Data-driven model predictive control for continuous pharmaceutical manufacturing

Abstract

This study demonstrates that the development of interpretable, data-driven models for pharmaceutical continuous manufacturing is feasible using a machine learning method called Dynamic Mode Decomposition with Control (DMDc). This approach facilitates adoption within Good Manufacturing Practice (GMP)-regulated areas in the pharmaceutical industry. Furthermore, since the pharmaceutical industry needs to be more operationally efficient to be profitable and sustainable, we present a real-time monitoring strategy framework using an interpretable DMDc dynamic model for the design and tuning of a model predictive control (MPC) system for granule size control in a twin-screw granulation process. This model exhibits low computational complexity without requiring first principles knowledge, while effectively capturing nonlinear dynamics of this multiple input multiple output (MIMO) system, with enhanced performance (e.g., R² > 0.93 for D50 predictions) in the reconstruction of unseen test data in comparison with benchmark data-driven methods for system identification. The DMDc-MPC was implemented and tested on setpoint tracking and disturbance rejection and the proposed advanced process control framework guaranteed both.