Data-driven modeling on design and frictional response of high-entropy alloys through material composition variation

Materials engineering informatics (MEI) provides a data-driven route for accelerating high-entropy alloy (HEA) design, but reliable prediction of phase constitution and tribological response remains challenging. In this study, an integrated machine-learning framework was developed for HEA phase classification and coefficient of friction (COF) prediction. For phase prediction, a refined literature-derived dataset, processed using physically meaningful descriptors, and modeled using XGBoost and random forest. XGBoost achieved the best classification performance, with an accuracy of 0.8476, balanced accuracy of 0.8461, and Macro-F1 of 0.8527. Feature-importance analysis identified VEC, ΔHmix, and Tm as the dominant phase-related descriptors. For COF prediction, 11 AlCoCrFeNiMo-based compositions were fabricated by arc melting and tested under different loads. Alloy descriptors, applied load, and historical COF values were used as model inputs for XGBoost, long short-term memory (LSTM), and gated recurrent unit (GRU) models. Among the nine model-optimization combinations, Bayesian-optimized XGBoost achieved the lowest validation RMSE of 0.028289 and the strongest test performance of R 2 0.938979 and CVRMSE 4.156%. Experimental validation on an FCC-dominant HEA showed a wear rate of 8.05×10⁻⁴ mm³N⁻¹m⁻¹ at 5 N. Contact-mechanics analysis indicated local elastic-plastic deformation under all loads, while frictional work increased from 333 to 1242 J. Worn-surface analysis revealed a transition from predominantly abrasive wear at lower load to greater plastic deformation, tribolayer disruption, and delamination at higher load. Overall, the proposed framework integrates phase prediction, COF modeling, and experimental validation, offering an efficient strategy for screening HEAs with desirable phase stability and tribological performance.