Modeling Nonergodic Ground Motions Using Graph Neural Networks

ABSTRACT We develop a nonergodic ground-motion model (GMM) for southern California by training a graph neural network (GNN) on approximately 400,000 simulated earthquakes from CyberShake. Our analysis focuses on long-period responses at rupture distances of 200 km or less. Each rupture is modeled as a graph, with nodes representing CyberShake sites connected to their closest neighbors. The resulting nonergodic, or region-specific, GMM includes site conditions, source properties, and path properties, as well as absolute source and site locations (via latitude and longitude), relative source-to-site distances, and intersite distances. The GNN model predictions are residuals relative to an ergodic GMM. Using the graph structure, the GNN learns spatially correlated residual patterns, such as source-to-site and local site interactions. Consequently, the GNN captures site-specific and spatially varying features, such as basin amplification and azimuthal variations of ground motion with respect to the fault, features that are often missed by traditional ergodic GMMs. In addition, the GNN model maintains magnitude scaling and period-dependent trends that are consistent with empirical relations across sources in southern California. We select the 2019 Mw 7.1 Ridgecrest earthquake as a case study to evaluate the performance of the GNN. By leveraging the spatial and physical characteristics of CyberShake simulations, the GNN achieves lower mean residuals than ergodic GMMs at T = 5.0 s and shows comparable performance at other periods. GNN-based nonergodic GMMs offer a promising framework for ground-motion modeling and hazard analysis.