Abstract
This study investigates the use of machine learning (ML) models to predict the ultimate load capacity of demountable shear connectors in steel-concrete composite structures. A dataset of 239 experimental and numerical records was assembled, incorporating critical features such as bolt diameter, bolt yield and ultimate strengths, concrete and grout compressive strengths, and multiple interfacial friction coefficients. Eight supervised ML algorithms were evaluated: Linear Regression, Ridge, Lasso, K-Nearest Neighbors (KNN), Support Vector Regression (SVR), Decision Tree, Random Forest, and XGBoost. Model performance was assessed using R², Mean Absolute Error (MAE), and Mean Squared Error (MSE). Among these, XGBoost and Random Forest delivered the best predictive accuracy, with R² values of 0.9477 and 0.9255, respectively, outperforming other methods across all evaluation metrics. SHAP analysis was employed to explain model behavior and identify the most influential features. The analysis revealed that bolt diameter, steel ultimate strength, and bolt-to-concrete friction were the most impactful predictors of shear capacity. These insights were supported by feature importance heatmaps generated from the top-performing models. Despite the robust performance of advanced algorithms, the study acknowledges challenges such as data imbalance and minimal influence of certain parameters, which may affect model generalizability.