Polymer Grade Transition Control via Reinforcement Learning Trained with a Physically Consistent Memory Sequence-to-Sequence Digital Twin

In this work, a memory layer sequence-to-sequence digital twin (ML-StSDT) of a high-density polyethylene (HDPE) reactor simulated by ASPEN DynamicsTM was constructed using simulated grade transition and steady-state operating data. A reinforcement learning control (RLC) algorithm was developed by training with the ML-StSDT. The RLC was able to control both grade transition and steady-state operation of the simulated plant. The RLC performs better or equally well when compared with the direct application of ML-StSDT in nonlinear model predictive control (NLMPC) but substantially reduces the computation load. Our results demonstrate the feasibility of deep learning models serving as a digital twin for RLC training in nonlinear process control applications.