Interpretable transfer learning accelerates the design of high-temperature-resistant copper alloys with limited data

Copper alloys are critical for high-temperature applications, including rocket combustion chamber linings and nuclear reactor deflectors, yet traditional smelt-cast alloys struggle under extreme thermal conditions. Recent advances in big data and machine learning have enabled the efficient design of high-performance copper alloys. Here, we propose a novel interpretable transfer learning model–the electrical conductivity-hardness artificial neural network (ECH–ANN)–that integrates feature and model transfer to accurately predict properties of dispersion-strengthened copper alloys via powder metallurgy. We further established a feature screening pipeline comprising correlation screening, recursive elimination, recursive addition, and exhaustive screening. Using SHapley Additive exPlanations (SHAP), we quantified the importance and impact of key features on performance. Guided by these insights, a high-performance Cu-5.0 vol% Cr 2 O 3 composite was designed and fabricated through in-situ internal oxidation, achieving a strength of 570 MPa, electrical conductivity of 73.6 % IACS, and a softening temperature of 1000 °C, approaching copper’s melting point. This work provides a data-efficient framework for designing advanced materials, while elucidating the relationships between component characteristics and material properties.