Identify best-suited computational approach for rapid, high-resolution 3D seismic wavefield prediction

Determine, via systematic benchmarking, which computational approach for seismic wavefield prediction—among reduced-order modeling based on interpolated proper orthogonal decomposition and machine-learning architectures such as neural operators, generative adversarial networks, and physics-informed neural networks—best achieves rapid time-to-solution and high-resolution accuracy for realistic regional three-dimensional velocity models that include lateral heterogeneity, topography, site effects, and viscoelastic attenuation.

Background

Three-dimensional physics-based simulations of seismic wave propagation are computationally expensive, making real-time ground motion prediction and uncertainty quantification challenging. A variety of machine-learning and interpolation approaches have been proposed to accelerate waveform modeling, including neural operators, GANs, and PINNs, yet direct analytical solutions are unavailable in heterogeneous 3D settings.

The text notes that no current method can perform regional-scale physics-based ground motion prediction while incorporating detailed structural heterogeneities and topographic effects at the required speeds. Given the importance of time-to-solution, it explicitly states uncertainty about which approach is best suited to produce rapid, high-resolution seismic wavefields in realistic regional domains.