Machine learning-assisted discovery of strong and conductive Cu alloys: Data mining from discarded experiments and physical features

Abstract

Copper alloys with high strength and electrical conductivity are ideal candidates for a wide range of civilian and engineering applications. Traditional methods of alloy design, such as “trial and error” experiments, are costly and time-consuming, limiting the discovery of Cu alloys. Machine learning (ML)-based technology facilitates a better understanding on inter-relationships within massive experimental datasets, and shortens the development cycle of new alloy systems. Herein, we propose a method of material design based on ML to discover high-performance Cu alloys. The ML models are trained from “discarded” experimental data that show undesirable hardness and/or electrical conductivity. We constructed effective Gaussian process regression-based models successfully from limited training data by engaging additional features. The predicted Cu alloys were experimentally synthesized and characterized, exhibiting superior hardness or electrical conductivity with respect to original training data. The increase of hardness is due to precipitation of second-phase Co-Ti and Fe-Ti compounds during aging, while the rise of electrical conductivity is caused by a purification effect of the precipitates on Cu matrix. The impact of physical features on ML models was further evaluated by a genetic algorithm. Our findings suggest that ML holds great potential for developing high-performance Cu alloys.