Parametric study, modified direct strength method and Deep-BP prediction for local-global buckling resistance of thin-walled aluminium alloy double-web members under axial compression

This study investigates the local-global buckling resistance of thin-walled aluminium alloy double-web members under axial compression using finite-element modelling, a modified direct strength method (DSM), and data-driven prediction. A refined finite-element model was developed in ABAQUS and validated against 48 axial compression tests on 6061-T6 and 6082-T6 aluminium alloy double-web members. The model reproduced the observed failure modes, load-midspan displacement responses, and ultimate resistances, with a mean finite-element-to-test resistance ratio of 1.037, a standard deviation (SD) of 0.021, and a maximum deviation of 7.088%. A database comprising 3792 numerical results was then generated by varying the alloy grade, buckling axis, member slenderness, and cross-sectional geometry. The results show that the normalised slenderness ratio governs the transition from cross-sectional strength or local-buckling control to global flexural-buckling control. Increasing either the outstand-flange width-to-thickness ratio or the web height-to-thickness ratio reduces the axial resistance by impairing local plate stability and flange-web interaction. A modified DSM formulation was subsequently proposed to account for local-global buckling interaction in double-web aluminium alloy members. In parallel, a Bayesian-optimised deep back-propagation (Deep-BP) neural network was developed for resistance prediction. For the training, validation, and test sets, respectively, the proposed model achieved NRMSE values of 0.0121, 0.0184, and 0.0181; MAPE values of 2.097%, 2.489%, and 2.370%; and R² values of 0.9962, 0.9935, and 0.9935. Reliability analyses further showed that both the proposed DSM formulation and the Deep-BP model outperform the existing design specifications, with the Deep-BP model exhibiting the lowest prediction dispersion.