Developing softening-resistant Cu-Cr alloys and understanding their mechanisms via mechanism-informed interpretable machine learning

Cu-Cr alloys are widely applied in electronic, aerospace and nuclear industries, due to their high strength and high conductivity. However, their terrible softening resistance limits wider applications. This paper presents a novel strategy of integrating mechanism features into interpretable machine learning (ML) to develop softening-resistant Cu-Cr alloys and to understand their mechanisms. First, the mechanism features were specially designed to describe mechanisms potentially vital to softening resistance, and they were obtained through first-principles calculations. Those mechanism features that described interfacial segregation and solute diffusion exhibited significant Gini importance during feature selection. Only integrated with them, did ML models achieve great performance, accurate predictions, and successful development of Cu-0.4Cr-0.10La/Ce (wt.%) alloys with excellent softening resistance. Then, the contributions of these mechanism features to the predictions were interpreted by a game theoretic approach, but unexpectedly, they were not fully consistent with interpretations that we expected from mechanism features. Finally, investigation targeted at these inconsistencies gave novel insights into softening resistance mechanisms. The Cu-Cr-La/Ce alloys’ excellent softening resistance was not induced by a prevailing mechanism of La/Ce atoms segregating at phase interfaces, nor by an expected mechanism of La/Ce atoms improving the Cr atom jump energy barriers. Instead, it was caused by a unique mechanism in which La/Ce atoms competed with Cr atoms for vacancies and therefore depleted the available vacancies for the Cr atom jump. This paper demonstrates a new paradigm of developing softening-resistant Cu-Cr alloys and understanding their mechanisms via mechanism-informed interpretable ML.