Data-driven and explainable AI models for evaluating bond strength in reinforced concrete at elevated temperatures

The bond between steel and concrete in reinforced concrete (RC) and fibre-reinforced concrete (FRC) structures is a multifaceted and intricate phenomenon. It refers to the adhesion and mechanical interlock between the steel reinforcement bars and the surrounding concrete matrix. The bond becomes more complex at elevated temperatures; however, having an accurate estimate is a crucial factor in design. Therefore, this paper employs advanced machine learning (ML) techniques to predict bond strength ( T b ) at both ambient and elevated temperatures from a 394-point experimental database, which includes additional variables such as fibre content, geometric ratios, and thermal parameter conditions. Seven models were built and assessed, including Linear Regression (LR), Gradient Boosting (GB), Extreme Gradient Boosting (XGBoost), Artificial Neural Network (ANN), k -nearest Neighbours (KNN), Decision Tree (DT), and Deep Learning (DLearning) Regressors. The GB, XGBoost, and DT models offered the best prediction results with R² above 0.95 for the testing datasets, lowest error metrics (mean absolute error (MAE) between 0.8 and 1.1 MPa), and highest reliability (a30%-index ≥ 90%), all outperforming those reported in earlier literature. According to SHapley Additive exPlanations (SHAP) analysis, the length-to-diameter ratio ( l d ) and failure surface temperature ( T ) dominated as the predictors, followed by concrete compressive strength ( f c ), and cover-to-diameter ratio ( c d ), which is according to the existing mechanics of bond and thermal degradation. This study presents resolutions regarding the promise of data-driven models to accurately, reliably, and interpretably predict bond strength in post-fire conditions, which is of great merit in terms of resilient design practice. Future work may investigate hybrid ML–mechanistic frameworks and the integration of full-scale fire testing to further enhance engineering applicability.