Group A, Poster #173, Ground Motions

Instant physics-based ground motion time series using reduced-order modeling of 3D wave propagation simulations

John Rekoske, Alice-Agnes Gabriel, & Dave A. May
Poster Image: 

Poster Presentation

2023 SCEC Annual Meeting, Poster #173, SCEC Contribution #13127 VIEW PDF
Physics-based ground motion estimates obtained by numerical simulation of wave propagation are increasingly being considered for seismic hazard assessment. However, simulations are computationally expensive, especially when considering ensembles of earthquake sources, and require high-performance computing (HPC). Extending an approach for instantaneous generation of ground motion maps (Rekoske et al., 2023), we investigate using a reduced-order model (ROM) to obtain synthetic seismograms that could be used for on-demand ground motion simulation. We generate the data required to construct our ROM using time-dependent wavefield output from 2D and 3D wave propagation simulations using SPECFEM2D... and SeisSol, respectively, for velocity models with heterogeneous elastic media and varying earthquake source locations. After building the ROM using radial basis function interpolation combined with proper orthogonal decomposition, we examine the time-domain, frequency-domain, and cross-correlation errors for a ROM that models ≤1.0 Hz elastic wave propagation. Due to the high computational cost of the simulations, we explore multiple sampling strategies of the source locations to develop a high-accuracy ROM for a fixed computational budget. We compare uniform sampling, depth-aware, velocity-aware, and an adaptive sampling approach using a Voronoi neighborhood algorithm. We show the number of forward simulations required to reach a certain error tolerance for each sampling approach. We also quantify the change in error when we vary the highest resolved frequency and the lowest S-wave speed considered in the velocity model. Once constructed, the ROMs can be made easily accessible to the community to enable instant computation of ground motion time series. For the Los Angeles region where high resolution community velocity models are available, we present a ROM based on the CVM-H velocity model that accurately matches simulated seismograms. We further show the value of this approach by comparing the instant ROM predictions against real ground motion data recorded in this region.