Explainable artificial intelligence approach for yield strength prediction in as-cast multi-principal element alloys

Abstract

In this paper, we develop an explainable artificial intelligence (XAI) approach to rapidly predict and explain the temperature-dependent yield strength (YS) trends of as-cast multi-principal element alloys (MPEAs) based on an ensemble of support vector regression (eSVR) models. Post-hoc interpretability of trained eSVR models based on SHapley Additive exPlanations, k-means clustering and Ceteris-Paribus profile plots reveal trends that explain the variation in YS as a function of temperature, constituent elements, and phase. According to our XAI algorithm, MPEAs with the highest room temperature YS should occur in multi-phase refractory metal based six-component systems, whose microstructure is made up of secondary phase precipitates within the body-centered cubic (BCC) matrix. The propensity for secondary phase formation is promoted by the optimal concentrations of Al and/or Si alloying additions. Even in MPEAs that form in the face-centered cubic structure, secondary phases are identified to be important for higher YS at room temperature. We uncover two design criteria that favor high room-temperature YS: (i) Choose compositions with equiatomic or near-equiatomic concentrations and (ii) Choose alloying elements with an optimal number of p-electrons in the valence configuration. High temperature YS trends associated with both single-phase and multi-phase BCC-based MPEAs are dominated by the temperature descriptor, which obfuscates the interpretation of other input variables.