Scholarlink

<p>Monotonic ductility is a fundamental parameter for evaluating the deformation capacity and structural redundancy of reinforced members. Concrete-encased steel beams combine the ductility of steel with the stiffness of concrete; however, conventional design approaches often neglect flange geometry effects. This study integrates finite element simulations under displacement-controlled monotonic loading with data-driven modeling to address this limitation. A dataset of 250 concrete-encased steel beams was generated, and machine learning models were developed to predict their monotonic ductility index (<i>D</i><sub><i>i</i></sub>). Results show that nonlinear approaches, such as Support Vector Regressor (<i>SVR</i>), Gradient Boosting Regressor (<i>GBR</i>), and Deep Neural Networks (<i>DNN</i>), outperform linear models. An optimized ensemble framework (<i>SVR</i>, <i>GBR</i>, and <i>DNN</i>) achieves a high predictive accuracy, with an <i>R</i><sup><i>2</i></sup> of 0.747, providing a robust tool for preliminary structural screening. Feature importance analysis reveals that the trained model assigns nearly five times higher importance to the normalized height ratio compared to the width ratio within the studied parametric bounds. The study utilizes SHapley Additive exPlanations (<i>SHAP</i>) to provide a mechanistic understanding of how flange geometry influences global ductility.</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

Hybrid finite element and machine learning approach for ductility prediction of concrete encased steel beams

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月22日

查看全文
  • Ahmed Youssef Kamal,
  • Rasha Abd El Ghany

摘要

Monotonic ductility is a fundamental parameter for evaluating the deformation capacity and structural redundancy of reinforced members. Concrete-encased steel beams combine the ductility of steel with the stiffness of concrete; however, conventional design approaches often neglect flange geometry effects. This study integrates finite element simulations under displacement-controlled monotonic loading with data-driven modeling to address this limitation. A dataset of 250 concrete-encased steel beams was generated, and machine learning models were developed to predict their monotonic ductility index (Di). Results show that nonlinear approaches, such as Support Vector Regressor (SVR), Gradient Boosting Regressor (GBR), and Deep Neural Networks (DNN), outperform linear models. An optimized ensemble framework (SVR, GBR, and DNN) achieves a high predictive accuracy, with an R2 of 0.747, providing a robust tool for preliminary structural screening. Feature importance analysis reveals that the trained model assigns nearly five times higher importance to the normalized height ratio compared to the width ratio within the studied parametric bounds. The study utilizes SHapley Additive exPlanations (SHAP) to provide a mechanistic understanding of how flange geometry influences global ductility.