Interpretable Deep Learning–Machine Learning Models for Accelerated Discovery of Elastic Properties in Refractory High‐Entropy Alloys
Abstract
Refractory high‐entropy alloys (RHEAs) offer mechanical performance at extreme temperatures, but their optimization is hindered by the vast compositional space and high cost of first‐principles calculations. This study establishes an interpretable machine‐learning benchmark for predicting nine elastic and mechanical properties of RHEAs. An EMTO‐CPA dataset comprising 2487 alloys is used to evaluate eight regression models: Gaussian process regression (GPR), support vector regression (SVR), shallow and deep neural networks, LightGBM, XGBoost, CatBoost, and Histogram‐based Gradient Boosting. Seven composition‐derived descriptors representing atomic size, electronic structure, and thermodynamic characteristics are employed. Hyperparameters are optimized using randomized search with five‐fold cross‐validation, while robustness is assessed through ten train–test evaluations using random seeds 42–51. Model performance is reported as mean ± standard deviation of MAE and R2. The results demonstrate target‐dependent performance: GPR performs particularly well for sws, SVR achieves favorable performance for several properties, and deep neural networks provide strong predictions for selected mechanical targets. SHAP analysis of optimized SVR models identifies valence electron concentration, average atomic radius, atomic size mismatch, and melting temperature as important contributors, whereas mixing entropy and mixing enthalpy generally show weaker contributions. Overall, the benchmark provides a reproducible framework for RHEA prediction and supports screening and alloy design.