Inverse Process-Structure-Property Mapping

Abstract

Workpieces for dedicated purposes must be composed of materials which have certain properties. The latter are determined by the compositional structure of the material. In this paper, we present the scientific approach of our current DFG funded project Tailored material properties through microstructural optimization: Machine Learning Methods for the Modeling and Inversion of Structure-property relationships and their application to sheet metals. The project proposes a methodology to automatically find an optimal sequence of processing steps which produce a material structure that bears the desired properties. The overall task is split in two steps: First find a mapping which delivers a set of structures with given properties and second, find an optimal process path to reach one of these structures with least effort. The first step is achieved by machine learning the generalized mapping of structures to properties in a supervised fashion, and then inverting this relation with methods delivering a set of goal structure solutions. The second step is performed via reinforcement learning of optimal paths by finding the processing sequence which leads to the best reachable goal structure. The paper considers steel processing as an example, where the microstructure is represented by Orientation Density Functions and elastic and plastic material target properties are considered. The paper shows the inversion of the learned structure-property mapping by means of Genetic Algorithms. The search for structures is thereby regularized by a loss term representing the deviation from process-feasible structures. It is shown how reinforcement learning is used to find deformation action sequences in order to reach the given goal structures, which finally lead to the required properties.