An interpretable machine learning approach for shear capacity prediction of prestressed UHPC bridge girders with GUI deployment

Shear failure in prestressed ultra-high-performance concrete (UHPC) bridge girders is abrupt and brittle, making reliable shear-capacity prediction essential for safe design. To overcome limitations in current design codes, this study leverages machine learning (ML) techniques, utilizing a comprehensive experimental database from the literature. Fourteen ML regressors were trained and tested under a consistent evaluation framework, and predictive performance was assessed using R², MAE, RMSE, and MAPE. Ensemble-based models provided the best generalization, with AdaBoost achieving the highest accuracy on the test set, followed by XGBoost and Random Forest. To benchmark practical design applicability, predictions from the best-performing ML models were compared with four widely used analytical approaches including NF P18–210, fib, AASHTO, and PCI. The analytical methods exhibited systematic bias and noticeably larger dispersion relative to the ML estimators with NF P18–210 tending to be conservative, while fib, AASHTO and PCI showed broader scatter with occasional overprediction, indicating higher uncertainty for data ranges represented in the database. Model interpretability was examined using SHAP, which identified geometric descriptors such as section size and depth as the dominant drivers of shear capacity, with secondary effects from material and reinforcement parameters. Finally, a web-based graphical interface was implemented to deploy the optimized AdaBoost predictor for rapid, user-friendly shear-strength estimation in practice.