Type selection and parameter optimization of FRP-steel-concrete composite columns based on machine learning interpretability analysis

FRP-steel-concrete composite members mainly exist in two forms: FRP-confined concrete-filled steel tubular columns (FRP-CFST) and FRP-concrete-steel double-skin tubular columns (DSTC). With high capacity, light weight, and good durability potential, these members are well-suited for marine applications such as offshore bridges and coastal structures, though long-term issues such as FRP interfacial degradation and steel corrosion require attention in actual marine environments. However, fragmented studies and the lack of a unified design system hinder the quantitative comparison and optimal selection of these members. To address this, this paper proposes an intelligent selection and optimization framework integrating experimental data, finite element simulations, and machine learning interpretability analysis. A comprehensive database of 585 samples was constructed from 308 experimental and 277 simulation datasets, covering a broad parameter space for both member types. The member type was treated as an explicit feature, and a unified CatBoost-based prediction model for axial compressive capacity was established, achieving an R 2 of 0.957 on the test set, significantly outperforming existing formulas. SHAP-based interpretability analysis revealed parameter contribution differences across member types, confirming consistency with mechanical principles. The prediction model was then integrated with the NSGA-III algorithm for integrated optimization of type selection and parameter configuration. Results indicate that high capacity preference favors solid sections with large diameters, whereas lightweight preference favors hollow sections with small diameters and thin-walled steel tubes. Compared with traditional designs, the optimized solutions reduce weight by 11%-22% and cost by 7%-8% with little to no reduction in load-bearing capacity.