Answers are grounded in this paper and cite their sources.
Machine learning (ML) has emerged as an advanced approach to reduce the time and cost associated with traditional trial-and-error experiments. However, the accuracy of ML models is closely related to data quality and availability, while HEA datasets are commonly small and imbalanced. This work proposes a Normality-Constrained Adaptive Synthetic Minority Oversampling Technique (NCA-SMOTE)-assisted ML framework to investigate the effects of Al, Cu, and Si co-doping on phase transformation and hardness. To prevent information leakage, data augmentation was restricted to the training data and was performed independently within each cross-validation training fold. For phase prediction, SMOTE improved the class-balanced performance of the RF classifier, as reflected by Macro-F1, while maintaining comparable overall accuracy. For hardness regression, Gaussian Process Regression (GPR) exhibited the best overall performance; NCA-SMOTE increased the test R2 from 0.906 to 0.916 and reduced the RMSE from 70.1 to 66.2 HV. Guided by the coupled phase-hardness framework, Si0.05Al0.5Cu0.3CrFeNi (FCC + BCC), Si0.05AlCu0.3CrFeNi (BCC), and Si0.25Al1.2Cu0.1CrFeNi (BCC) were screened and experimentally prepared. SHapley Additive exPlanations (SHAP) further identified valence electron concentration (VEC) as the dominant descriptor for phase classification and atomic size difference (δ) as the leading descriptor for hardness. This work provides a compact, data-augmented, and interpretable ML strategy for targeted HEA design.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: