Machine learning prediction of axial compressive behavior of circular FRP-concrete-steel double-skin tubular stub columns

Hybrid FRP-Concrete-Steel Double-Skin Tubular Columns (DSTCs) are innovative composite structural components that combine FRP, concrete, and steel to achieve exceptional seismic performance and corrosion resistance. While extensive experimental research and finite element simulations have validated their performance, the use of machine learning (ML) to predict their axial compressive behavior quickly and accurately remains underexplored. In this study, a comprehensive database was constructed, including 126 experimental datasets from 20 published studies and 2970 finite element (FE) simulation datasets. This integrated database was employed to train and evaluate multiple ML models, including XGBoost, Random Forest, LightGBM, Artificial Neural Network, and an Ensemble Learning model, focusing on predicting two critical parameters that govern the ultimate state of confined concrete in DSTCs (i.e., axial compressive peak strength and ultimate strain). To enhance the interpretability of the ML predictions, a semi-empirical theoretical model was combined with the optimal ML model to reconstruct stress-strain curves for confined concrete in circular DSTCs. Comparative analyses revealed that the Ensemble Learning model achieved the highest predictive accuracy, outperforming other ML models. The SHapley Additive exPlanation (SHAP) method was utilized to interpret the optimal model, identifying the most influential features governing axial compressive behavior, such as material properties, geometric dimensions, and confinement effects. Furthermore, a graphical user interface (GUI) was developed to enable real-time, user-friendly predictions of axial compressive peak strength, ultimate strain, and stress-strain curves of confined concrete in DSTCs. This study not only advances the understanding of DSTC behavior under axial compression but also provides a practical design tool for structural engineers. The integration of ML, theoretical modeling, and user-interface technology offers a novel paradigm for predictive modeling and structural design optimization.