Exploration of ductility for refractory high entropy alloys via interpretive machine learning

Abstract

Refractory high-entropy alloys (RHEAs) having excellent thermo-mechanical properties always suffer from room temperature brittleness, due to the inherent brittleness of refractory metal elements. RHEAs with compressive ductility ≥30 % hold particular commercial value. While ductility can be engineered through alloying composition and processing condition optimization, this remains challenging due to the virtually infinite and vast unexplored compositional space. Data-driven machine learning (ML) can significantly reduce experimental costs and time by establishing robust correlations between RHEA compositions and ductility. This study constructs an ML model for accurate ductility prediction from sparse compositional data, accelerating the design of ductile RHEAs within infinite compositional space. Two ML algorithms, decision tree (DT) and CatBoost, are trained using physical parameters, with CatBoost demonstrating superior performance in RHEA ductility classification. Integration of the shapley additive explanations (SHAP) model with sample data reveals key relationships between physical parameters and RHEA ductility, providing theoretical insights for developing high-ductility RHEAs. The analysis identifies δχ, T_m, VEC, and ΔH_mix as critical factors in RHEA ductility prediction, with a quasi-linear relationship observed between ΔH_mix and ductility. Through the prediction and subsequent preparation of TiZrNbTa_x (x=0.3, 0.5, 0.7, 1) alloys, the experimental results confirmed the reliability of the proposed model and key parameters. This approach establishes a novel methodology for exploring and developing RHEAs with excellent ductility.