Machine learning models for predicting load-slip curves of shear stud connectors in solid concrete slabs

The resistance, stiffness, and ductility of steel-concrete composite structures depend on the distribution of connector shear forces, which is affected by the connector deformations under the load. This paper presents several machine learning (ML) models with optimized hyperparameters for predicting load-slip curves of headed shear studs in solid slabs composed of lightweight or normal weight concrete with a compression strength ranging from 30 to 113 MPa. The developed models are based on the following ML algorithms: extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), gradient boosting with categorical features support (CatBoost), and natural gradient boosting (NGBoost). The first three models predict deterministic (mean) values of the relative shear load within the stud based on the specified slip; concrete compressive strength; stud tensile strength, diameter, and height; slab reinforcement position; and concrete density. The NGBoost model produces mean and probabilistic values of the stud relative shear load from the same inputs. The models were developed using a database of 10,950 load and slip measurements from 180 push tests compiled by the present authors from the literature. Model predictions were interpreted using the SHapley Additive exPlanations (SHAP) method, which indicated that the relative shear load is most significantly affected by slip, followed by concrete compressive strength and stud tensile strength, which were found to be significantly less important for model predictions. The shear stiffness and slip capacity determined from the predicted curves in accordance with prEC4 and AISC 360 showed good agreement with those obtained from the experimental curves. The developed ML models outperformed several existing load-slip models. A web application was developed and published online to predict load-slip curves using the CatBoost and NGBoost models, and to determine the shear stiffness and slip capacity from the predicted curves using various criteria. The proposed models provide important information for future numerical and analytical studies investigating the dependence of the resistance, stiffness, and ductility of steel-concrete composite structures on stud deformations, thereby facilitating the development of improved design provisions.