Divide and conquer： Machine learning accelerated design of lead-free solder alloys with high strength and high ductility

Abstract

Abstract The attainment of both high strength and high ductility is always the goal for structure materials, because the two properties generally are mutually competing, called strength-ductility trade-off. Nowadays, the data-driven paradigm combined with expert domain knowledge provides the state-of-the-art methodology to design and discovery for structure materials with high strength and high ductility. To enhance both strength and ductility, a joint feature is proposed here to be the product of strength multiplying ductility. The strategy of “divide and conquer” is developed to solve the contradictory problem, that material experimental data of mechanical behaviors are, in general, small in size and big in noise, while the design space is huge, by a newly developed data preprocessing algorithm, named the Tree-Classifier for Gaussian Process Regression (TCGPR). The TCGPR effectively divides an original dataset in a huge design space into three appropriate sub-domains and then three Machine Learning (ML) models conquer the three sub-domains, achieving significantly improved prediction accuracy and generality. After that the Bayesian sampling is applied to design next experiments by balancing exploitation and exploration. Finally, the experiment results confirm the ML predictions, exhibiting novel lead-free solder alloys with high strength high ductility. Various material characterizations were also conducted to explore the mechanism of high strength and high ductility of the alloys.