Scholarlink

<p>High-entropy alloys (HEAs) favor solid solution phases and exhibit superior properties in irradiation environments. However, the radiological risks of Co necessitate Co-free designs. In this work, different data augmentation techniques, i.e., synthetic minority oversampling technique (SMOTE) and Wasserstein generative adversarial network (WGAN) with gradient penalty, are evaluated for predicting body-centered cubic (BCC) phase stability in Co-free HEAs. Principal component analysis (PCA) reduces dimensionality. A simple multilayer perceptron trained on these augmented datasets achieved cross-validation accuracies of 84.39% (SMOTE) and 86.82% (WGAN). Shapley additive explanations (SHAPs) and PCA analyses show valence electron concentration’s dominance of SMOTE-generated data per Hume-Rothery rules. WGAN-generated data offer balanced thermodynamic contributions for enhanced generalizability. These results show that for predicting phases in materials with small datasets, simple algorithm supported by data augmentation techniques can be as effective as complex algorithms. This work can guide HEA design for niche-subset applications and broaden the utility of machine learning for small datasets.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Interpretable machine learning for Co-free high-entropy alloy phase prediction: balancing the role of valence electron concentration

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月23日

查看全文
  • Wenyi Huo

摘要

High-entropy alloys (HEAs) favor solid solution phases and exhibit superior properties in irradiation environments. However, the radiological risks of Co necessitate Co-free designs. In this work, different data augmentation techniques, i.e., synthetic minority oversampling technique (SMOTE) and Wasserstein generative adversarial network (WGAN) with gradient penalty, are evaluated for predicting body-centered cubic (BCC) phase stability in Co-free HEAs. Principal component analysis (PCA) reduces dimensionality. A simple multilayer perceptron trained on these augmented datasets achieved cross-validation accuracies of 84.39% (SMOTE) and 86.82% (WGAN). Shapley additive explanations (SHAPs) and PCA analyses show valence electron concentration’s dominance of SMOTE-generated data per Hume-Rothery rules. WGAN-generated data offer balanced thermodynamic contributions for enhanced generalizability. These results show that for predicting phases in materials with small datasets, simple algorithm supported by data augmentation techniques can be as effective as complex algorithms. This work can guide HEA design for niche-subset applications and broaden the utility of machine learning for small datasets.