Data-driven model predictive control for continuous pharmaceutical manufacturing.

Area of Science:

• Pharmaceutical Manufacturing
• Chemical Engineering
• Machine Learning

Background:

• Pharmaceutical continuous manufacturing requires operational efficiency for profitability and sustainability.
• Adoption of advanced modeling techniques is crucial for Good Manufacturing Practice (GMP)-regulated environments.
• Existing data-driven methods may lack interpretability or struggle with complex nonlinear dynamics.

Purpose of the Study:

• To demonstrate the feasibility of interpretable, data-driven models for pharmaceutical continuous manufacturing using Dynamic Mode Decomposition with Control (DMDc).
• To present a real-time monitoring strategy framework for designing and tuning a Model Predictive Control (MPC) system.
• To achieve precise granule size control in a twin-screw granulation process.

Main Methods:

• Application of Dynamic Mode Decomposition with Control (DMDc), a machine learning technique, for system identification.
• Development of an interpretable DMDc dynamic model capturing nonlinear dynamics of a multiple input multiple output (MIMO) system.
• Integration of the DMDc model with a Model Predictive Control (MPC) system for advanced process control.

Main Results:

• The DMDc model achieved high performance (R² > 0.93 for D50 predictions) in reconstructing unseen test data.
• The model demonstrated low computational complexity and did not require first-principles knowledge.
• The DMDc-MPC framework was successfully implemented and tested for setpoint tracking and disturbance rejection.

Conclusions:

• Interpretable, data-driven models using DMDc are feasible for pharmaceutical continuous manufacturing.
• The developed DMDc-MPC framework offers a robust solution for real-time monitoring and advanced process control.
• This approach enhances operational efficiency and ensures granule size consistency in pharmaceutical production.