Interpretable hybrid machine learning frameworks for predicting the axial load capacity of concrete filled double skin steel tube columns

Abstract

Predicting the axial load capacity of concrete-filled double-skin steel tube (CFDST) columns is difficult because of the complex mechanical interactions between the concrete core and the dual steel tubes. Although machine learning (ML) methods provide advantages in prediction, current models frequently lack sufficient hyperparameter optimization and mechanistic interpretability. To resolve these limitations, novel hybrid ML frameworks are proposed integrating eXtreme Gradient Boosting (XGBoost) with advanced metaheuristic algorithms, specifically the Bald Eagle Search (BES) and Aquila Optimizer (AO). A database containing 153 experimental samples defined by eight geometric and material parameters was used to develop and validate the models. Systematic optimization comparisons demonstrated that the XGB-BES framework provides high generalization and accuracy, producing a coefficient of determination (R²) of 0.998, a Root Mean Square Error (RMSE) of 62.054 kN, and an A20 index of 0.957 when tested on validation data. Furthermore, the inherent opacity of complex ML models was addressed using post-hoc interpretability methods such as SHAP (Shapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations). Analyses of global and local feature importance demonstrated that the diameter and thickness of the outer tube most strongly affect the structural capacity, which connects the computational predictions to fundamental structural mechanics principles. These findings indicate that the developed XGB-BES model functions as an accurate and mechanically transparent instrument for the evaluation and design of CFDST columns.