A Machine Learning-Based Method for Predicting Bond Strength Between FRP Bars and Concrete

This study investigates the bond strength between fiber-reinforced polymer （FRP） bars and concrete. A database was constructed by integrating 487 pull-out test data points from 16 literature sources， featuring 12 input variables such as fiber type， forming process， and surface characteristics. Seven machine learning algorithms， such as RF， XGBoost， and LightGBM， were employed for prediction. The influence mechanisms of the 12 variables on the bond strength were analyzed and compared. The prediction results indicated the compressive strength of concrete， along with the tensile strength and elastic modulus of the FRP bars， as the key factors affecting the bond strength. Among them， tree models showed superior performance to non-tree models， with the LightGBM algorithm exhibiting high accuracy and efficiency. SHAP value evaluation revealed that forming processes （e.g.， the RB process） had a greater influence on the bond strength than fiber type did. Furthermore， the significance of all involved variables was quantitatively ranked using the LightGBM algorithm， showing that the mechanical properties of the materials and the surface characteristics of the FRP bars had a more significant effect on the bond strength.