Designing Laves-phase RFe₂-type alloy with excellent magnetostrictive performance by physics-informed interpretable machine learning

Laves-phase RFe₂-type (R = rare earth) magnetostrictive materials have tremendous application potential in smart devices. However, efficiently unearthing novel RFe₂-type compounds with huge magnetostriction in experiments remains challenge due to the vast compositional space. Herein, we employ a physics-informed interpretable machine learning-based strategy to facilitate the design of targeted alloys. A home-built dataset is obtained through constructing composition-physical parameters-magnetostriction relationship. By comparing different models, the XGBoost (XGB) regression model is selected to predict magnetostriction of quaternary Tb_xDy_1-xFe_yV_2-y alloys. The results demonstrate that the optimal performance occurs in the composition range of 0.23–0.38 for Tb content and 0.01–0.08 for V content. The predicted properties are then verified by the measured results of a series of synthesized samples. Additionally, a model interpretability based on SHapley Additive exPlanations (SHAP) values manifests that volume magnetic susceptibility and bulk modulus exert the greatest impact on magnetostriction. This work offers a recipe to swiftly designing RFe₂-type materials with giant magnetostriction.