Multi-objective optimization for high-performance Fe-based metallic glasses via machine learning approach

Abstract

Fe-based metallic glasses (MGs) are a class of promising soft magnetic materials that have received great attention in transformer industries. However, it is challenging to achieve a balance between saturation magnetization (Bs), glass-forming ability and plasticity due to their contradictory correlations in Fe-based MGs, which severely hinders the development of new Fe-based MGs with advanced performances. Inspired by the significant development in machine learning technology, we herein propose a multi-objective optimization strategy to search for Fe-based MGs with optimal combinations of critical casting size (Dmax), Bs, and plasticity. The objective functions are built in combination with neural network models for predicting Dmax and Bs, as well as empirical formula for plasticity. The effect of number of hidden layers is investigated and the dropout regularization method employed to improve the prediction performance. Our results show that the predictions of Bs and Dmax by using alloy composition as the sole input perform well, as evidenced by their r2 values of 0.963 and 0.874, respectively. Multi-objective optimization based on the genetic algorithm is executed to obtain the Pareto front and Pareto-optimal solutions. The Pareto-optimal alloys predicted for the Fe83C1BxSiyP16-x-y and FexCoyNi72-x-yB19.2Si4.8Nb4 systems are in good agreement with those reported in experiments. This work thus showcases potential applications for the design of high-performance Fe-MGs against conflicting objectives.