Advanced finite element and machine learning-based prediction of axial behavior of FRP-confined double-skin tubular columns

Abstract Double-Skin Tubular Columns (DSTCs) with Fiber-Reinforced Polymer (FRP) confinement consist of an outer FRP tube, an inner steel tube, and a concrete core. Although extensive research has focused on hollow DSTCs, investigations on DSTCs with concrete-filled inner steel tubes remain limited. Considering their expanding use in structural applications, it is essential to understand how concrete infill and its mechanical properties influence overall performance. In this study, predictive models for axial load capacity and confined ultimate strain are developed by integrating finite element modeling (FEM) with machine learning (ML) techniques. A combined dataset of 85 experimental tests and 67 FEM simulations was compiled, from which five influential parameters were selected as ML input features. The predictive capability of advanced ML algorithms—Gradient Boosting (GB), Random Forest (RF), Convolutional Neural Networks (CNNs), and Long Short-Term Memory (LSTM)—was examined alongside traditional approaches such as Multiple Linear Regression (MLR), Support Vector Regression (SVR), and SVR enhanced with Empirical Mode Decomposition (EMD). Model performance was assessed using regression error characteristic curves, SHAP-based interpretability, and statistical metrics across 107 validated simulated and experimental cases. The EMD-SVR, GB, and RF models demonstrated superior accuracy for predicting confined ultimate load, achieving R 2 values of 0.99, 0.989, and 0.960, respectively. For confined strain prediction, GB and EMD-SVR achieved R 2 values of 0.690 and 0.99. Overall, the developed models provide engineers with reliable tools to estimate axial load capacity and deformation characteristics under varying design conditions, thereby enhancing material optimization and structural performance evaluation.