Prediction, interpolation and extrapolation of the bonding strength between FRP bars and UHPC with machine learning

Accurately predicting the bond strength between Fiber-Reinforced Polymer (FRP) bars and Ultra-High-Performance Concrete (UHPC) is essential for the design of FRP–UHPC components. This study develops a simplified bond strength prediction model based on the extrapolation analysis of machine learning (ML) models to enable reliable design beyond existing experimental data. An ensemble ML framework was constructed and optimized through data preprocessing and algorithmic comparison. The model’s extrapolation capability was verified using Basalt Fiber Reinforced Polymer (BFRP)–UHPC pull-out tests, demonstrating strong consistency between predictions and experiments. When trained on a dedicated FRP–UHPC database, the model achieved Root Mean Squared Error (RMSE) = 3.64, Mean Absolute Error (MAE) = 2.58, and coefficient of determination (R²) = 0.96; when expanded to include FRP–Normal Concrete (NC) and Steel–UHPC datasets, it maintained robust accuracy (RMSE = 9.97, MAE = 6.40, R² = 0.92). Based on parameter analysis of 77,200 sets of data points, a simplified bond strength calculation formula was derived using the least-squares method, incorporating the effects of concrete strength, cover thickness, bar type, surface profile, elastic modulus, diameter, anchorage length, and specimen type. The proposed formula achieved RMSE = 6.69, MAE = 4.80, and R² = 0.49, offering a practical tool for engineering design and providing a foundation for future FRP–UHPC research.