Simple Stacking Ensemble Machine Learning Approach for Vickers Hardness Prediction of Arc Vacuum Melted High-Entropy Alloys

High-entropy alloys (HEAs) show excellent mechanical properties, but their enormous compositional space makes experimental and computational exploration expensive. This study predicts the Vickers hardness (HV) of HEAs from nominal elemental compositions (at.%) using a composition-based stacking ensemble machine learning approach without the use of handcrafted descriptors. A dataset of 120 vacuum arc-melted (VAM) HEAs containing 15 constituent elements (Re, Ru, Al, Si, Ti, Cr, Fe, Co, Ni, Zr, Nb, Mo, Hf, Ta, and W) was compiled. Four baseline models, k-nearest neighbors (KNN), support vector regression (SVR), extreme gradient boosting (XGB), and artificial neural networks (ANN), were evaluated. SVR achieves the highest baseline performance (R2 = 0.90). Stacking ensembles with the SVR meta-learner improved the weaker base learner performance by 17.20% and 12.50% for KNN-SVR and ANN-SVR respectively, relative to their base models. The best-performing stacking ensemble, XGB-SVR, only results in a minimal improvement of 2.90% compared to SVR. Ru, Re, and Al as the strongest positive contributors to hardness, as well as Fe, Co, and Ni as the strongest negative contributors, are identified with an interpretability test using permutation feature importance (PFI) and Shapley additive explanations (SHAP). Validation on the held-out compositions showed prediction errors below 7%. This study indicates that the composition-only models could form the basis for an effective and interpretable approach for rapid HEAs hardness screening. However, the investigated stacking strategy offers limited benefit compared to the best-performing baseline model.