FCC vs. BCC phase selection in high-entropy alloys via simplified and interpretable reduction of machine learning models

Abstract

Machine learning (ML) models for phase selection in high-entropy alloys often suffer from an inherent lack of interpretability despite remarkable advancements of late. Here we present a mathematical expression for the probability of occurrence of FCC and BCC phases in HEAs that was obtained through statistical analysis of the experimental data and ML models. The model presented here utilizes a logistic function to isolate the effect of valence electron count on FCC and BCC phase occurrence and models the residuals as a function of six other physical and thermodynamic descriptors. Thus, the complex ML model is replaced by a simplified and interpretable mathematical function. The proposed model is quite consistent with the ML model and experimental database and enables a direct quantitative estimation of feature contributions towards phase occurrence probabilities leading to insights into the decision-making process learnt by the ML model for phase selection in high-entropy alloys.