Scholarlink

<p>Accurate calculation of the compressive stiffness of micropiles (<InlineEquation ID="IEq1"><EquationSource Format="TEX">\(\:K_{s}\left(C\right)\)</EquationSource></InlineEquation>) is essential for forecasting load–displacement behavior and maintaining foundation serviceability in geotechnical structures. Conventional analytical and numerical methods frequently oversimplify soil-structure interaction and require substantial calibration, thereby limiting their applicability across diverse ground conditions. This paper presents a data-driven predictive approach that combines supervised machine learning techniques with a field-based micropile (<InlineEquation ID="IEq2"><EquationSource Format="TEX">\(\:MP\)</EquationSource></InlineEquation>) test database to address these limitations. A comprehensive dataset of 393 in-situ MP compression experiments was compiled after statistical preprocessing, including normalization, randomization, and outlier elimination based on the interquartile range criterion. Nine geotechnical and geometric characteristics were utilized as predictors of <InlineEquation ID="IEq3"><EquationSource Format="TEX">\(\:K_{s}\left(C\right)\)</EquationSource></InlineEquation>. Five ensemble learning models—Gradient Boosting (<InlineEquation ID="IEq4"><EquationSource Format="TEX">\(\:GB\)</EquationSource></InlineEquation>), Light Gradient Boosting (<InlineEquation ID="IEq5"><EquationSource Format="TEX">\(\:LGB\)</EquationSource></InlineEquation>), Histogram-based Gradient Boosting (<InlineEquation ID="IEq6"><EquationSource Format="TEX">\(\:HGB\)</EquationSource></InlineEquation>), Extreme Gradient Boosting (<InlineEquation ID="IEq7"><EquationSource Format="TEX">\(\:XGB\)</EquationSource></InlineEquation>), and Categorical Boosting (<InlineEquation ID="IEq8"><EquationSource Format="TEX">\(\:CB\)</EquationSource></InlineEquation>)—were created and refined with the Parrot Optimization Algorithm (<InlineEquation ID="IEq9"><EquationSource Format="TEX">\(\:POA\)</EquationSource></InlineEquation>) for hyperparameter optimization. The <InlineEquation ID="IEq10"><EquationSource Format="TEX">\(\:XG{B}_{POA}\)</EquationSource></InlineEquation> algorithm demonstrated the greatest prediction reliability. Comparative analyses demonstrated that <InlineEquation ID="IEq11"><EquationSource Format="TEX">\(\:XG{B}_{POA}\)</EquationSource></InlineEquation> decreased prediction error by 10–22% compared to other boosting models while ensuring enhanced convergence stability. The proposed <InlineEquation ID="IEq12"><EquationSource Format="TEX">\(\:POA\)</EquationSource></InlineEquation>-optimized boosting framework offers a precise, interpretable, and computationally efficient method for calculating <InlineEquation ID="IEq13"><EquationSource Format="TEX">\(\:K_{s}\left(C\right)\)</EquationSource></InlineEquation> directly from field data. This hybrid modeling methodology reconciles empirical testing with predictive analytics, providing a pragmatic solution for performance-oriented <InlineEquation ID="IEq14"><EquationSource Format="TEX">\(\:MP\)</EquationSource></InlineEquation> design and foundation system optimization in geotechnical engineering.</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

An interpretable AI framework using XGB-POA for micropile compressive stiffness prediction

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月1日

查看全文
  • Mohammadreza Ahmadi Golsefidi,
  • Mahzad Esmaeili-Falak,
  • Hossein Sarbaz

摘要

Accurate calculation of the compressive stiffness of micropiles (\(\:K_{s}\left(C\right)\)) is essential for forecasting load–displacement behavior and maintaining foundation serviceability in geotechnical structures. Conventional analytical and numerical methods frequently oversimplify soil-structure interaction and require substantial calibration, thereby limiting their applicability across diverse ground conditions. This paper presents a data-driven predictive approach that combines supervised machine learning techniques with a field-based micropile (\(\:MP\)) test database to address these limitations. A comprehensive dataset of 393 in-situ MP compression experiments was compiled after statistical preprocessing, including normalization, randomization, and outlier elimination based on the interquartile range criterion. Nine geotechnical and geometric characteristics were utilized as predictors of \(\:K_{s}\left(C\right)\). Five ensemble learning models—Gradient Boosting (\(\:GB\)), Light Gradient Boosting (\(\:LGB\)), Histogram-based Gradient Boosting (\(\:HGB\)), Extreme Gradient Boosting (\(\:XGB\)), and Categorical Boosting (\(\:CB\))—were created and refined with the Parrot Optimization Algorithm (\(\:POA\)) for hyperparameter optimization. The \(\:XG{B}_{POA}\) algorithm demonstrated the greatest prediction reliability. Comparative analyses demonstrated that \(\:XG{B}_{POA}\) decreased prediction error by 10–22% compared to other boosting models while ensuring enhanced convergence stability. The proposed \(\:POA\)-optimized boosting framework offers a precise, interpretable, and computationally efficient method for calculating \(\:K_{s}\left(C\right)\) directly from field data. This hybrid modeling methodology reconciles empirical testing with predictive analytics, providing a pragmatic solution for performance-oriented \(\:MP\) design and foundation system optimization in geotechnical engineering.