Interpretable machine learning for functional properties prediction of Fe–Si–Al alloys via feature engineering

This study investigates the impact of composition-based featurization on ensemble machine learning models for predicting magnetic (coercivity H_C, saturation polarization J_S) and transport (electrical resistivity ρ) properties of ternary Fe–Si–Al alloys. Matminer.WenAlloys and CBFV.Oliynyk featurization schemes are compared against chemical composition using Random Forest, Extra Trees, Gradient Boosting, and XGBoost regressors. It is shown that featurization enhances prediction accuracy for H_C and J_S, while composition-based models outperform featurization-based models for ρ. In addition, an evaluation of the featurization scheme for the interpretation of models was carried out. SHAP analysis reveals distinct feature importance patterns: H_C predictions are associated with electronegativity-related descriptors, J_S with Fe compositional dilution, and ρ with descriptors consistent with impurity and electron scattering mechanisms. Comparison of bagging and boosting algorithms reveals that both converge on the same primary descriptors, but differ in their supporting features, with bagging relying on compositional proxies and boosting on auxiliary descriptors for local prediction refinement. The results demonstrate that physically meaningful featurization enables interpretable predictions on small experimental datasets, advancing data-driven design of soft magnetic materials.