Physics-guided GAN-based prognostic framework for seismic hysteresis loop and damage pattern in RC bridge columns

Hysteretic behavior, damage patterns, and structural performance, e.g., Damage index (DI), are the critical seismic characteristic of reinforced concrete (RC) bridge columns. In this study, a novel bridge seismic hysteresis loop and damage pattern prognostic framework is proposed, including a sequence-to-sequence hysteresis prediction network based on Gated Recurrent Units (GRU) and a damage pattern generative model based on a conditional generative adversarial network (CGAN), namely HysGRU and HysGAN, respectively. Through training on 160 hysteresis loops, the physics-guided HysGRU model, which incorporates constraints derived from numerical simulation, achieves a root mean square error (RMSE) of 47.57 kN and an R 2 score of 94.55%. This represents substantial improvements over purely simulation, with a 63.53% reduction in RMSE and a 35.56% increase in R 2 score, demonstrating the effectiveness of integrating physics-based constraints into the deep learning framework. Moreover, the physics features of hysteretic behavior are extracted as a condition of HysGAN. To evaluate the performance of HysGAN, a pre-trained DI prediction network is adopted to calculate the R 2 scores from the DI values prediction of the predicted damage patterns. Extensive experiments have shown that conditioning with latent features can improve quality, stability, and generalization capability. HysGAN achieves R 2 scores of 91.65% and 91.62% on validation and test datasets, significantly improving over the baseline model with increases of 4.68% and 54.62%, respectively. Upon completion of the model training process, engineers can automatically derive the hysteresis loop and damage patterns given specific column design parameters and selected drift ratios with the proposed HysGRU and HysGAN, facilitating an efficient assessment of seismic design and capacity. • Forecasting of hysteresis loop and damage pattern for RC bridge column is presented. • The HysGRU, guided by physics simulation, enhances the prediction of hysteresis loop. • The HysGAN, with physics latent features as input, synthesizes the damage pattern. • Incorporation of physics information boosts the generalizability under limited data. • Damage quantification using the proposed HysGAN is presented.