Application of Conditional Dynamic Variational Autoencoder for Simulating Ground Motions in the Geysers Geothermal Field
Description:
Accurate simulation of ground motions caused by induced seismicity is imperative for engineering operations to mitigate large events and for advancing our understanding of source physics through correlating with stress, strain and fluid flow monitoring. We focus on the Geysers geothermal field in northern California, whose induced seismicity activities have been linked to geothermal energy production. Physics-based simulations require a large computational cost and are susceptible to uncertainties in subsurface velocity models. Deep generative models are attractive alternatives that directly utilize observations. Based on the conditional dynamic variational autoencoder for ground motion simulation (CD-VAE-GMG) model developed for small natural earthquakes in the San Francisco Bay Area, we aim to generate region-specific waveforms associated with induced seismic events. We introduce a novel dynamic Variational Autoencoder design to ensure the statistical features of the waveform envelopes and spectra shapes to be consistent with observations in the time-frequency and time domains simultaneously.
Our model is trained to approximate the probability distribution of a vast dataset of ground motions in the Geysers geothermal field. It simulates three-component waveforms conditioned on earthquake magnitudes, epicentral distances, and earthquake depths. The training dataset comprises waveforms from over 30,000 earthquakes with magnitudes ranging from 0 to 4, recorded at 90 stations. Preliminary results demonstrate the effectiveness of the model. Generated seismic waveforms match the observed waveforms in both time and frequency domains, but some discrepancies remain in peak values and spectral shapes. We are improving the performance through data expansion, model tuning, and comprehensive validation. This study showcases the potential of machine learning techniques in simulating ground motion waveforms caused by the induced seismicity and enhancing their prediction. We further envision that our approach will become valuable for improving earthquake hazard assessment and gaining insights into seismic source physics.
Session: How Well Can We Predict Broadband Site-Specific Ground Motion and Its Spatial Variability So Far? [Poster Session]
Type: Poster
Date: 5/1/2024
Presentation Time: 08:00 AM (local time)
Presenting Author: Zhengfa
Student Presenter: No
Invited Presentation:
Authors
Zhengfa Bi Presenting Author Corresponding Author zfbi@lbl.gov Lawrence Berkeley National Laboratory |
Pu Ren ren.pu@northeastern.edu Lawrence Berkeley National Laboratory |
Rie Nakata rnakata@lbl.gov Lawrence Berkeley National Laboratory |
Nori Nakata nnakata@lbl.gov Lawrence Berkeley National Laboratory |
|
|
|
|
|
Application of Conditional Dynamic Variational Autoencoder for Simulating Ground Motions in the Geysers Geothermal Field
Session
How Well Can We Predict Broadband Site-Specific Ground Motion and Its Spatial Variability So Far?