Control of batch and continuous crystallization processes using reinforcement learning

In crystallization processes, the control of particle size distribution, shape and purity are crucial to achieve the targeted critical quality attributes of the final drug product and meet the pharmaceutical regulatory requirements. This work presents novel optimal trajectory tracking control strategies for batch and continuous cooling crystallization processes using reinforcement learning (RL). The cooling crystallization of paracetamol in water was used as a case study. A model-based reinforcement learning technique is implemented to achieve large crystal size by reducing the deviation from targeted reference trajectories namely process temperature, supersaturation and particle size. This multioutput tracking control strategy was development to address quality and performance challenges commonly encountered in batch and continuous crystallization processes. Various training strategies and reward functions were investigated to enhance the learning capabilities and robustness of the reinforcement-learning-based control. Despite the computational costs inherent to reinforcement learning, the later demonstrated robust control capabilities compared the benchmark control strategies such as model predictive control.