Experimental testing and machine learning to predict the load-slip behavior of stud connectors in steel-UHPC composite structures

Machine learning techniques have demonstrated significant potential in predicting the shear mechanical performance of stud connectors. However, predicting the load-slip curves of stud connectors in steel-ultrahighperformance concrete (UHPC) composite structures remains a challenge. While some empirical models have been developed to describe the load-slip behavior of stud connectors, most are fitted to limited databases, leading to inadequate generalizability. This study presents a series of push-out tests on stud connectors encased in steelUHPC composite structures, along with a comprehensive analysis of the load-slip curve characteristics derived from the experimental results. Subsequently, eight machine learning models were trained and tested using a database comprising 289 instances from push-out tests. Fourteen key factors were selected as input parameters, and three characteristic values including shear strength, initial shear stiffness, and peak slip were chosen as output parameters. The results indicated that the Extreme Gradient Boosting (XGBoost) model exhibited excellent performance in predicting shear strength and peak slip, with corresponding R2 values of 0.988 and 0.814, respectively, while the Categorical Boosting (CatBoost) model performed best in predicting initial shear stiffness with an R2 value of 0.867. Feature importance analysis using the Shapley Additive Explanations (SHAP) method highlighted that stud diameter was a critical factor affecting shear strength and initial shear stiffness, while stud height was a critical factor influencing the peak slip. Finally, based on the predicted characteristic values, an effective model was established for predicting the load-slip curve of stud connectors in steel-UHPC composite structures, which was validated to have good accuracy and applicability via the test results.