Generative Modeling of Earthquake Ground Motion

Reliable earthquake ground-motion time histories are essential for seismic hazard analysis, structural design, and the risk assessment of distributed infrastructure systems. However, observed records of large earthquakes remain sparse, while physics-based simulations are computationally expensive and difficult to deploy at the scale required for uncertainty quantification. This thesis develops generative modeling frameworks for efficient, scenario-specific synthesis of earthquake ground motion, with an emphasis on neural operators, function-space learning, and scalable regional wavefield generation. The work begins with conditional generative adversarial neural operators for broadband three-component ground-motion synthesis at individual sites, conditioned on earthquake source, path, and site parameters. It then develops broader methodological foundations for functional generative modeling, including operator flow matching for stochastic-process learning and mesh-informed neural operators for domain- and discretization-agnostic generation on complex geometries. Building on these advances, the thesis introduces Ground-Motion Flow, a physics-inspired latent operator flow-matching framework for regional coherent ground-motion synthesis. By generating wavefields in a physics-aligned latent representation and reconstructing full spatiotemporal motions through neural operators, the framework produces spatially coherent earthquake wavefields conditioned on physical parameters with orders-of-magnitude acceleration relative to direct numerical simulation. Together, these studies establish a progression from single-site waveform synthesis to regional-scale, uncertainty-aware ground-motion wavefield generation. The results demonstrate the potential of functional generative models to complement physics-based simulation and empirical ground-motion modeling, opening a path toward fast, probabilistic, and scalable seismic hazard workflows for resilient infrastructure systems.