Generative AI for Spectrum-Compatible Accelerogram Synthesis Using a Conditional VAE–GAN

Reliable nonlinear time-history analysis requires artificial ground motions that are compatible with prescribed response spectra while retaining the temporal, frequency-domain, and nonstationary characteristics of recorded earthquakes. This study develops a conditional variational autoencoder–generative adversarial network (CVAE–GAN) for directly generating spectrum-compatible acceleration time histories. The model takes a target 5%-damped absolute acceleration response spectrum as the conditioning input and produces diverse 60-s accelerograms without relying on seed records or iterative spectral-matching procedures. A large waveform dataset was constructed using 141,582 east–west acceleration records from the Japanese K-NET and KiK-net strong-motion networks. PhaseNet was employed to identify P-wave arrivals, after which the records were uniformly resampled to 50 Hz and processed into fixed-length waveforms. The proposed architecture combines a probabilistic encoder, a conditional waveform generator, and a spectrum-conditioned discriminator. Variational inference establishes a continuous latent space for stochastic sampling, while adversarial training based on the Wasserstein loss with gradient penalty improves waveform realism and training stability. Evaluation using earthquakes within the training magnitude range of 4.5≤Mw≤7.54.5 \leq M_w \leq 7.54.5≤Mw​≤7.5 shows that more than 99% of the generated response spectra achieve coefficients of determination greater than 0.8, with root-mean-square errors predominantly below 0.1 m/s20.1\,\mathrm{m/s^2}0.1m/s2. The generated accelerograms also reproduce major characteristics of recorded motions, including the evolution of strong shaking, coda-wave attenuation, characteristic Fourier-frequency content, and realistic phase and phase-difference distributions. Additional evaluations for earthquakes outside the training magnitude range indicate reasonable generalization to both lower- and higher-magnitude events, although limitations remain for underrepresented long-period and long-duration records. The proposed framework provides an efficient data-driven approach for generating large ensembles of realistic, diverse, and spectrum-compatible accelerograms for structural seismic analysis, performance-based design, and probabilistic seismic assessment.