment starts, the nonlinear model computes states solely
until the filter/controller initialization is reached (after 20 min in our
practice in order to allow the sensor to settle down). Once the ℎ-DEKF

17
S.D. Salas et al.
Fig. 7. Results for the constant 𝑀𝑤 trajectory OL. (a) 𝑀𝑤 and 𝑀𝑛 time evolution, (b) final chain length distribution estimated by ℎ-DEKF and experimental results,
(c) estimated chain length distribution along the reaction.
is activated, estimated states from the nonlinear estimator are available
and the control action starts. For closed loop experiments, the controller
updates the process behavior in each control interval (𝛥𝑡𝛽 ). Finally, at
the end of each experiment, a sample of the final product is taken to
measure its MWD using the GPC standard calibration.
The filter/controller module was implemented in the Python 2.7
environment allowing full connectivity with the ACOMP server. Fig. 4
explains the functionality of the module. A cascade structure improves
the resolution of the filter, which consequently guarantees an adequate
control action. The filter remains in the inner loop at a higher frequency
than the outer loop where the control action takes place. The ACOMP
system updates measurements every 1 s; thus, the relation 1 s < 𝛥𝑡𝛼 <
𝛥𝑡𝛽 must occur for the satisfactory operation of the module.
To verify the efficacy and robustness of the proposed framework
and module, open-loop (OL), PID only and PID coupled with ℎ-DEKF
experiments are carried out. Experiments took place at the Tulane Center
for Polymer Reaction Monitoring and Characterization.
4. Results & discussion
4.1. System stability & ℎ-DEKF tuning
Following an auto-tune error-driven strategy to set the ℎ-DEKF free-
parameters, optimal 𝒆𝒘 are searched using the ParLMSRBF-R algorithm
and real data from one OL experiment. The algorithm evaluates a total
of 512 points using Eqs. (25) and (26). Sixteen promising candidate
18
Control Engineering Practice 78 (2018) 12–23
points are evaluated simultaneously in each iteration taking advantage
of parallel computing. Fig. 5a shows the convergence profile of the
cost function. The eigenvalues of the error dynamics matrix allow us
to analyze the filter stability. The stability analysis is required for all
proposed trajectories in order to generalize the validity of the method
for closed-loop experiments. Fig. 5b shows the eigenvalues distribution
for all trajectories in a realimaginary plane. All eigenvalues fall inside
the unitary circle and show stable behavior in all trajectories making
closed-loop experiments promising.
𝝈 𝑣 2 = [5, 100, 10−12 ]
𝒆𝒘 = [0.08, 0.08, 0.0356, 0.0, 0.0, 0.0056]
𝑷0 = 100𝑸
4.2. Open-loop experiments
OL experiments permit the verification of the ℎ-DEKF performance
using real data from the ACOMP system. During these experiments,
the process behavior follows strictly the trajectories obtained by the
offline dynamic optimization. Thus, for isothermal operation 𝐹𝑚 and 𝐹𝑖
are set equal to the optimal profiles provided by gPROMS considering
variations every 2.5 min. The nonlinear model initializes while the
reaction starts in the pilot plant. After 20 min, nonlinear state estimation
starts. Figs. 6–8 illustrate the ℎDEKF action for all OL scenarios. Each
S.D. Salas et al.
Fig. 8. Results for the decreasing 𝑀𝑤 trajectory OL. (a) 𝑀𝑤 and 𝑀𝑛 time evolution, (b) final chain length distribution estimated by ℎ-DEKF and experimental results,
(c) estimated chain length distribution along the reaction.
figure includes estimated properties, experimental values and optimal
trajectories. Figs. 6a, 7a & 8a illustrate relevant information related
to the filter competency for state estimation and signal processing.
Estimated 𝑀𝑤 points are provided with raw measurements to illustrate
the noise reduction features of the filter. In addition, the ℎDEKF has the
ability of estimating the time-evolution of 𝑀𝑛 which is an unmeasured
property. The estimated 𝑀𝑤 and 𝑀𝑛 are drawn in parallel with respect
to their optimal trajectories, represented with dashed lines, showing
physical consistency between estimated and calculated values.
Another important aim of the module is the monitoring of the MWD
during polymerization. Figs. 6b, 7b & 8b illustrate a comparison of the
estimated and measured MWD final value. The experimental values
obtained using the standard calibration GPC method show adequate
estimation ability even though they seem to not capture the initial low
mass section and the long high mass tail estimated due to equipment and
column calibration limitations. The experimental MWD is acquired after
the experiment finalizes, taking between 15 to 30 min to obtain experi-
mental results. In addition, Figs. 6c, 7c & 8c depict the predicted MWD
time-evolution, which evidences the real-time monitoring capability of
proposed framework.
In summary, all the evaluated trajectories show good performance
in state estimation and signal processing. Once the filter switches
on, estimated properties move from the nonlinear model towards the
measurements and provide a smoother projection of a given polymer
property while reducing noise. Regarding the online prediction ability
of the MWD, the results are good in predicting the trends of this
19
Control Engineering Practice 78 (2018) 12–23
critical property. The capability of noise reduction is important towards
enhancing the control action, which is the scope of the next section.
4.3. Closed-loop experiments
After validating the ℎ-DEKF performance in OL experiments, the
weight-average molecular weight control response is tested in two
sets of closed loop experiments, using only PID control and the PID
integrated with the filter. The control objective is to follow 𝑀𝑤 optimal
set point trajectories by manipulating the 𝐹𝑚 . Hence, 𝐹𝑖 is equal to
its optimal profile trajectory in all cases. Small variations of 𝐹𝑚 along
the optimal input trajectories should only be necessary to achieve the
target value. In the first set of experiments (PID), the controller collects
raw 𝑀𝑤 measurements and computes an average value for a 30 s time
interval. The mentioned procedure intents to reduce noise from the
data by using a simple average. The second set of experiments (PID
+ ℎ-DEKF) proceed as explained in Section 3.2 (Framework & Module
Functionality). In this case, the estimated 𝑀𝑤 are the controller inputs.
Nonlinear state estimation and control action initialize 20 min after the
reaction starts.
Figs. 9–11 contrast the performance of OL, PID, and PID + ℎ-
DEKF experiments under equal operating conditions. In general, results
show the performance of three operating philosophies towards standard-
ized polymers. These setups include OL optimal operation (following
a recipe), PID control, and PID control combined with a nonlinear
estimator. Figs. 9a, 10a & 11a reveal raw 𝑀𝑤 measurements, estimated
S.D. Salas et al. Control Engineering Practice 78 (2018) 12–23

Fig. 9. Experimental results for increasing 𝑀𝑤 trajectory closed-loop. (a) 𝑀𝑤 and 𝑀𝑛 in different experiments, (b) Measured flow rate of monomer in different
experiments, (c) Final chain length distribution: dynamic optimization, estimated by ℎ-DEKF and measured by standard GPC, (d) MWD evolution estimated values.

𝑀𝑤 and 𝑀𝑛 (computed by the filter for PID + ℎ-DEKF experiments), Table 2

and set-point trajectories. In terms of performance, the PID + ℎ- PID + ℎ-DEKF control action improvement.
DEKF shows a better response than the PID alone, improving not only Trajectory description Improvement after:
the 𝑀𝑤 trajectory as compared to the optimal set-point, but also in 20 min 90 min 120 min
providing smoother responses in terms of the manipulated variables. Increasing 𝑀𝑤 50.5% 69.2% 74.0%
Figs. 9b, 10b & 11b display the measured values of 𝐹𝑚 (manipulated Constant 𝑀𝑤 44.6% 11.2% 25.1%
variable). The sum of the absolute difference between the set point and Decreasing 𝑀𝑤 56.5% 24.9% 18.2%

raw measurements computed every one second permits to quantify the

control action improvement of the PID + ℎ-DEKF. Table 2 denotes the
percentages of improvement considering three different time intervals:
after initialization (after 20 min), for the second half (after 90 min) and
for the last third of the experiment (after 120 min).
Figs. 9c, 10c & 11c compare the target, estimated and measured
values of the final MWD. In all scenarios, the filter demonstrates good
prediction ability, especially for the constant 𝑀𝑤 trajectory. The ℎ-DEKF
provides extra information on the MWD time-evolution as observed in
Figs. 9d, 10d & 11d indicating the monitoring ability of the framework
as well as illustrating the variation of this property along the semi-batch
run. In summary, signal processing or measurements reconciliation
represents a remarkable advantage in terms of control performance
while nonlinear state estimation provides the opportunity of achieving
full polymer characterization in real-time.
5. Conclusions & future work
Advanced control and online monitoring of polymerization reactions
are noteworthy areas of interest for the industry and academia. This

20
S.D. Salas et al.
Fig. 10. Experimental results for constant 𝑀𝑤 trajectory closed-loop. (a) 𝑀𝑤 and 𝑀𝑛 in different experiments, (b) Measured flow rate of monomer in different
experiments, (c) Final chain length distribution: dynamic optimization, estimated by ℎ-DEKF and measured by standard GPC, (d) MWD evolution, estimated
values.
paper addressed some of the challenges that the polymer industry
faces during product synthesis and operation. With that determination,
a dynamic filter for nonlinear systems using state-of-the-art polymer
characterization tools provided the basis for enhanced control and
comprehensive polymer monitoring. In particular, a framework that
integrates dynamic optimization, an auto-tuned error-driven ℎ-DEKF
as nonlinear state estimator and feedback control using online mea-
surements from the ACOMP system was implemented for full polymer
characterization and to achieve weigh-average molecular weight control
in free-radical polymerizations.
The dynamic optimization of a multi-objective problem provided
optimal trajectories/recipes for a given target. The performance of the
ℎ-DEKF improved by using a stochastic global optimization technique,
which minimizes a cost function that includes the covariance error ma-
trix of the measurements and the squared difference between estimated
and measured values. An in-house filter/controller module, formulated
and implemented in a Python 2.7 environment, was incorporated into
21
Control Engineering Practice 78 (2018) 12–23
the ACOMP system server providing full functionality and connectivity
as well as expanding the sensor capabilities.
The proposed framework was tested experimentally under a range
of possible scenarios. Overall, the validation experiments highlighted its
excellent performance and robustness for different operating conditions.
The ACOMP system provided real-time data of polymer properties,
which constitutes the basis for implementing a nonlinear state esti-
mation strategy towards full real-time polymer characterization. The
ability of improving the quality of the raw noisy online data allowed the
improvement of the control action providing a new path to formulate
and achieve polymers with target properties.
Acknowledgments
The information, data, or work presented herein was funded in part
by the Office of Energy Efficiency and Renewable Energy (EERE), U.S.
S.D. Salas et al.
Fig. 11. Experimental results for decreasing 𝑀𝑤 trajectory closed-loop. (a) 𝑀𝑤 and 𝑀𝑛 in different experiments, (b) Measured flow rate of monomer in different
experiments, (c) Final chain length distribution: dynamic optimization, estimated by ℎ-DEKF and measured by standard GPC, (d) MWD evolution, estimated
values.
Department of Energy, under Award Number DE-EE0005776. Addi-
tional support came from the National Science Fundation (NSF) EPS-
1430280 and the Louisiana Board of Regents. The authors acknowledge
support from Mr. Phong Pham for automation and integration and Mr.
Colin Bellmore for software development.
Disclaimer. The information, data, or work presented herein was funded
in part by an agency of the United States Government. Neither the
United States Government nor any agency thereof, nor any of their em-
ployees, makes any warranty, express or implied, or assumes any legal
liability or responsibility for the accuracy, completeness, or usefulness of
any information, apparatus, product, or process disclosed, or represents
that its use would not infringe privately owned rights. Reference herein
to any specific commercial product, process, or service by trade name,
trademark, manufacturer, or otherwise does not necessarily constitute
or imply its endorsement, recommendation, or favoring by the United
States Government or any agency thereof. The views and opinions of
22
Control Engineering Practice 78 (2018) 12–23
authors expressed herein do not necessarily state or reflect those of the
United States Government or any agency thereof.
Appendix A. Supplementary data
Supplementary material related to this article can be found online at
https://doi.org/10.1016/j.conengprac.2018.06.004.
References
Beyer, M. A., Grote, W., & Reinig, G. (2008). Adaptive exact linearization control of batch
polymerization reactors using a Sigma-Point Kalman Filter. Journal of Process Control,
18(7), 663–675.
Bonniwell, J. L., Schneider, S. C., & Yaz, E. E. (2015). H∞-property of the discrete-time
extended Kalman filter with stochastic 𝓁2 disturbances. In IEEE decis contr P. Vol. 54
(pp. 6992–6997).
Chan, Z., Ngan, H. W., Fung, Y. F., & Rad, A. B. (2001). An advanced evolutionary
algorithm for parameter estimation of the discrete Kalman filter. Computer Physics
Communications, 142, 248–254.
S.D. Salas et al.
Cho, H. S., Chung, J. S., & Lee, W. Y. (2000). Control of molecular weight distribution for
polyethylene catalyzed over ziegler-natta/metallocene hybrid and mixed catalysts.
Journal of Molecular Catalysis A - Chemical, 159, 203–213.
Crowley, T. J., & Choi, K. Y. (1997a). Calculation of molecular weight distribution from
molecular weight moments in free radical polymerization. Industrial & Engineering
Chemistry Research, 36, 1419–1423.
Crowley, T. J., & Choi, K. Y. (1997b). Discrete optimal control of molecular weight
distribution in a batch free radical polymerization process. Industrial & Engineering
Chemistry Research, 36, 3676–3684.
Eaton, J. W., & Rawlings, J. B. (1990). Feedback-control of chemical processes using online
optimization techniques. Computers & Chemical Engineering , 14, 469–479.
Ellis, M. F., Taylor, T. W., & Jensen, K. F. (1994). Online molecular weight distribution
estimation and control in batch polymerization. AIChE Journal, 40, 445–462.
Florenzano, F. H., Strelitzki, R., & Reed, W. F. (1998). Absolute, on-line monitoring of
molar mass during polymerization reactions. Macromolecules, 31, 7226–7238.
Galdeano, R., Asteasuain, M., & Sánchez, M. C. (2011). Unscented transformation-based
filters: performance comparison analysis for the state estimation in polymerization
processes with delayed measurements. Macromolecular Reactor Engineering , 5(7–8),
278–293.
Gentekos, D. T., Dupuis, L. N., & Fors, B. P. (2016). Beyond dispersity: deterministic control
of polymer molecular weight distribution. Journal of the American Chemical Society,
138(6), 1848–1851.
Gentric, C., Pla, F., & Corriou, J. P. (1997). Experimental study of the nonlinear geometric
control of a batch emulsion polymerization reactor. Computers & Chemical Engineering ,
21, S1043–S1048.
Gentric, C., Pla, F., Latifi, M. A., & Corriou, J. P. (1999). Optimization and non-linear
control of a batch emulsion polymerization reactor. Chemical Engineering Journal,
75(1), 31–46.
Ghadipasha, N., Geraili, A., Romagnoli, J. A., Castor, C. A., Drenski, M. F., & Reed, W. F.
(2016). Combining on-line characterization tools with modern software environments
for optimal operation of polymerization processes. Processes, 4(1), 5.
Ghadipasha, N., Romagnoli, J. A., Tronci, S., & Baratti, R. (2015). On-line control of crystal
properties in nonisothermal antisolvent crystallization. AIChE Journal, 61(7), 2188–
2201.
Ghadipasha, N., Zhu, W., Romagnoli, J. A., McAfee, T., Zekoski, T., & Reed, W. F.
(2017). Online optimal feedback control of polymerization reactors: application to
polymerization of acrylamide-water-potassium persulfate (KPS) system. Industrial and
Engineering Chemistry Research, 56, 7322–7335.
Giz, A., Catalgil-Giz, H., Alb, A. M., Brousseau, J. L., & Reed, W. F. (2001). Kinetics
and mechanisms of acrylamide polymerization from absolute, online monitoring of
polymerization reaction. Macromolecules, 34, 1180–1191.
Gopalakrishnan, A., Kaisare, N. S., & Narasimhan, S. (2011). Incorporating delayed and
infrequent measurements in extended Kalman filter based nonlinear state estimation.
Journal of Process Control, 21(1), 119–129.
Haseltine, E. L., & Rawlings, J. B. (2005). Critical evaluation of extended Kalman filtering
and moving-horizon estimation. Industrial & Engineering Chemistry Research, 44(8),
2451–2460.
Hashemi, R., & Engell, S. (2016). Effect of sampling rate on the divergence of the extended
Kalman filter for a continuous polymerization reactor in comparison with particle
filtering. IFAC PapersOnLine, 49(7), 365–370.
Hashemi, R., Kohlmann, D., & Engell, S. (2016). Optimizing control and state estimation of
a continuous polymerization process in a tubular reactor with multiple side-streams.
Macromolecular Reactor Engineering , 10, 415–434.
Hedengren, J. D., & Eaton, A. N. (2015). Overview of estimation methods for industrial
dynamic systems. Optimization Engineering , 1–24.
Heidemeyer, P., & Pfeiffer, J. (2002). Special requirements on compounding technology
for bimodal polyolefines and their industrial application. In Macromol symp. Vol. 181
(pp. 167–176).
Isayev, A. I. (2011). Encyclopedia of polymer blends, Volume 2: Processing. Wiley.
Kalman, R. E. (1960). A new approach to linear filtering and prediction problems. Journal
of Basic Engineering - Transactions of the ASME, 82(1), 35–45.
23
Control Engineering Practice 78 (2018) 12–23
Kalman, R. E., & Bucy, R. S. (1961). New results in linear filtering and prediction theory.
Journal of Basic Engineering - Transactions of the ASME, 83(3), 95–108.
Kozub, D. J., & MacGregor, J. F. (1992). State estimation for semi-batch polymerization
reactors. Chemical Engineering Science, 47 (5), 1047–1062.
Kreft, T., & Reed, W. F. (2009a). Predictive control and verification of conversion kinetics
and polymer molecular weight in semi-batch free radical homopolymer reactions.
European Polymer Journal, 45(8), 2288–2303.
Kreft, T., & Reed, W. F. (2009b). Predictive control of average composition and molec-
ular weight distributions in semi-batch free radical copolymerization reactions.
Macromolecules, 42, 5558–5565.
Leu, G., & Baratti, R. (2000). An extended Kalman filtering approach with a criterion
to set its tuning parameters: application to a catalytic reactor. Computers & Chemical
Engineering , 23, 1839–1849.
Li, R., Corripio, A. B., Henson, M. A., & Kurtz, M. J. (2004). On-line state and parameter
estimation of EPDM polymerization reactors using a hierarchical extended Kalman
filter. Journal of Process Control, 14(8), 837–852.
McAfee, T., Leonardi, N., Montgomery, R., Siqueira, J., Zekoski, T., Drenski, M. F.,
et al. (2016). Automatic control of polymer molecular weight during synthesis.
Macromolecules, 49(19), 7170–7183.
McKeen, L. W. (2014). The effect of temperature and other factors on plastics and elastomers.
William Andrew.
Nicholson, B., López-Negrete, R., & Biegler, L. T. (2014). On-line state estimation of
nonlinear dynamic systems with gross errors. Computers & Chemical Engineering , 70,
149–159.
Norquay, S. J., Palazoglu, A., & Romagnoli, J. A. (1998). Model predictive control based
on wiener models. Chemical Engineering Science, 53(1), 75–84.
Qin, S. J. (2014). Process data analytics in the era of big data. AIChE Journal, 60(9),
3092–3100.
Reed, W. F., & Alb, A. M. (2013). Monitoring polymerization reactions: From fundamentals
to applications. Wiley.
Regis, R. G., & Shoemaker, C. A. (2007). A stochastic radial basis function method for
the global optimization of expensive functions. Informs Journal on Computing , 19(4),
497–509.
Regis, R. G., & Shoemaker, C. A. (2009). Parallel stochastic global optimization using
radial basis functions. Informs Journal on Computing , 21(3), 411–426.
Romagnoli, J. A., & Palazoglu, A. (2005). Introduction to process control. CRC Press.
Romagnoli, J. A., & Sanchez, M. C. (1999). Data Processing and reconciliation for chemical
process operations, Vol. 2. Academic Press.
Salas, S. D., Geraili, A., & Romagnoli, J. A. (2017). Optimization of renewable energy
businesses under operational level uncertainties through extensive sensitivity analysis
and stochastic global optimization. Industrial & Engineering Chemistry Research, 56(12),
3360–3372.
Schlegel, M., Stockmann, K., Binder, T., & Marquardt, W. (2005). Dynamic optimization
using adaptive control vector parameterization. Computers & Chemical Engineering , 29,
1731–1751.
Schuler, H., & Schmidt, C. U. (1992). Calorimetric-state estimators for chemical reactor
diagnosis and control: review of methods and applications. Chemical Engineering
Science, 47 (4), 899–913.
Simon, D. (2006). Optimal state estimation: Kalman, H infinity, and nonlinear approaches.
Wiley.
Srinivasan, N., Kasthurikrishnan, N., Cooks, R. G., Krishnan, M. S., & Tsao, G. T. (1995).
Online monitoring with feedback-control of bioreactors using a high ethanol tolerance
yeast by membrane introduction mass-spectrometry. Analytica Chimica Acta, 316,
269–276.
Tatiraju, S., & Soroush, M. (1997). Nonlinear state estimation in a polymerization reactor.
Industrial and Engineering Chemistry Research, 36(7), 2679–2690.
Weiss, G. H., Romagnoli, J. A., & Islam, K. A. (1996). Data reconciliation - an industrial
case study. Computers & Chemical Engineering , 20(12), 1441–1449.