Buckling behavior and interpretable machine-learning of CFRP-strengthened circular steel tubes under axial compression

This study develops a physics-informed and interpretable data-driven framework for predicting the axial buckling capacity of circular steel tubes strengthened with longitudinal CFRP strips. A validated finite element (FE) model against test results was used to generate 196 parametric samples, and these data were combined with 190 experimental results from literature to establish a 386-sample database for machine-learning modeling. The FE analysis shows that CFRP strengthening increases the sectional flexural stiffness and can improve the ultimate load by up to 49.57% for slender members. For medium-length columns, a distinct capacity jump of up to 30.12% is observed, which is attributed to a transition from global flexural buckling to localized instability induced by the longitudinal stiffness constraint of high-modulus CFRP strips. Four prediction models (XGBoost, BPNN, SVR and Stacking) were optimized using four swarm-intelligence algorithms to overcome the sensitivity of model performance to hyperparameter selection. Pearson correlation analysis identifies the slenderness ratio as the dominant capacity-reducing variable ( r = -0.81), whereas the number of CFRP layers, the diameter-to-thickness ratio and the steel yield strength show positive correlations with buckling resistance ( r = 0.25, 0.25 and 0.13, respectively). Among all optimized models, AVOA-XGBoost provides the best overall performance, with a testing R 2 of 0.925, a MAE of 18.40 kN and a RMSE of 30.11 kN. SHAP analysis further confirms that the learned feature effects are consistent with structural stability theory. The proposed framework offers both mechanistic insight into CFRP-assisted instability control and a practical surrogate model for the rehabilitation design of steel tubular members.