Data-Driven Shear Capacity Prediction of Studs Embedded in UHPC for Steel–UHPC Composite Structures

This study employs ensemble learning algorithms to accurately predict the shear capacity of studs embedded in ultrahigh performance concrete (UHPC), a critical aspect of structural engineering. Using a comprehensive data set featuring six UHPC slab features, seven stud features, and one shear pocket feature, 35 ensemble models are developed across five cases and seven algorithms (bagging, random forest, AdaBoost, gradient boosting, XGBoost, CatBoost, and LightGBM). The Shapley additive explanations (SHAP) analysis is conducted on the top-performing models to gain insights into their decision-making mechanisms. Results indicate that transforming UHPC slab geometry parameters into a volume, alongside different diameter ratios, proves to be an effective feature engineering strategy. The gradient boosting algorithm consistently demonstrates superior performance in cases with transformed features and all features, while CatBoost outperforms in other cases. Importantly, all ensemble models surpass the performance of mechanics-based equations, as evidenced by the metrics mean absolute error (MAE), root-square-mean error (RSME), and coefficient of determination R2. For instance, the best gradient boosting model achieves an MAE of 11.243 kN, RSME of 14.874 kN, and R2 of 0.988, compared to 27.230 kN, 53.314 kN, and 0.869, respectively, in the best empirical equation. Additionally, insights from the SHAP analysis suggest a steel fiber volume of 2.2% in UHPC, along with recommended stud transverse spacing to diameter (st/dst) and longitudinal spacing to diameter (sl/dst) ratios of 4.5 and 12, respectively, for the design of grouped studs in UHPC slabs. The study underscores the importance of evaluating different ensemble algorithms, highlighting the need for a comprehensive approach in model selection and analysis.