CD-VAE-GMG: Conditional Dynamic Variational Autoencoder for Earthquake Ground Motion Generation
Description:
Simulating earthquake ground motions is crucial for assessing seismic hazards and ensuring the safety and resilience of critical infrastructure. However, obtaining ground motion data across a wide geographic area with high spatial sampling is challenging and resource-intensive, especially for regions with infrequent seismic activity. This work proposes a novel Conditional Dynamic Variational Autoencoder (CD-VAE) framework to predict new seismograms from unobserved earthquakes at new sources and site locations. Specifically, we leverage the Short-Time Fourier Transform (STFT) to extract the time-frequency information and utilize the probabilistic autoencoder to learn the latent distributions based on the amplitude spectrogram. Both prior and posterior distributions are constructed with variational sequential models (i.e., recurrent neural networks) to facilitate capturing temporal dynamics. Furthermore, we apply the phase retrieval method to estimate the phase information from the reconstructed amplitude spectrogram and use inverse STFT to recover the corresponding waveforms. Moreover, our model is conditioned on specific physical parameters, such as earthquake magnitudes and the coordinates of sources and stations. This method is easy and stable for optimization compared with previous Generative Adversarial Network (GAN) models. We use a rich set of small-magnitude earthquake records from the San Francisco Bay Area. From 1.4 million horizontal component traces recorded at 266 sensors for 737 earthquakes, we select 5194 traces based on S/N ratios. The generated ground motions follow the PGV-magnitude and PGV-distance relationships of the observations. The Fourier amplitude spectra values show good agreement between observed and generated data over an entire range of frequencies (2-15 Hz) used in this study. We further checked the kinematics of the waveforms by picking P and S wave arrivals using PhaseNet. S wave arrivals are well reproduced while generated P arrivals tend to be earlier than the observation. These evaluations demonstrate the effectiveness of the proposed methods.
Session: How Well Can We Predict Broadband Site-Specific Ground Motion and Its Spatial Variability So Far? - I
Type: Oral
Date: 5/1/2024
Presentation Time: 08:45 AM (local time)
Presenting Author: Pu
Student Presenter: No
Invited Presentation:
Authors
Pu Ren Presenting Author Corresponding Author pren@lbl.gov Lawrence Berkeley National Laboratory |
Ilan Naiman naimani@post.bgu.ac.il Ben-Gurion University of the Negev |
Maxime Lacour maxlacour@berkeley.edu University of California, Berkeley |
Rie Nakata rnakata@lbl.gov Lawrence Berkeley National Lab |
Nori Nakata nnakata@lbl.gov Lawrence Berkeley National Lab |
Zhengfa Bi zfbi@lbl.gov Lawrence Berkeley National Lab |
Osman Malik oamalik@lbl.gov Lawrence Berkeley National Lab |
Dmitriy Morozov dmorozov@lbl.gov Lawrence Berkeley National Lab |
Omri Azencot aizencot@gmail.com Ben-Gurion University of the Negev |
N. B Erichson erichson@lbl.gov Lawrence Berkeley National Lab, Berkeley, California, United States |
Michael W Mahoney mmahoney@stat.berkeley.edu Lawrence Berkeley National Lab, Berkeley, California, United States |
CD-VAE-GMG: Conditional Dynamic Variational Autoencoder for Earthquake Ground Motion Generation
Category
How Well Can We Predict Broadband Site-Specific Ground Motion and Its Spatial Variability So Far?