From Open Loop to Real-time Recipe Optimization for Complex Industrial Batch Processes

We propose a fundamental-model-based optimization framework for open-loop and real-time batch recipe optimization using mathematical programing (MP) and reinforcement learning (RL). The dynamic fundamental model gives rise to a complex nonlinear differential algebraic equations (DAE) system. We introduce a decomposition-based initialization algorithm for solving the large-scale nonlinear program (NLP) resulting from the discretization of the DAE system. The proposed MP and RL-based approaches are implemented to optimize the recipe of a semi-batch process in the Dow Chemical Company. For open-loop optimization, we find the optimal profiles of two input variables and the batch length that maximize average profit. For real time optimization, we train the RL agent using the fundamental model with uncertainties. The trained agent can interact with the actual process and provide control actions in real time.