Scholarlink

<p>Accurate estimation of shear wave velocity (V<sub>s</sub>) is critical for offshore geotechnical design, yet direct measurements remain sparse due to cost and logistical constraints. Empirical correlations of V<sub>s</sub> from cone penetration test data are derived for specific conditions, thus introducing uncertainty when applied more generally. Machine learning (ML) based correlations have become popular, yet to date have prioritised accuracy over interpretability. To address this gap and enhance transparency, this study integrates the computationally efficient XGBoost technique with SHapley Additive exPlanations (SHAP) to predict V<sub>s</sub> and attribute prediction contributions to individual input features. A combined open-source dataset of 7485 paired cone penetration test data with pore pressure measurement (CPTu) and V<sub>s</sub> measurements was integrated and used for training and validation. SHAP analysis on a testing dataset of 1526 samples shows that depth, corrected cone resistance (q<sub>t</sub>), and sleeve friction (f<sub>s</sub>) are the most influential features, with depth increasing in importance when the V<sub>s</sub> predicted deviates from the mean V<sub>s</sub> in the training dataset. Compared to a widely used empirical correlation, the ML approach demonstrated superior accuracy across most cases, while also offering insight into decision-making logic. This study highlights the value of interpretable ML in offshore site investigations, with this specific CPTu-V<sub>s</sub> interpretable ML model particularly relevant to bottom-fixed foundation designs for offshore wind developments, which are governed by stiffness criteria, in V<sub>s</sub> data-limited projects.</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 XGBoost-based predictions of shear wave velocity from CPTu data

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月20日

查看全文
  • Héctor Marín-Moreno,
  • James Willis,
  • Yuting Zhang,
  • Susan Gourvenec

摘要

Accurate estimation of shear wave velocity (Vs) is critical for offshore geotechnical design, yet direct measurements remain sparse due to cost and logistical constraints. Empirical correlations of Vs from cone penetration test data are derived for specific conditions, thus introducing uncertainty when applied more generally. Machine learning (ML) based correlations have become popular, yet to date have prioritised accuracy over interpretability. To address this gap and enhance transparency, this study integrates the computationally efficient XGBoost technique with SHapley Additive exPlanations (SHAP) to predict Vs and attribute prediction contributions to individual input features. A combined open-source dataset of 7485 paired cone penetration test data with pore pressure measurement (CPTu) and Vs measurements was integrated and used for training and validation. SHAP analysis on a testing dataset of 1526 samples shows that depth, corrected cone resistance (qt), and sleeve friction (fs) are the most influential features, with depth increasing in importance when the Vs predicted deviates from the mean Vs in the training dataset. Compared to a widely used empirical correlation, the ML approach demonstrated superior accuracy across most cases, while also offering insight into decision-making logic. This study highlights the value of interpretable ML in offshore site investigations, with this specific CPTu-Vs interpretable ML model particularly relevant to bottom-fixed foundation designs for offshore wind developments, which are governed by stiffness criteria, in Vs data-limited projects.