Scholarlink

<p>The precise estimation of the shear strength of reinforced concrete (RC) beams constructed with recycled aggregate concrete (RAC) is essential for the secure and sustainable design of structural components. The utilization of construction and demolition waste (CDW) to produce recycled aggregate concrete (RAC) is an attractive approach from both an environmental and budgetary perspective. However, this study proposes a single and novel hybrid machine learning framework to predict the shear strength of RAC beams using a dataset compiled from published experimental studies and validated numerical models. Ensemble learning techniques such as Support Vector Regression (SVR), Gradient Boosted Regression Trees (GBRT), CatBoost, Decision Tree (DT), and Bagging Regressor (BR) were developed to train models using six input features: 28-day compressive strength of concrete (fc), the percentage of recycled coarse aggregate (RCA), the effective depth of the beam cross-section (d), the width of the beam cross-section (b), the percentage of longitudinal reinforcement (rhow), the shear span to effective depth ratio (a/d) and output parameter is the shear strength of the specimen (V<sub>test</sub>). Furthermore, evaluating model performance, we used R<sup>2</sup>, RMSE, MAPE, and MAE metrics on a robust database that was divided into training (70%) and testing (30%) phases. Results show that the hybrid models outperform standalone algorithms, with the hybrid GBRT model combination achieving the highest prediction accuracy throughout both stages using R<sup>2</sup> (0.869, 0.998), RMSE, MAE, and MAPE (20.033, 12.753, and 15.598%), respectively. Additionally, Shapley Additive Explanations (SHAP) analysis was employed to determine significant input characteristics and clarify how they affect the Beam Shear Strength Prediction. The presence of the width of the beam cross-section and the shear span to effective depth ratio contributes the highest positive influence to the outcome. This study demonstrates that hybrid ML approaches can reliably capture nonlinear interactions among RAC beam variables, offering a powerful alternative to empirical formulas. Moreover, a Graphical User Interface (GUI) was developed to enable designers to effectively and economically forecast beam shear strength, and the experimental findings substantially impact the construction industry by facilitating a more accurate and reliable implementation of RAC.</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

Beam shear strength prediction of recycled aggregate concrete using explainable artificial intelligence

  • 期刊论文

  • 发表日期：

    2025年8月19日

查看全文
  • Sourov Paul,
  • Lesanoor Rahman,
  • Asmaul Husna Chara,
  • Md Mahmuduzzaman,
  • Abul Kashem,
  • Md Naim,
  • Rafrad Bhuiyan,
  • Somir Chandra Malo

摘要

The precise estimation of the shear strength of reinforced concrete (RC) beams constructed with recycled aggregate concrete (RAC) is essential for the secure and sustainable design of structural components. The utilization of construction and demolition waste (CDW) to produce recycled aggregate concrete (RAC) is an attractive approach from both an environmental and budgetary perspective. However, this study proposes a single and novel hybrid machine learning framework to predict the shear strength of RAC beams using a dataset compiled from published experimental studies and validated numerical models. Ensemble learning techniques such as Support Vector Regression (SVR), Gradient Boosted Regression Trees (GBRT), CatBoost, Decision Tree (DT), and Bagging Regressor (BR) were developed to train models using six input features: 28-day compressive strength of concrete (fc), the percentage of recycled coarse aggregate (RCA), the effective depth of the beam cross-section (d), the width of the beam cross-section (b), the percentage of longitudinal reinforcement (rhow), the shear span to effective depth ratio (a/d) and output parameter is the shear strength of the specimen (Vtest). Furthermore, evaluating model performance, we used R2, RMSE, MAPE, and MAE metrics on a robust database that was divided into training (70%) and testing (30%) phases. Results show that the hybrid models outperform standalone algorithms, with the hybrid GBRT model combination achieving the highest prediction accuracy throughout both stages using R2 (0.869, 0.998), RMSE, MAE, and MAPE (20.033, 12.753, and 15.598%), respectively. Additionally, Shapley Additive Explanations (SHAP) analysis was employed to determine significant input characteristics and clarify how they affect the Beam Shear Strength Prediction. The presence of the width of the beam cross-section and the shear span to effective depth ratio contributes the highest positive influence to the outcome. This study demonstrates that hybrid ML approaches can reliably capture nonlinear interactions among RAC beam variables, offering a powerful alternative to empirical formulas. Moreover, a Graphical User Interface (GUI) was developed to enable designers to effectively and economically forecast beam shear strength, and the experimental findings substantially impact the construction industry by facilitating a more accurate and reliable implementation of RAC.