Coercivity-Aware Machine Learning Discovery of Rare-Earth-Free Soft Magnetic Alloys with First-Principles Magnetic Benchmarks

Abstract

Coercivity-aware machine learning is developed to screen rare-earth-free soft magnetic alloys using Curie temperature and coercivity as coupled design targets. A compiled experimental magnetic-material dataset is evaluated using composition-grouped partitioning to avoid overlap of identical compositions between training and testing, and elemental fractions and composition-weighted elemental descriptors are examined for prediction of the two magnetic properties. Residual-based prediction intervals and distance-to-training filters are then incorporated to screen Fe-Co-Ni-Mn-Al-Si compositions. Composition-grouped evaluation shows lower predictive accuracy than conventional random partitioning, demonstrating the importance of evaluating transfer to previously unseen compositions. Combining weighted elemental descriptors with elemental fractions improves Curie-temperature prediction, whereas the additional descriptors do not improve coercivity prediction, consistent with the strong dependence of coercivity on processing and microstructure. Uncertainty-aware screening substantially narrows the candidate space and identifies Fe-Co-rich compositions for further evaluation. First-principles calculations independently show strong magnetic polarization and magnetovolume effects in representative Fe-Co structures, while finite-temperature atomistic spin simulations place the Fe-Co benchmark systems on high magnetic-ordering-temperature scales. The combined approach provides a computational framework for prioritizing rare-earth-free soft-magnetic chemistries while distinguishing composition-level screening from candidate-specific physical validation.