Machine learning superalloy microchemistry and creep strength from physical descriptors

Abstract

We propose an element-agnostic set of descriptors to model superalloy properties with Gaussian process regression. Furthermore, we develop a correction method to deliver the best and most physical predictions for microchemistry in multi-phase alloys. The models' performance in predictions is confirmed for superalloy microchemistry, microstructure, and strength properties. When including new, unseen elements in the test data, the models still give good predictions; such extrapolations into new chemical-space would be impossible with component-based descriptors.