Reinforcement learning strategy for control of microstructure evolution in phase field models

Abstract

We propose a reinforcement learning-based approach for controlling microstructure evolution in phase field models. A key feature of our approach is the replacement of the high-fidelity but computationally expensive phase field simulation with an inexpensive surrogate model of acceptable fidelity. We apply our approach to a 2-D model problem based on the Allen-Cahn equation that represents a two-phase microstructure with a single, scalar order parameter. The control parameters for this model are spatially and temporally varying temperature (T) and bias fields (h). Our approach generates sets of T and h parameters that yield final microstructures qualitatively matching target microstructures. We discuss further developments required to apply this approach to optimal processing of designer microstructures.