This study evaluates the efficiency of an explainable ensemble learning framework in precisely predicting the bond strength between steel sections with different surface treatments and various concrete types. Besides seven numerical features, two categorical features, surface treatment and concrete type, that significantly affect bond strength but were omitted in existing equations were considered as inputs for such framework. A database comprising 302 push-out test results was carefully constructed. Four standalone machine learning models, multiple linear regression (MLR), multilayer perceptron artificial neural network (MLP-ANN), support vector machine (SVM), classification and regression tree (CART), as well as two latest ensemble models, adaptive boosting (Adaboost), light gradient boosting machine (LightGBM) were developed. Their performance was thoroughly compared in terms of learning ability, computational cost, predictive accuracy, and residuals. Comparative study show that the ensemble model LightGBM outperforms other models and their predictive performance can be ranked as LightGBM>AdaBoost>CART>SVM>MLP-ANN>MLR. The superior reliability of the ensemble model is further confirmed through comparisons with empirical equations. The SHapley Additive exPlanations method (SHAP) is also employed to explain the contributions of essential features to the individual prediction by LightGBM. Finally, a LightGBM-SHAP-based web application with a user-friendly interface was created, which enables convenient and efficient bond performance estimation without cumbersome programming.