Recurrent Neural Network-based Model Predictive Control for Continuous Pharmaceutical Manufacturing (1807.09556v1)

Abstract: The pharmaceutical industry has witnessed exponential growth in transforming operations towards continuous manufacturing to effectively achieve increased profitability, reduced waste, and extended product range. Model Predictive Control (MPC) can be applied for enabling this vision, in providing superior regulation of critical quality attributes. For MPC, obtaining a workable model is of fundamental importance, especially in the presence of complex reaction kinetics and process dynamics. Whilst physics-based models are desirable, it is not always practical to obtain one effective and fit-for-purpose model. Instead, within industry, data-driven system-identification approaches have been found to be useful and widely deployed in MPC solutions. In this work, we demonstrated the applicability of Recurrent Neural Networks (RNNs) for MPC applications in continuous pharmaceutical manufacturing. We have shown that RNNs are especially well-suited for modeling dynamical systems due to their mathematical structure and satisfactory closed-loop control performance can be yielded for MPC in continuous pharmaceutical manufacturing.

• The paper demonstrates that RNNs can effectively model complex, nonlinear process dynamics in pharmaceutical manufacturing.
• The study applies a two-layer RNN architecture that achieved an RMSE of 0.0083, ensuring robust performance in MPC scenarios.
• The research highlights the practical feasibility of integrating data-driven RNNs with MPC to enhance control precision and transient response.

Recurrent Neural Network-based Model Predictive Control for Continuous Pharmaceutical Manufacturing: An Overview

The paper "Recurrent Neural Network-based Model Predictive Control for Continuous Pharmaceutical Manufacturing" explores the application of Recurrent Neural Networks (RNNs) in Model Predictive Control (MPC) within the domain of continuous pharmaceutical manufacturing. This work underscores the potential utility of RNNs for generating robust data-driven models that can significantly enhance the efficacy of MPC systems in managing the intricate dynamics of pharmaceutical processes.

Introduction

Continuous manufacturing in the pharmaceutical industry aims to address key challenges such as reducing waste, lowering costs, minimizing footprints, and shortening lead times. Advanced process control strategies, particularly those involving Model Predictive Control (MPC), are seen as critical enablers for achieving these objectives. Traditional physics-based models, although insightful, often fall short in handling the complex and nonlinear nature of pharmaceutical reactions. This gap underscores the importance of data-driven models, with a specific interest in Artificial Neural Networks (ANNs) and, more recently, RNNs given their proficiency in capturing temporal dynamics.

Methodology

The research focuses on the application of RNNs for system identification within MPC frameworks. The authors implemented their approach using a single Continuous-Stirred Tank Reactor (CSTR) for illustrative purposes. The CSTR involves complex reaction kinetics characterized by reversible reactions ($A \rightleftarrows R \rightleftarrows S$), with reaction rates sensitive to both temperature and flow rate manipulations. The challenge is to model these dynamics effectively to optimize the production yield of the desired product, $R$.

System Identification

The paper describes a process for collecting operational data through perturbations in manipulated variables to create a rich training dataset. RNNs are then trained using this data, leveraging the back-propagation through time (BPTT) algorithm to optimize the network weights. Importantly, the paper explores the use of Long Short-Term Memory (LSTM) cells within the RNN architecture to address issues with training stability due to gradient vanishing and exploding problems.

Model Predictive Control Formulation

For MPC, the RNN model predicts future system outputs over a prediction horizon. The MPC optimization problem balances deviations from set points with the control effort, ensuring constraints on the manipulated variables and their rates of change are respected. The authors used Sequential Quadratic Programming (SQP) for solving the non-linear optimization problems intrinsic to MPC.

Experimental Setup and Results

Two primary control scenarios were studied: (1) system start-up from low-product conditions, and (2) recovery from an upset state. The RNN-MPC system's performance was benchmarked against a traditional NMPC that used the true plant model.

• System Identification Results: Various configurations of RNNs were tested, with the best performance achieved using a two-layer RNN with 1,000 nodes per layer, as indicated by an RMSE of 0.0083 on unseen test data.
• Closed-Loop Control Performance: The RNN-MPC demonstrated stable and effective control in both scenarios, with performance closely mimicking that of the benchmark NMPC. Notably, the RNN-MPC achieved superior transient response and steady-state accuracy when appropriately tuned, contrasting with traditional linear MPC approaches which failed under the same conditions.

Implications and Future Work

This research highlights several key implications:

• Practical Feasibility: The paper shows that RNNs can effectively model the complex dynamics of pharmaceutical processes, thereby enabling robust and efficient MPC strategies. This is particularly relevant for processes where traditional models are cumbersome or infeasible.
• Enhanced Control Performance: Implementing RNN-MPC can result in superior control performance, especially in handling the non-linearities and input multiplicities of reaction kinetics in continuous manufacturing.

Moving forward, expanding this methodology to encompass multiple reactors and more complex reaction schemes could further solidify the utility of RNNs in pharmaceutical manufacturing. Future research may also explore the integration of other advanced machine learning techniques and optimization algorithms to enhance both system identification and control performance further.

Conclusion

The paper provides a comprehensive examination of using RNNs within MPC frameworks to manage the complexities of continuous pharmaceutical manufacturing. The results compellingly argue for the adoption of data-driven modeling techniques to achieve precise and agile control in pharmaceutical production systems. By establishing the practical viability and benefits of RNN-MPC, this research sets a foundation for broader applications and further innovations in the field.