Accelerated design of [formula omitted] alloys with targeted magnetostructural properties through interpretable machine learning

Abstract

Data-driven machine learning (ML) models are developed to rapidly predict the magnetostructural transition temperature (Tt) and thermal hysteresis in the vast search space of MnNiSi-MTX solid solutions. Each alloy is represented in three unique ways based on: (1) chemical compositions, (2) descriptors describing average elemental properties, and (3) crystal structure and magnetic descriptors derived from density functional theory. The trained ML models are experimentally validated through two newly synthesized alloy compositions. While the Tt predictions show good agreement with experimental measurements, the thermal hysteresis predictions were inconclusive. The global and local behaviors of the trained models are then examined using novel post hoc model interpretability techniques. These techniques show that the ML models have learned the salient features that are known to govern the Tt of MTX compounds and give unique insights into how the model behaves for different alloying elements. The outcome of this work has major implications in the design of novel MTX-solid solutions with targeted Tt, in addition to demonstrating new techniques that make the ML results more interpretable. An interactive web application is built, where the researchers can query our trained ML models and design previously unexplored alloys with the desired Tt and thermal hysteresis properties.