Scholarlink

<p>Accurate estimation of axial pile bearing capacity remains challenging due to soil heterogeneity and the limitations of empirical design methods. This study compares <i>TabNet</i>, an interpretable deep learning architecture for tabular data, with <i>XGBoost</i>, a reference standard machine learning model for predicting axial pile capacity. Both models were trained on 440 load tests from Olson’s <i>Axial Pile Capacity</i> database with explicitly layered soil representations and tested using 184 load tests from the same database. Models were optimized using a consistent <i>Bayesian</i> framework and evaluated under strict site-separated training and testing to assess true generalization. While <i>XGBoost</i> achieved near-perfect training performance, its accuracy degraded on unseen sites, indicating overfitting. In contrast, <i>TabNet</i> showed weaker training fit but superior test performance. Feature-importance and stepwise mask analyses reveal that <i>TabNet</i> distributes importance more evenly across pile geometry and layered soil properties and provides intrinsic interpretability aligned with geotechnical reasoning. The results demonstrate that <i>TabNet</i> offers improved generalization and transparency compared to <i>XGBoost</i> for predicting axial pile capacity under realistic geotechnical uncertainty. </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: a comparison of TabNet and XGBoost for axial pile capacity prediction using layered soil representations

  • 期刊论文

  • 发表日期：

    2026年7月10日

查看全文
  • Baturalp Ozturk,
  • Magued Iskander

摘要

Accurate estimation of axial pile bearing capacity remains challenging due to soil heterogeneity and the limitations of empirical design methods. This study compares TabNet, an interpretable deep learning architecture for tabular data, with XGBoost, a reference standard machine learning model for predicting axial pile capacity. Both models were trained on 440 load tests from Olson’s Axial Pile Capacity database with explicitly layered soil representations and tested using 184 load tests from the same database. Models were optimized using a consistent Bayesian framework and evaluated under strict site-separated training and testing to assess true generalization. While XGBoost achieved near-perfect training performance, its accuracy degraded on unseen sites, indicating overfitting. In contrast, TabNet showed weaker training fit but superior test performance. Feature-importance and stepwise mask analyses reveal that TabNet distributes importance more evenly across pile geometry and layered soil properties and provides intrinsic interpretability aligned with geotechnical reasoning. The results demonstrate that TabNet offers improved generalization and transparency compared to XGBoost for predicting axial pile capacity under realistic geotechnical uncertainty.