Reconstructing Seismograms via Self-Supervised Learning: Methodology and Applications
Description:
Seismograms are studied to decipher earthquake mechanisms, resolve Earth’s structures, and assess seismic hazards. However, fully exploiting them remains challenging for their imbalanced and contaminated spectral contents. Typical earthquakes of moderate size (M2-4) are usually deficient in clean low-frequency signal (< 1 Hz) , but have sufficient signal at high-frequency (> 1 Hz). Conversely, some earthquake signals may be depleted in high-frequency information, such those recorded in the past at low sampling rates or those looking at attenuated signals. Additionally, current numerical methods and knowledge of the Earth’s structure generally limit accurate simulations of seismograms to low frequencies. These limitations prevent seismologists from discovering detailed earthquake source physics and structural information. Therefore, decontaminating low frequencies from seismic noise and recovering high frequencies in seismograms are critical for better utilizing seismic signals. Here, we use a deep neural network to investigate the relationship between low- and high-frequency seismograms. After self-supervised learning of a small volume of high-quality broadband seismograms, our deep-learning model can reconstruct high-quality relative amplitudes and phases of waveforms with high similarity to the true ones for either low-frequency or high-frequency seismograms. In particular, our model can reconstruct high-quality low-frequency signals from high-frequency ones; it can also reconstruct high-quality high-frequency signals from low-frequency ones. Our seismogram reconstruction method can improve the utility of seismic data and hence allow for resolving the source mechanism of noisy earthquakes, enhancing structural imaging, and evaluating earthquake ground motions.
Session: Opportunities and Challenges for Machine Learning Applications in Seismology [Poster]
Type: Poster
Date: 4/19/2023
Presentation Time: 08:00 AM (local time)
Presenting Author: Congcong Yuan
Student Presenter: Yes
Invited Presentation:
Authors
Congcong Yuan Presenting Author Corresponding Author cyuan@g.harvard.edu Harvard University |
Youzuo Lin ylin@lanl.gov Los Alamos National Laboratory |
Marine Denolle mdenolle@uw.edu University of Washington |
John Shaw shaw@eps.harvard.edu Harvard University |
|
|
|
|
|
Reconstructing Seismograms via Self-Supervised Learning: Methodology and Applications
Category
Opportunities and Challenges for Machine Learning Applications in Seismology