Scholarlink

<p>This study proposes a physics-guided extreme gradient boosting (PG‑XGB) model to predict the axial capacity of circular concrete‑filled double‑skin steel tubular (CFDST) columns. For this purpose, a database of 243 tests was collected to evaluate seven existing empirical design equations. Accordingly, the best empirical equation is incorporated into the XGB model’s loss function. The resulting PG‑XGB model adaptively balances data-driven learning and physics-based predictions through two weighting parameters. Different training–test splits are examined to assess the effect of training size on model performance. The optimal PG‑XGB model achieves <i>R</i><sup>2</sup> values of (0.998 and 0.998), <i>A</i>10 values of (0.995 and 1.0), RMSE values of (101.492 and 117.086&#xa0;kN), and MAE values of (53.138 and 78.214&#xa0;kN) on the training and test subsets, respectively. Compared with the best empirical equation and several baseline tree-based models, including standard XGB, gradient boosting, AdaBoost, random forest, decision tree, and artificial neural network, the PG‑XGB model consistently improves correlation metrics and substantially reduces prediction errors, demonstrating the benefit of integrating mechanics into machine-learning-based axial capacity prediction of CFDST columns.</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 Efficient Physics-Guided Extreme Gradient Boosting Model for Axial Capacity Prediction of Circular CFDST Columns

  • 期刊论文

  • 发表日期：

    2026年7月6日

查看全文
  • Viet-Linh Tran,
  • T. Nguyen-Thoi

摘要

This study proposes a physics-guided extreme gradient boosting (PG‑XGB) model to predict the axial capacity of circular concrete‑filled double‑skin steel tubular (CFDST) columns. For this purpose, a database of 243 tests was collected to evaluate seven existing empirical design equations. Accordingly, the best empirical equation is incorporated into the XGB model’s loss function. The resulting PG‑XGB model adaptively balances data-driven learning and physics-based predictions through two weighting parameters. Different training–test splits are examined to assess the effect of training size on model performance. The optimal PG‑XGB model achieves R2 values of (0.998 and 0.998), A10 values of (0.995 and 1.0), RMSE values of (101.492 and 117.086 kN), and MAE values of (53.138 and 78.214 kN) on the training and test subsets, respectively. Compared with the best empirical equation and several baseline tree-based models, including standard XGB, gradient boosting, AdaBoost, random forest, decision tree, and artificial neural network, the PG‑XGB model consistently improves correlation metrics and substantially reduces prediction errors, demonstrating the benefit of integrating mechanics into machine-learning-based axial capacity prediction of CFDST columns.