SHAP-interpreted XGBoost prediction framework for the shear capacity of externally bonded FRP-strengthened RC beams

Abstract

Externally bonded fiber-reinforced polymer (EB-FRP) strengthening improves the shear capacity of reinforced concrete (RC) beams, but prediction remains challenging because shear resistance depends on interacting geometric, material, reinforcement, strengthening, and failure-mechanism variables. This study develops a SHAP-interpreted XGBoost framework for predicting the ultimate shear capacity of unanchored EB-FRP-strengthened RC beams using a 315-specimen experimental database. Three formulations were examined to distinguish scenario-based and design-stage use: a 21-input model with the failure-mode descriptor treated as a prescribed scenario variable, a 20-input direct regression model excluding this descriptor, and a two-stage pipeline in which the failure mode is first predicted from design-stage variables and then used for shear-capacity prediction. Model performance was evaluated using statistical metrics, prediction-ratio bias indicators, feature-selection stability, SHAP interpretation, baseline ML models, and design provisions. The complete two-stage pipeline achieved the best testing performance, with R 2 = 0.9415, RMSE = 26.266 kN, and MAE = 18.369 kN, while the direct 20-input model achieved the highest A20 of 81.05%. SHAP and parametric checks confirmed mechanically plausible trends, including shear-span, effective-bond-height, transverse-reinforcement, and FRP-strain effects, supporting practical design-stage EB-FRP shear-capacity assessment within the database range.