Generative adversarial networks for enhanced performance prediction of square CFST members under axial tension

Taking square concrete-filled steel tubular (CFST) members under axial tension as the research object, a three-dimensional mesoscopic finite element model was established based on the experimental results of six specimens. Ten parametric models were further developed to investigate the effects of section size, confinement coefficient, and slenderness ratio on tensile performance. In addition, code-based comparisons and machine learning predictions were carried out. The results indicate that the finite element simulations agree well with the test results, with the ratios of simulation results to test results all being below 0.95, indicating that the simulation predictions are within a reasonable range of the experimental data, which reflects good agreement. The parametric analysis shows that when the confinement coefficient increases from 0 to 0.99, the maximum load rises from 182 kN to 895 kN; when the slenderness ratio increases from 8 to 20, the maximum load exhibits an overall decreasing trend. The code comparison shows that the predictions from the Chinese code are closer to the finite element results, with an average error of approximately 4.57%. To improve prediction accuracy with limited data, a Generative Adversarial Network (GAN)-based data augmentation method was employed. Using both original and WGAN-GP-augmented data, predictive models were developed. Among these models, the Random Forest model achieved the best overall performance. On the augmented test set, the coefficients of determination (R2) for ultimate load and displacement prediction reached 0.997 and 0.9855, respectively. The findings provide a reference for tensile performance analysis and rapid assessment of this type of member, demonstrating the effectiveness of GAN-based data augmentation in enhancing predictive accuracy.