Group B, Poster #224, Ground Motions

Reduced-order modeling for physics-based ground motion simulation

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

Poster Presentation

2022 SCEC Annual Meeting, Poster #224, SCEC Contribution #12224 VIEW PDF
With the increase of machine learning (ML) applications in seismology, new research is needed to develop hybrid frameworks that combine ML with physics-based modeling. While ML methods have shown promising results for extracting information from large geophysical datasets, their black-box quality raises concerns about physical inconsistencies, especially for seismic hazard and earthquake early warning (EEW) applications. In contrast to ML, physics-based modeling is firmly rooted in first-order principles, though it generally does not incorporate real observations. Moreover, physics-based modeling is too computationally expensive for real-time predictions, and may be unable to generate suffic...ient training data to fully train certain ML algorithms. The complementary strengths of ML and physics-based modeling motivates new research to develop hybrid models that combine the two.

To that end, here we explore the applicability of combining data-driven reduced order models (ROMs) with physics-based forward simulations of ground motion to predict peak ground velocity (PGV) maps. Specifically, we evaluate the interpolated proper orthogonal decomposition-based ROM by considering the southern California region and simulating synthetic earthquakes in which the focal mechanism and depth are taken as the model parameters. We explore the ROM accuracy when using radial basis functions, neural networks, k-nearest neighbors, and decision tree interpolators, and when either a 1D or 3D subsurface velocity model is employed. We validate this reduced-order modeling approach with recorded PGV data from the M4.5 2020 South El Monte earthquake. We find that the reduced-order models accurately predict the recorded PGV data, and we compare this prediction accuracy against NGA-West2 ground motion model predictions. We envision that these ROMs will be useful for real-time ShakeMap generation and EEW algorithms that could benefit from physics-based ground motion prediction.
SHOW MORE

Citation: Rekoske, J., Gabriel, A., & May, D. (2022, 09). Reduced-order modeling for physics-based ground motion simulation. Poster Presentation at 2022 SCEC Annual Meeting.
Relevant SCEC Themes:
Data-Intensive Computing
Ground Motion Simulation
Post-Earthquake Rapid Response