Interpretable machine learning workflow for evaluation of the transformation temperatures of TiZrHfNiCoCu high entropy shape memory alloys

Abstract

Machine learning approaches (ML) based on data-driven models are conducive to accelerating the assessments of the martensitic transformation peak temperature (Tp) of TiZrHfNiCoCu high entropy shape memory alloys (HESMAs) over a huge composition space. In this work, an interpretable machine learning workflow was established through dataset construction, feature selection, modeling and validation, and model interpretation. We identified a set of key feature combinations closely related to Tp, by exploiting Pearson correlation selection, univariate feature selection, and forward feature elimination. The established ML model was then used to estimate the Tp of three newly synthesized alloys, with their prediction relative errors of less than 3 % in comparison with the experimental measurements. The behaviors of our ML model were interpreted by the Shapley Additive exPlainations (SHAP) approach, demonstrating the crucial role of CV22 (Allred Rochow electronegativity) in the prediction of Tp. In addition, the ML model in combination with our designed interpretation strategy was further used to investigate the effects of alloying elements on the Tp, which showed that the TiZrHfNiCoCu HESMAs with 9 ≦ Co (mol%) ≤ 10 and 15 ≦ Cu (mol%) ≤ 17 have pronounced positive effects on Tp.