Tree-based machine learning models for predicting the bond strength in reinforced recycled aggregate concrete

To address the ever-increasing environmental degradation caused by concrete construction, utilizing recycled aggregate (RA) in concrete mixes offers a significant solution. This study aims to assess the bond strength of both plain and deformed steel rebars in recycled aggregate concrete (RAC) using machine learning (ML) methods. The ML models employed include Decision Tree (DT), AdaBoost, CatBoost, Gradient Boosting, and Extreme Gradient Boosting (XGB). A comprehensive dataset of 158 pull-out tests from previous studies was collected. The features investigated associated with both concrete and rebar characteristics, namely recycled and natural coarse aggregates (RCA and NCA), fine aggregates, cement, water, the water-to-cement ratio (w/c), concrete compressive strength ( \({f}_{c}{\prime})\) $f_c\prime )$ , yield strength of steel rebar \(({f}_{y})\) $(f_y)$ , rebar type and diameter, and bond length. The findings highlighted that, before hyperparameter tuning, the CatBoost regressor, outperformed the other ML models with \({R}^{2}\) $R^2$ score and RMSE value of 0.94, and 3, respectively. However, after hyperparameter tuning, the XGBoost regressor was the most accurate model, achieving an impressive \({R}^{2}\) $R^2$ score of 0.94, and an RMSE value of 3. Furthermore, according to the Shapley values applied to the XGB model, the features \({f}_{c}{\prime}\) $f_c\prime$ , \({f}_{y}\) $f_y$ , and bond length were found to have the highest impact on the bond strength of the studied specimens. Whereas, the RAC replacement level has minimal impact on the target value.