Composition-based machine learning approach for A _f prediction in NiTiZr alloys

Abstract NiTiZr, which is among high-temperature shape memory alloys, stands out with its cost-effectiveness and superior properties. However, determining the optimal composition through traditional trial-and-error methods is time-consuming and costly. In this study, a comprehensive machine learning (ML) framework was developed to predict the austenite finish temperature ( A _f ), a critical phase transformation parameter of NiTiZr alloys. Using experimental data compiled from the literature, five different algorithms were trained and their performances were compared. SHAP-based explainability analyses were used to identify the most influential factors on transformation temperatures. The developed model was also tested on an independent experimental sample and showed trend-level agreement with the experimentally observed transformation behavior. The findings show that ML models based solely on chemical composition, independent of processing history, can provide a useful framework for preliminary screening of candidate NiTiZr alloys. This study presents a composition-based and interpretable ML framework tailored to A _f prediction in the NiTiZr system.