Bayesian-optimized tree-based models for predicting the shear strength of U-shaped externally bonded FRP-strengthened RC beams

Abstract

The rehabilitation of aging concrete infrastructure using externally bonded fiber-reinforced polymer (EB-FRP) systems has emerged as a crucial solution in civil engineering. Yet, accurate prediction of their shear-strengthening contribution remains challenging due to complex failure mechanisms and behavioral uncertainties. This study investigates the application of Tree-based machine learning models for predicting the shear strength contribution of U-shaped EB-FRP systems in reinforced concrete beams. Three distinct approaches—Decision Tree, Random Forest, and CatBoost—were developed and evaluated using a refined database of 189 experimental specimens, encompassing diverse beam configurations and strengthening parameters. The methodology incorporates Bayesian optimization through the Optuna framework for systematic hyperparameter tuning, ensuring optimal model performance. The CatBoost model demonstrated superior predictive capabilities, maintaining exceptional consistency across training (R² = 0.92, VAF = 92.55%) and testing phases (R² = 0.90, VAF = 89.91%), significantly outperforming Decision Tree and Random Forest models. Comparative analysis against current design guidelines (ACI 440.2R-17, fib Bulletin 90, and TR-55) revealed substantial improvements in prediction accuracy, with the CatBoost model reducing mean absolute error by approximately 65% compared to code provisions. The results highlight the potential of advanced machine learning techniques in capturing the complex nonlinear relationships governing FRP shear contribution, offering a reliable tool for preliminary design and validation of strengthening systems. This study contributes to the growing integration of data-driven approaches in structural engineering practice, particularly in the context of FRP-strengthening applications.