Modeling Nonergodic Ground Motions using a Graph Neural Network

Eduardo Arzabala, Kyle B. Withers, Morgan P. Moschetti, Tim Clements, & Ian W. McBrearty

Submitted September 7, 2025, SCEC Contribution #14585, 2025 SCEC Annual Meeting Poster #TBD

We use a deep learning approach, specifically Graph Neural Networks (GNNs), to develop a non-ergodic ground-motion model (GMM) from CyberShake. Unlike ergodic GMMs, non-ergodic GMMs require region-specific data or simulations for calibration. We use 2 million simulated earthquakes to train a non-ergodic GNN model for southern California. GNNs excel at learning from inherent spatial relationships, such as source-to-site interactions in our case. This non-ergodic GMM includes location-specific effects (e.g., near-surface shear wave velocity) and path effects (e.g., rupture distance). This non-ergodic GNN model also incorporates (1) the absolute locations (i.e., latitude and longitude) of sources, sites, and the relative distance between them, and (2) the relative distance between sites. By using input parameters consistent with those in ergodic GMMs, as well as new parameters, we ensure compatibility with existing ground-motion predictions and seismic hazard workflows. Preliminary results indicate that the GNN-based model captures site-specific effects like basin amplification of ground motions and azimuthal variations of ground motion with respect to the fault, features that are often missed by traditional GMMs. Future work includes validating the model with empirical earthquake observations. These results suggest that GNNs offer a powerful and flexible alternative to other non-ergodic GMM approaches. The GNN model has the ability to generalize to different earthquake sources in southern California and may provide useful insights for future updates of GMMs.

Citation
Arzabala, E., Withers, K. B., Moschetti, M. P., Clements, T., & McBrearty, I. W. (2025, 09). Modeling Nonergodic Ground Motions using a Graph Neural Network. Poster Presentation at 2025 SCEC Annual Meeting.

Related Projects & Working Groups
Ground Motions (GM)