Mechanism-aware physics-guided XGBoost model for shear strength prediction of FRP-strengthened RC beams

Shear strengthening of existing reinforced concrete beams via externally bonded fiber-reinforced polymer (FRP) systems is widely practiced; however, accurate prediction of shear capacity remains challenging due to complex interaction mechanisms, regime-dependent failure behavior, and the limited generalization of existing design-code equations. Code-based models calibrated on restricted experimental data often exhibit bias and large scatter across varying geometries, shear span-depth ratios, and FRP layouts, whereas purely data-driven machine-learning approaches typically lack physical interpretability. To address these limitations, a mechanism-aware physics-guided machine-learning framework is proposed for predicting the shear strength of FRP-strengthened reinforced concrete beams. The framework decomposes total shear resistance into a mechanics-based concrete contribution and a data-driven residual component modeled via XGBoost, enabling variance reduction and physically informed learning. A compiled database of 275 experimental beam tests is classified into low \:a/d≤\:2.5 and high \:a/d>2.5 shear span-depth regimes to reflect distinct dominant failure mechanisms. In the high $$\:a/d$$ regime, the model achieves R² = 0.927, RMSE = 19.33 kN, and MAE = 14.58 kN, whereas in the low $$\:a/d$$ regime it attains R² = 0.771, RMSE = 72.67 kN, and MAE = 44.69 kN, consistent with greater mechanism variability. Explainable-AI analyses confirm regime-dependent parameter influence and saturation of FRP effectiveness, demonstrating mechanically coherent and interpretable predictions. Potential applications to reliability-based assessment and design-code calibration require additional uncertainty quantification and external validation beyond the scope of this study.