Active learning accelerates the discovery of high strength and high ductility lead-free solder alloys

Developing new materials through traditional trial-and-error is time-consuming and costly due to the vast potential compositions to explore. Moreover, some desirable properties are mutually exclusive, presenting challenges for materials design. Herein, we employed an active learning strategy to address the strength-ductility trade-off in SAC105 lead-free solders. We developed two Gaussian regression models, one for strength and one for elongation representing ductility. The Gaussian Upper Confidence Boundary algorithm, a Bayesian optimization method, was used to balance exploitation and exploration by considering prediction uncertainty. The variance weight was designed to decay with adaptive iterations. The two models were linearly combined and the maximal combined model recommended alloys on the Pareto front for experiments. After just three iterations, a new 91.4Sn-1.0Ag-0.5Cu-1.5Bi-4.4In-0.2Ti low-silver SAC solder was discovered, exhibiting 73.94±5.05 MPa strength and 24.37±5.92% elongation, representing the best comprehensive mechanical properties. Weldability experiments also showed excellent performance. Based on Ab initio molecular dynamics, a precise machine learning potential revealed the alloying mechanism in Sn-In solid solutions with quantum accuracy in systems of over 103 atoms, closely matching experimental X-ray diffraction pattern. To facilitate materials informatics development, all active learning algorithms was made open-source in our designed framework, Bgolearn.