Tailoring glass forming ability and coercivity in iron-based nanocrystalline alloys through machine learning-driven feature interpretation

Abstract

Nanocrystalline soft magnetic materials developed from precursors with high glass forming ability (GFA), combined with their low coercivity (H_c), are beneficial for advanced electronic devices. Herein, machine learning (ML) models with 519 original data and 22 features were developed for predicting GFA and H_c. Four-step feature screening was utilized to obtain important features predicting GFA, and top four features (e.g., VEC, VEC1, c_B, and ΔT) remained. Among three distinct ML algorithms, the XGBoost outperforms SVM and RF, achieving an AUC of 0.93. The accuracy of the classification model is up to 96.7 %. The SHAP analysis revealed that when the valance electron concentration (VEC) was below 7.29, the VEC excluding Fe, Ni, and Co elements (VEC1) was above 0.71, boron content (c_B) among 6.85 ∼ 11.54 at.%, and the crystallization temperature intervals (ΔT) above 121 °C, the precursors easily obtain fully amorphous structures. Three optimized alloys were designed by XGBoost and NSGA-II algorithm, exhibiting good GFA and low H_c of 1.34 ∼ 1.40 A/m. The errors between ML-predicted and experimental H_c values were lower values compared to those previously reported alloys, underscoring the reliability and robustness of the developed ML algorithm.