Explainable multi-model machine learning framework for tensile property prediction of FeNiCrCoCu high-entropy alloys

Reliable and interpretable prediction of tensile properties is critical for accelerating the design of high-entropy alloys (HEAs), yet many existing machine learning approaches emphasize predictive accuracy without sufficient consideration of physical interpretability or deployment efficiency. In this study, we propose X-HEA-ML, an explainable multi-model machine learning framework for predicting key tensile properties of FeNiCrCoCu high-entropy alloys, namely Young’s modulus, ultimate tensile strength (UTS), and yield strength. A dataset of 918 samples was generated using molecular dynamics simulations, incorporating compositional, microstructural, and thermomechanical descriptors. The framework integrates a diverse set of regression models, including linear models, decision trees, support vector machines, ensemble learners, Gaussian Process Regression (GPR), neural networks, and kernel-based methods. All models are trained under identical conditions and evaluated using a multi-objective benchmarking strategy that combines prediction accuracy (RMSE, MAE), variance-explanation capability (R²), and deployment-oriented efficiency metrics (prediction speed, training time, and model size). Across all tensile properties, Gaussian Process Regression (GPR) consistently achieved the highest predictive accuracy and variance-explanation capability among the evaluated regression models, with Squared Exponential and Matern 5/2 kernels yielding R² values exceeding 0.98 for Young’s modulus and UTS, and 0.96 for yield strength. The results further suggest that probabilistic kernel-based learning is well suited for modeling the smooth but highly nonlinear structure–property relationships represented in the investigated FeNiCrCoCu molecular dynamics dataset, highlighting the potential of uncertainty-aware regression for interpretable and reliable materials informatics. While compact neural networks and polynomial SVMs provide competitive accuracy–efficiency trade-offs, GPR demonstrated the most balanced performance among the evaluated models in terms of predictive accuracy, robustness, and computational practicality. To ensure transparency, Shapley value–based explainable machine learning is employed to quantify feature contributions and directional effects. The analysis reveals a consistent feature hierarchy across all properties, identifying Cu and Cr contents as the dominant factors associated with the observed tensile-response trends, while grain size and temperature exhibit secondary influence within the investigated MD parameter space. By combining systematic multi-model benchmarking, deployment-aware evaluation, and explainable learning, the proposed X-HEA-ML framework extends conventional black-box prediction by providing interpretable insights that may support tensile-property prediction, alloy screening, and future data-driven HEA design.