Abstract: In retrofitting systems with external bonded carbon fiber-reinforced polymer (CFRP) materials, the interfacial performance between CFRP and steel plays an important role in the strengthening efficiency which is mainly characterized by the ultimate load bearing capacity. Semi-empirical prediction models with fracture mechanics and test results have been proposed. Accuracy and feasibility of these models are affected by experimental conditions and the number of experimental results. In this study, a database of bond strength comprising of 587 CFRP-to-steel single/double-lap shear joints was first collected. A prediction model for the failure mode of CFRP-to-steel interfaces based on the XGBoost algorithm was proposed. With prior physical knowledge involved, a prediction model for the ultimate load bearing capacity of CFRP-to-steel interfaces was established through a physics and data dual-driven method. The input characteristics and output results of the models were analyzed for interpretability using the Shapley Additive Explanations (SHAP) and Partial dependence Plot (PDP) methods. The models exhibited reasonable results, with the accuracy of 97.95% for failure mode predictions, as well as the fitting coefficient R² of 0.944 and RMSE of 4.829 for ultimate load bearing capacity predictions. This study provides useful suggestions for the analysis of CFRP strengthened steel structures.