Prediction of hysteretic shapes for reinforced concrete columns using conditional generative adversarial networks

The hysteretic response of reinforced concrete (RC) columns is essential for understanding the seismic capacity of RC buildings. Traditional methods for estimating this response often rely on experimental tests or numerical simulations, which are labor-intensive, costly, and sometimes unstable. This study presents HysGAN-RC, a conditional generative adversarial network model developed to predict the hysteretic behavior of RC columns. The model was trained using 258 hysteretic loops of RC column specimens encompassing flexure, flexure-shear, and shear failure modes. By incorporating the Wasserstein distance with a gradient penalty, HysGAN-RC improves the baseline CGAN performance. The model’s effectiveness was evaluated using Intersection over Union (IoU) and Fréchet Inception Distance (FID) metrics under various conditions. Numerical experiments show that HysGAN-RC achieves a mean IoU of 0.68 and an FID of 7.2 when utilizing ten RC design parameters, including geometric, material, and reinforcement details. The model’s performance is further supported by a mean Structural Similarity Index of 0.90, indicating strong agreement between generated and real hysteretic curves in both visual shape and mechanical behavior. Notably, HysGAN-RC can predict hysteretic shapes across varying design conditions without requiring prior knowledge of the failure type, positioning it as a promising tool for high-fidelity seismic assessment of RC columns.