Modeling and optimization of composition-property relationships in copper alloys using explainable machine learning

Abstract To address the inherent trade-off between strength and electrical conductivity in copper alloys, this study proposes an interpretable modeling framework based on automated machine learning. Utilizing a dataset of 465 experimental records containing both tensile strength and electrical conductivity measurements, with chemical composition as the sole input variable, we systematically compared multiple regression algorithms using the PyCaret platform. Interpretability and visualization were achieved through SHAP and t-SNE analyses. Results demonstrate that nonlinear ensemble models significantly outperform traditional linear models, with the Gradient Boosting Regressor achieving determination coefficients (R 2 ) of 0.94 and 0.97 for strength and conductivity predictions on the test set, respectively, while maintaining low error rates. SHAP analysis revealed that elements such as Ni and Si contribute positively to strength, whereas Cu, Ni, and Sn are critical factors influencing conductivity. t-SNE dimensionality reduction further identified performance distribution characteristics of typical alloy systems: Cu-Cr-Zr alloys excel in medium-strength and high-conductivity regions, Cu-Ni-Si systems exhibit high-strength potential, and Cu-Zn alloys occupy low-performance zones. This work provides an efficient and interpretable approach for modeling composition-property relationships in copper alloys, offering theoretical support and data-driven insights for the rapid design and optimization of next-generation high-strength, high-conductivity copper alloys.