Abstract Accurately predicting the ultimate shear capacity of perfobond rib (PBL) connectors is of great significance for the design of steel–concrete composite structures. This study predicted the ultimate shear capacity of PBL using machine learning (ML) methods. Initially, a dataset comprising 233 sets of PBL push-out test data was established. To enhance data quality, an Isolation Forest was used to identify and eliminate outliers from the dataset. Subsequently, four ML models—XGBoost, DT, RF, and ANN—were trained on this dataset to predict the ultimate shear capacity of PBL. By comparing and analyzing the prediction results, XGBoost demonstrated the best predictive performance with an R 2 value of 0.97, outperforming other models. Then, a visual analysis, including SHAP and PDP, was conducted on the XGBoost model, revealing the contribution levels of different features to the predicted values. The analysis found that the number of perforated holes ( n ) had the greatest impact. Moreover, based on the analysis of visualizations, recommended ranges of values for the input features are provided to maximize the ultimate shear capacity of the PBL connectors. In comparison with traditional formulas, the trained ML models exhibit superior accuracy. The MAE of XGBoost is approximately 10% of that of the traditional formulas, and its RMSE value is less than 20% of those.
Hu et al. (Wed,) studied this question.