Machine learning assisted optimization of soft magnetic properties in ternary Fe–Si–Al alloys

The transition from the traditional "post-analysis" strategy for developing soft magnetic materials to an innovative "pre-design" one is highly desirable for the development of advanced electrical devices. In this work, we present the creation of a machine learning (ML) model capable of accurately predicting the soft magnetic properties (JS, HC, μ, ρ) of Fe–Si–Al alloys based on their composition. Through extensive ML experiments employing various algorithms commonly utilized in ML-assisted materials science, including SVM, RFR, KNR, XGB, and others, we achieved high accuracy in predictions, as indicated by R2 values close to 1. The best models were used to predict the properties of the 22800 FeSiAl alloys with the Al and Si content up to 15 wt % and step 0.1 %. Out of this vast compositional space, five alloys were selected for experimental validation, demonstrating the high quality of the predictions. The performance of ML models for specific properties is analyzed in terms of the nature of the distribution of the data used for training. In addition, some composition-properties correlations in the Fe–Si–Al system were analyzed and discussed.