Machine-learning prediction model for bond strength evolution of corroded rebar-concrete interface

The mechanical performance of concrete structures under corrosive environments largely depends on the bond behavior between rebar and concrete. Existing studies primarily focus on predicting the peak bond stress, paying limited attention to the complete degradation process of bond strength. To address the deterioration of bond strength caused by internal rebar corrosion in concrete structures, a comprehensive bond–slip dataset was constructed based on extensive pull-out test data from existing literature. Nine input features were selected: corrosion rate, bond length, rebar diameter, concrete compressive strength (both cube and cylinder), concrete cover thickness, rebar yield strength, rebar type, and slip. This dataset captures the full evolution of bond strength at the corroded rebar–concrete interface as a function of slip. A bond–slip prediction model for corroded rebar was developed using a stacking GBDT–SVR (Gradient Boosting Decision Tree–Support Vector Regression) machine learning approach. Feature importance analysis was conducted using the SHAP method. The results showed a strong agreement between predicted and actual values. Performance metrics such as R², σ, η, and γ confirmed the high accuracy of the model, with prediction outside 0.8–1.2 confidence band only at low bond stress values. Compared to traditional empirical formulas, the proposed model demonstrates superior precision.