A machine learning strategy for rapid design of preparation parameters in zero-sample complex alloy

To address the zero-sample challenge in preparation parameter design for newly developed alloys, a novel machine learning strategy that integrates basic dataset construction with Bayesian optimization, was proposed. The impact of basic sample dataset construction methods, optimization benchmarks and multi-objective utility functions on Bayesian optimization was investigated. It was found that the combination of orthogonal design, linear benchmark, and shifted multiplicative utility function exhibits the best optimization performance. The strategy was then applied to a new Cu−Ni−Co−Si alloy with ultra-low Co content (0.7 wt.% Co), previously designed by our research team. Rapid optimization of six preparation parameters in the two-stage deformation and aging process of the zero-sample alloy was achieved through only 23 experiments. The measured ultimate tensile strength and electrical conductivity of the new alloy were 878 MPa and 44.0%(IACS), respectively, reaching the comprehensive performance level of the Cu−Ni−Co−Si alloy (C70350 alloy) containing 1.0−2.0 wt.% Co.