Axial Capacity Prediction of CFST Columns Based on Machine Learning and Transfer Learning

This study presents a comprehensive investigation into predicting the axial bearing capacity of Concrete-Filled Steel Tubular (CFST) columns with machine learning and transfer learning. Nine different machine learning models, including traditional regression methods, ensemble techniques, and gradient boosting methods, were systematically evaluated and compared. Feature engineering based on mechanical principles of CFST was conducted to enhance model performance. The result showed that the XGBoost model demonstrated the best predictive accuracy with a test R2 of 0.9936. SHapley Additive exPlanations (SHAP) analysis revealed that diameter-to-thickness ratio, confinement factor, and concrete strength were the most influential parameters affecting axial capacity. In addition, transfer learning was employed to extend the knowledge from circular CFST columns to rectangular ones, achieving an R2 of 0.9624 through parameter transfer with retraining. This research demonstrates machine learning effectively predicts performance of CFST and demonstrates the effectiveness of knowledge transfer between different structural configurations.