Scholarlink

<p>Accurate prediction of the swelling pressure (<i>P</i><sub>s</sub>) of bentonite is essential for the design and long-term safety of geotechnical barriers in high-level radioactive waste (HLW) disposal systems. For this purpose, this study proposed a data-driven framework for both point and interval predictions of <i>P</i><sub>s</sub> by integrating data augmentation, machine learning (ML) and explainable techniques. A dataset comprising 755 samples with <i>P</i><sub>s</sub> and 8 key influencing parameters of <i>P</i><sub>s</sub> was augmented using the synthetic minority over-sampling technique (SMOTE) to address data imbalance and enhance model generalization. Three ML models, namely extreme gradient boosting (XGBoost), long short-term memory (LSTM), and gated recurrent unit (GRU), were trained on both the original and augmented datasets. Among them, the XGBoost-SMOTE model achieved the best point prediction performance. Interval predictions were generated by combining point prediction results with kernel density estimation (KDE), and the XGBoost-SMOTE model yielded the best performance at 95%, 93%, and 90% confidence levels, with corresponding coverage width criterion (<i>CWC</i>) values of 0.5368, 0.4790, and 0.4211, respectively. For further validation of the superior predictive capability of the XGBoost-SMOTE model, the predicted value of <i>P</i><sub>s</sub> was compared with the value calculated by the diffuse double layer (DDL) theory. Furthermore, shapley additive explanations (SHAP) analysis and partial dependence plots (PDPs) identified initial void ratio as the most influential parameter and revealed the underlying relationships between input parameters and <i>P</i><sub>s</sub>. This study provided an approach for improving the understanding and accurate prediction of the swelling behavior of bentonite buffer materials.</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 with augmented data for predicting bentonite swelling pressure

  • 期刊论文

  • 发表日期：

    2026年6月29日

查看全文
  • Wenyan Shi,
  • Feng Zhang,
  • Changdong Li,
  • Qingbing Liu,
  • Long Xu,
  • Zhangrong Liu,
  • Hangjing Zhou,
  • Meichen Liu,
  • Yuxian Jiang

摘要

Accurate prediction of the swelling pressure (Ps) of bentonite is essential for the design and long-term safety of geotechnical barriers in high-level radioactive waste (HLW) disposal systems. For this purpose, this study proposed a data-driven framework for both point and interval predictions of Ps by integrating data augmentation, machine learning (ML) and explainable techniques. A dataset comprising 755 samples with Ps and 8 key influencing parameters of Ps was augmented using the synthetic minority over-sampling technique (SMOTE) to address data imbalance and enhance model generalization. Three ML models, namely extreme gradient boosting (XGBoost), long short-term memory (LSTM), and gated recurrent unit (GRU), were trained on both the original and augmented datasets. Among them, the XGBoost-SMOTE model achieved the best point prediction performance. Interval predictions were generated by combining point prediction results with kernel density estimation (KDE), and the XGBoost-SMOTE model yielded the best performance at 95%, 93%, and 90% confidence levels, with corresponding coverage width criterion (CWC) values of 0.5368, 0.4790, and 0.4211, respectively. For further validation of the superior predictive capability of the XGBoost-SMOTE model, the predicted value of Ps was compared with the value calculated by the diffuse double layer (DDL) theory. Furthermore, shapley additive explanations (SHAP) analysis and partial dependence plots (PDPs) identified initial void ratio as the most influential parameter and revealed the underlying relationships between input parameters and Ps. This study provided an approach for improving the understanding and accurate prediction of the swelling behavior of bentonite buffer materials.