Simulation of near-fault strong ground motion with generative adversarial network

This paper proposes a generative adversarial network (GAN) for feature extraction and simulation generation of seismic data. Specifically, a Wasserstein-distance-based GAN model has been adopted using the well-acknowledged NGA-West2 ground motion American database of the Pacific Earthquake Engineering Research Center (PEER). Notably, this study adopts multi-dimensional convolution/deconvolution to extract and reconstruct multi-dimensional features from one-dimensional data, ensuring the generated data exhibit required features simultaneously. The validity and diversity of the generated data are verified by four aspects, such as time–frequency power spectrum, data similarity, and seismic intensity parameter distribution map. The results show that the generated data by the proposed method have good and diverse features and waveforms.