Multi-objective optimization of temperature profile using reinforcement learning in batch crystallization process

Abstract The temperature profile is a critical parameter in industrial batch cooling crystallization processes which directly impacts the crystal size distribution (CSD) and subsequent down streaming operations. Conventional cooling approaches, including natural and linear cooling, often fail to achieve optimal results. Linear cooling may trigger excessive primary nucleation by crossing metastable limits, while natural cooling leads to inconsistent nucleation and growth, resulting in smaller and non-uniform crystals. Hence, optimized cooling profiles are used to achieve the desired target CSD. This study employs reinforcement learning (RL) to optimize the temperature profile for batch cooling crystallization for two systems: paracetamol in water and potassium dihydrogen phosphate (KDP) in aqueous solution. RL dynamically adjusts the cooling trajectory by iteratively interacting with a process model, aiming to maximize mean crystal size while minimizing the coefficient of variation. A comprehensive mathematical model incorporating population balance equations, mass balance, and energy balance is developed to simulate the crystallization process. For the paracetamol system, the RL-optimization strategy resulted in a 107.5 % increase in mean crystal size compared to the natural cooling profile. In the case of the KDP system, a 12 % increase in mean crystal length is achieved relative to a linear cooling profile, along with a significant reduction in the coefficient of variation, indicating improved crystal size uniformity. The optimization results obtained using RL are also compared with that from a genetic algorithm for both cases, and RL demonstrated superior performance. This work underscores the potential of RL in advancing broader applications in chemical process optimization.