Machine learning models for predicting load-slip curves of shear studs welded in deck slab ribs transverse to beams

Deformations of shear studs welded within deck slab ribs, transverse to the supporting steel beams, influence the performance and resistance of the beams under loads. This paper presents six machine learning models with optimized hyperparameters for predicting load-slip curves of such shear connectors, including XGBoost, LightGBM, CatBoost, NGBoost, SVR, and ANN. In contrast to other developed models predicting deterministic mean values of the relative shear load, the NGBoost model produces both mean and probabilistic estimates from the specified slip, normal force on the slab face, transverse stud spacing within rib, distance from the rib top corner to the nearest stud center, concrete compressive strength, stud shank diameter, and stud height after welding. The models were developed using a database of 7306 load and slip measurements from the published results of 198 push tests. Model predictions were interpreted using the SHAP method, which indicated that the relative shear load is most strongly affected by slip, followed by the other model features, whose contributions are considerably less important than slip. The sensitivity of the model outputs to uncertainties in feature values, as well as the feature interactions in the developed models, was evaluated and discussed. The shear stiffness and slip capacity obtained from the predicted curves in accordance with EC4 and AISC 360 showed good agreement with those obtained from the experimental curves. The CatBoost and NGBoost models, which were found to be more accurate than the others, have been integrated into an online web application. It predicts load-slip curves from user inputs and determines shear stiffness and slip capacity from the curves using various criteria. The proposed models, when used within their applicability limits outlined in the paper, offer valuable tools for future numerical and analytical studies exploring how resistance, stiffness, and ductility of steel-concrete composite structures depend on stud deformations, thereby facilitating the development of improved design provisions. These models also enable designers to predict stud deformation characteristics without physical testing.