Machine Learning-Assisted Discovery of Empirical Rule for Martensite Transition Temperature of Shape Memory Alloys.

Shape memory alloys (SMAs) derive their unique functional properties from martensitic transformations, with the martensitic transformation temperature () serving as a key design parameter. However, existing empirical rules, such as the valence electron concentration (VEC) and lattice volume (V) criteria, are typically restricted to specific alloy families and lack general applicability. In this work, we used a data-driven methodology to find a generalizable empirical formula for in SMAs by combining high-throughput first-principles calculations, feature engineering, and symbol regression techniques. Key factors influencing were first identified and a predictive machine learning model was subsequently trained based on these features. Furthermore, an empirical formula of = 82(ρ¯·MP¯)-700 was derived, where ρ¯ and MP¯ represent the weight-average value of density and melting point, respectively. The empirical formula exhibits strong generalizability across a wide range of SMAs, such as NiMn-based, NiTi-based, TiPt-based, and AuCd-based SMAs, etc., offering practical guidance for the compositional design and optimization of shape memory alloys.