Reinforcement learning-based approach for optimizing solvent-switch

Abstract

In chemical and pharmaceutical industries, process control optimization is a crucial step to improve economical efficiency and the environmental impact. The current state-of-practice heavily relies on expert knowledge and extensive lab experiments. This not only increases the development time but also limits the discovery of new strategies. In this study, we propose Reinforcement Learning-based optimization approach for solvent-switch processes. We utilize a digital twin as the environment for a process designed to switch the THF to 1-propanol. A reward function is created for minimizing the process time and constraints are implemented using logarithmic barrier functions. A PPO agent is trained on the environment. The agent proposed a novel strategy that combines two conventionally separate phases, evaporation and constant volume distillation. This strategy resulted in an overall cost decrease of 24.9% compared to the baseline strategy. Moreover, results were verified experimentally on a pilot plant of Johnson & Johnson (J & J).