Seismic demand modeling leveraing conditional generative adversarial network

Prediction of structural seismic responses under varying ground motion intensities is critical for advancing seismic resilience and optimizing structural designs.Surrogate models offer a costeffective alternative to computationally expensive high-fidelity numerical simulations.However, past surrogate modeling research largely focused on supervised machine learning techniques (Random Foresting, ANN, k-NN).Generative AI models, particularly the generative adversarial network (GAN), have received emerging research interest, but have yet to be explored for seismic surrogate modeling.This study proposes a surrogate modeling approach leveraging conditional variational auto-encoder generative adversarial network (CVAE-GAN) as a robust means to predict the engineering demand parameters (EDPs) of bridges subjected to varying ground motion intensity measures (IMs).We consider a case-study dataset of highway bridge seismic responses encompassing various input variables including intensity measures, geometrical properties, and material properties and six EDPs.The CVAE-GAN framework integrates an encoder, decoder, and discriminator to ensure accurate and realistic predictions.The encoder maps input conditioning structural features and a noise vector into a latent space representation, which is then fed into the decoder to reconstruct the structural responses.The discriminator refines the predictions even more by differentiating real from generated data, using a composite loss function combining the mean squared error for reconstruction accuracy with Kullback-Leibler (KL) divergence as a regularization term in latent space and adversarial loss for statistical realism.These features give cGANs significant advantages compared with conventional supervised machine learning paradigms in a context such as seismic surrogate modeling because they build the underlying data distribution in an adversarial framework and give rise, therefore, to context-aware predictions that better capture the inherent variability across responses of structural systems.The training through adversarial means engendered by a generator-discriminator duo has a factored regularizing effect on the model, insulating it against those topmost forms of overfitting indicative of conventional, supervised techniques.Besides, the representation in the encoder latent space partially lessens the impact of the highly dimensional data through true compression of essential information, thereby ensuring computational efficiency and focusing on critical dependencies.Through addressing computational demands and overcoming the limitations of existing techniques, the cGANbased surrogate model makes hazard analysis and performance-based design within structural engineering scalable and accurate.