WEB Adaptive microstructure-processing-path optimization by reinforcement learningWednesday (01.01.2020) 00:30 - 00:45 Part of:
Specific material properties are typically related with a variety of microstructures. Thus, the result of adequate microstructure optimization is a set of microstructures for a given goal property. A process can reach these microstructures by applying a yet unknown sequence of deformation steps, which has to be found.
We present a novel reinforcement-learning-based approach to find the most efficient (e.g. shortest) process path from a start microstructure to one of the target microstructures. Reinforcement learning builds a family of adaptive decision optimization algorithms, which can be used to learn optimal process control while interacting with the process. It learns by approximating functions of expected future reward for an optimal process control regarding a target microstructure. The resulting control selects each processing step attempting to maximize the approximated future expected reward.
The algorithm we present is model free and generalizes over target microstructures. Due to the generalization capability of these approximations, the shortest path to any microstructure of the target set can be found. Moreover, the proposed approach can transfer learned process control knowledge across various target microstructure sets, and thus to unseen goal properties and even property types.