Multi-objective design of Ni-B-Al master alloy by adaptive machine learning-driven aluminothermic reduction experiment

Designing Ni-B-Al master alloy greatly benefits from solving the refractory problem of B element and the burning loss problem of Al element in the preparation of nickel-based superalloy. To quickly and accurately tailor the chemical compositions of Al and B elements, developing new multi-objective optimization technologies to replace traditional high-cost, long-cycle trial-and-error method is necessary and challenging. In the present study, we developed an adaptive close-loop strategy that essentially is a process of machine learning-driven thermit reduction experiments to simultaneously optimize two composition targets, namely Al and B contents in the alloys. We also proved that the multi-objective effective global optimization (EGO) algorithm supported by the centroid-based utility function is effective in advancing the Pareto front quickly, and it only drives five thermit reduction experiments to find the expected Ni-B-Al master alloy (Al content of 5.26 wt% and B content of 14.28 wt%). The design principles of Ni-B-Al master alloy were also revealed, which is valuable for profoundly understanding the relationship between experimental conditions and product performance. To our knowledge, machine learning is the first to be applied in master alloy development, and the success of the multi-objective EGO algorithm in the experimental field provides a valuable reference for accelerating the design of new alloys in other materials systems.