Conditional generative adversarial networks for spectrum-compatible ground motion generation of scenario earthquakes

Obtaining acceleration time histories compatible with the design response spectrum is essential for structural seismic analysis . However, traditional methods for generating spectrum-compatible acceleration time histories often fail to incorporate physical characteristics, and the response spectrum cannot adequately represent key features such as spectral characteristics and non-stationarity. Therefore, this paper proposes a spectrum-compatible ground motion generation model, referred to as the conditional generative adversarial network-response spectrum (CGAN-RS). The model integrates a response spectrum discriminator (RS-D) and conditional variables, including moment magnitude, rupture distance, fault type, and time-averaged shear-wave velocity at the top 30 m, to ensure spectrum compatibility and physical consistency in the generated acceleration time histories. By integrating the response spectrum discriminator (RS-D) and conditional variables, the CGAN-RS model ensures spectrum compatibility and physical consistency in the generated acceleration time histories. Comparative analysis in the time-frequency domain between generated and recorded data under identical conditions reveals high similarity in peak ground acceleration , waveform characteristics, Arias intensity, and Fourier amplitude spectra . Furthermore, under identical scenario earthquakes, the CGAN-RS model achieves significant improvements in the coefficient of determination (R 2 ) between the target spectrum and the mean and median spectra of 50 generated spectra, with increases of 39.9 % and 19.7 % over the traditional CGAN model. Additionally, the mean absolute error (MAE) and the mean squared error (MSE) demonstrate substantial improvements. For the mean spectrum, the MAE is reduced by 62.23 %, and the MSE by 77.37 %. For the median spectrum, the MAE is reduced by 66.70 %, and the MSE by 83.99 %.