Scholarlink

<p>Intergranular stress corrosion cracking (IGSCC) of austenitic stainless steels remains a critical degradation mechanism in light water reactors (LWRs), driven by coupled mechanical, electrochemical, and microstructural factors. Accurate prediction of crack growth rates (CGR) is essential for effective monitoring and mitigation. Here, we develop a comprehensive machine learning (ML) framework to model CGR in both boiling water reactors (BWRs) and pressurized water reactors (PWRs) using a curated database, integrating experimental measurements with ML-based and physics-informed preprocessing. Categorical boosting (CatBoost) was used for point prediction and uncertainty quantification (UQ) of CGR, and compared with natural gradient boosting (NGBoost), Gaussian process regression (GPR), and TabNet. CatBoost outperformed other models and, when combined with post-hoc calibration, provided reliable uncertainty estimates. Shapley Additive Explanations (SHAP) applied to CatBoost revealed both individual and interaction effects of features on CGR and enabled comparison of CGR behavior between BWR and PWR environments.</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

Machine learning for modeling intergranular stress corrosion cracking of stainless steels in light water reactors with uncertainty quantification and explainability

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月11日

查看全文
  • Dayu Fajrul Falaakh,
  • Chi Bum Bahn

摘要

Intergranular stress corrosion cracking (IGSCC) of austenitic stainless steels remains a critical degradation mechanism in light water reactors (LWRs), driven by coupled mechanical, electrochemical, and microstructural factors. Accurate prediction of crack growth rates (CGR) is essential for effective monitoring and mitigation. Here, we develop a comprehensive machine learning (ML) framework to model CGR in both boiling water reactors (BWRs) and pressurized water reactors (PWRs) using a curated database, integrating experimental measurements with ML-based and physics-informed preprocessing. Categorical boosting (CatBoost) was used for point prediction and uncertainty quantification (UQ) of CGR, and compared with natural gradient boosting (NGBoost), Gaussian process regression (GPR), and TabNet. CatBoost outperformed other models and, when combined with post-hoc calibration, provided reliable uncertainty estimates. Shapley Additive Explanations (SHAP) applied to CatBoost revealed both individual and interaction effects of features on CGR and enabled comparison of CGR behavior between BWR and PWR environments.