Prediction of magnetocrystalline anisotropy constants in FeCoNi thin films using density of states features

We developed an integrated dataset by combining experimentally measured magnetocrystalline anisotropy constants (K1) of FeCoNi alloy thin films with density of states (DOS) calculated from first-principles calculations without spin–orbit coupling (SOC). The DOS was transformed into statistical descriptors and directly used as features for machine learning (ML) models to predict K1. The highest predictive performance was achieved when using DOS features within approximately ± 0.1 eV of the Fermi level, yielding a sign prediction accuracy of 83%. This result demonstrates that low K1 materials can be efficiently identified from the DOS even without the SOC. Furthermore, feature importance analysis revealed that the minority-spin DOS near the Fermi level plays the dominant role in the model predictions, reflecting the fact that most alloys in the dataset are Ni–rich with fcc structures, where contributions from majority-spin DOS are negligible.