Modeling Ground Motion Using Conditional Generative Model

Maxime Lacour, Pu Ren, Rie Nakata, Nori Nakata, & Michael M. Mahoney

Submitted September 7, 2025, SCEC Contribution #14845, 2025 SCEC Annual Meeting Poster #TBD

Recent approaches to model spatially varying (non-ergodic) ground-motion models use Gaussian processes to learn the spatial variability of ground motion from observations and extrapolate them at new source and site locations. Gaussian processes are fully characterized by their mean and covariance functions, which are defined analytically in each model and therefore remain relatively simple (Sung et. al. 2023, Lacour et. al. 2024). These simple forms restrict the spatial extrapolation of Gaussian processes to patterns constrained by their spatial correlation function.

We present a novel approach to model ground motion using neural-network based generative models, called Conditional Generative Model for Fourier Amplitude Spectra (CGM-FAS). CGM-FAS is a FAS version of the previous CGM-GM framework that successfully generates ground motion waveforms using a conditional variational autoencoder (Ren et al., 2024), and higher resolution wavefields using a diffusion model (CGM-Wave; Bi et al., 2025). Using CGM-FAS has several key advantages over Gaussian processes. Firstly, neural networks can learn the spatial variations without the restriction of the simple functional form required for Gaussian processes, allowing us to predict maps of ground motion that have more complex features than using Gaussian processes. CGM-FAS also significantly reduces the computational cost to generate ground motion maps when compared to using Gaussian processes. Most of the computational efforts using neural networks are required during the training phase, and forward predictions only involve basic linear algebra operations through the network.

We illustrate the CGM-FAS approach in the San Francisco Bay Area region using small-magnitude earthquakes (M < 4) for which FAS data is available within the frequency range [2Hz; 15Hz]. We compare our predictions with the recent non-ergodic empirical model from Lacour et. al. (2024) developed for that same region.

Citation
Lacour, M., Ren, P., Nakata, R., Nakata, N., & Mahoney, M. M. (2025, 09). Modeling Ground Motion Using Conditional Generative Model. Poster Presentation at 2025 SCEC Annual Meeting.

Related Projects & Working Groups
Ground Motions (GM)