Novel Ensemble Prediction Framework for Fast and Accurate Phase Prediction in FeNiCoCrAlCu High‐Entropy Alloys by Integrating Data Augmentation and Interpretable Methods

ABSTRACT

Accurate phase prediction in high‐entropy alloys (HEAs) remains challenging because of vast compositional spaces and limited available data. Targeting the complex FeNiCoCrAlCu system, we propose a novel machine learning framework. To overcome the small‐sample limitation of the dataset (204 samples), a physics‐constrained Gaussian mixture model (GMM) was utilized for data augmentation. Combined with 13 physical descriptors, a Pearson correlation–principal component analysis (PCA) strategy was employed to extract 8 optimal principal components. The reduced principal component space was further mapped back to the original feature domain to identify physically meaningful original descriptors. In this way, the problem of multicollinearity was effectively alleviated. Subsequently, a stacking ensemble model integrating 12 base algorithms and an XGBoost meta‐learner was constructed. It demonstrated strong robustness in multiphase classification, achieving a classification accuracy of 0.8377, an F1 score of 0.8377, and an average AUC of 0.9551, thereby significantly outperforming single algorithms. Furthermore, SHapley Additive exPlanations (SHAP) analysis elucidated the dominant physical parameters governing phase stability, successfully bridging data‐driven predictions with underlying physical metallurgy. This framework enables high‐accuracy phase prediction and provides theoretical support for designing novel high‐performance alloys.