(184x) Recurrent Neural Network-Based Model Predictive Control for Continuous Pharmaceutical Manufacturing | AIChE

(184x) Recurrent Neural Network-Based Model Predictive Control for Continuous Pharmaceutical Manufacturing

Authors 

Wang, X. - Presenter, National University of Singapore
Wong, W. C., Georgia Institute of Technology
In the recent past, the pharmaceutical industry has witnessed exponential growth in momentum in transforming operations away from the conventional batch processes [1, 2] towards continuous manufacturing [3, 4]. This fundamental shift is necessary to allow this industry to effectively address key challenges such as increasing margin, achieving superior Critical Quality Attributes (CQAs), extending manufacturing range (e.g., for Active Pharmaceutical Ingredients that are not amenable to conventional batch processes) while reducing cost, waste, footprint, lead-times, in the face of ever-increasing regulatory requirements globally.

With the aforementioned developments in continuous manufacturing technologies and shift towards Quality by Design paradigms, there is increasing demand for more advanced model identification and process control strategies in the pharmaceutical industry. Model Predictive Control (MPC) has represented a step-change in the way process industries approach and deploy advanced process control solutions at scales ranging from local unit operations to plant-wide control. In similar fashion, for continuous pharmaceutical manufacturing, Model Predictive Control (MPC) is considered to be a key technology that enables this vision by being able to rigorously address technical requirements such as the control of highly non-linear reactions (e.g., reversible reactions, side reactions, etc.) in the presence of tight constraints.

Within industry, empirical, data-driven models are oftentimes used in MPC, as first-principles models, whilst desirable in some situations, are costly to obtain. This may be especially so in pharmaceutical manufacturing and may
not be necessary nor practical. In the context of data-driven approaches, while it is generally recognized that Neural Networks (NNs) have broad-ranging capabilities of approximating non-linear functions, their usage as part of MPC has not found wide applicability.

Our contribution is in demonstrating the applicability of a special class of NNs, termed Recurrent Neural Networks (RNNs), for MPC applications for continuous pharmaceutical manufacturing applications. In particular, RNNs possess
inherent state-space structures that naturally lend themselves well for time-series modeling for predictive control. While MPC has been proposed and investigated in the control of CQAs in continuous pharmaceutical manufacturing [5], it has rarely been applied towards the control of the reactor, which is the heart of the process. RNNs in the context of closed-loop MPC has also been studied in but not for Continuous Pharmaceutical Manufacturing.

We study a single, Multi-Input Multi-Output CSTR example (per [6], [7]) which experiences input multiplicity due to reversible kinetics (A ↔ R ↔ S). The optimal set-point has been carefully selected to optimize production against
excessive downstream separations cost. This set-point is close to a point of inflexion where the system gain changes in sign with respect to reactor temperature. We show how a deep RNN can be learned and present associated closed-loop performance results for two scenarios that require the RNN-based Non-Linear Model Predictive Controller to move from either side of the inflexion point towards the desired set-point. As a result of input multiplicity, it is noted that a single linear controller is not expected to work well for the problem of concern [8]. We show favorable results of the RNN-MPC controller when compared against a Non-Linear MPC benchmark that uses the true plant model for control in terms of closed-loop performance and computational time.

Reference:

[1] Glasnov T. Continuous- flow chemistry in the research laboratory. Springer; 1st ed. 2016 edition (June 2, 2016); 2016.

[2] Poechlauer P, Colberg J, Fisher E, Jansen M, Johnson MD, Koenig SG, et al. Pharmaceutical Roundtable Study Demonstrates the Value of Continuous Manufacturing in the Design of Greener Processes. Organic Process Research and Development 2013; 17(12):1472–1478.

[3] Lakerveld R, Benyahia B, Heider PL, Zhang H, Wolfe A, Testa CJ, et al. The Application of an Automated Control Strategy for an Integrated Continuous Pharmaceutical Pilot Plant. Organic Process Research and Development 2015; 19(9):1088– 1100.

[4] Schaber SD, Gerogiorgis DI, Ramachandran R, Evans JMB, Barton PI, Trout BL. Economic Analysis of Integrated Continuous and Batch Pharmaceutical Manufacturing: A Case Study. Industrial & Engineering Chemistry Research 2011 9; 50(17):10083–10092.

[5] Mesbah A, Paulson JA, Lakerveld R, Braatz RD. Model Predictive Control of an Integrated Continuous Pharmaceutical Manufacturing Pilot Plant. Organic Process Research & Development 2017 6; 21(6):844–854. http://pubs.acs.org/ doi/10.1021/acs.oprd.7b00058.

[6] Seki H, Ooyama S, Ogawa M. Nonlinear Model Predictive Control Using Sucessive Linearization - Application to Chemical Reactors. Trans of the Society of Instrument and Control Engineers 2004; E-3(1):66–72.

[7] Koppel LB. Input multiplicities in nonlinear, multivariable control systems. AIChE Journal 1982; 28(6):935–945.