Machine learning of the mechanical properties and data-driven design of lead-free solder alloys

The development of lead-free solder alloys with high-strength, high-toughness, good creep, thermal-fatigue resistance, and good soldering properties are essential for the industrial development of modern electronics and microelectronics. Improving the strength of solder alloys usually induces a reduction in ductility due to the tradeoff between strength and ductility, which ultimately results in deterioration in toughness. However, it is challenging to optimize these two mutually restrictive properties of solder alloys simultaneously while comprehensively improving the mechanical properties. In this study, two kinds of machine learning models were established for Sn-based solder alloys. The alloy composition and atomic scale parameters were used as input features, respectively, and tensile strength and fracture elongation were the target properties. Two machine learning models with atomic features exhibited high tensile strength and fracture elongation prediction accuracies, respectively. A virtual sample space was constructed for Sn-based alloys according to the range and variation step size of the alloying elements in the original training dataset, and the average absolute SHAP values were used to rank and select features. With the L norm of strength and ductility as an index of the comprehensive mechanical property, two virtual samples on the Pareto front were selected for experimental verification. The experimental results were highly consistent with the ML predictions, and the selected samples exhibited improved properties in terms of their comprehensive mechanical properties. The results of this study can be used as guidance for the multi-objective optimization design of complex alloys.