Vertical continuation of seismic waveforms through the shallow structure with neural operators

Shuye Huang, & Yehuda Ben-Zion

Published September 8, 2024, SCEC Contribution #13661, 2024 SCEC Annual Meeting Poster #192

Seismic waveforms at the surface are strongly affected by nonlinear wave propagation in the geotechnical top layer, topography, various other site effects, and local noise sources, which limit the ability to both simulate ground motion and interpret seismograms recorded by surface sensors. As a denoising strategy, downward continuation of surface waveforms is sometimes applied. Similarly, upward continuation of waveforms recorded by borehole sensors may be used to simulate more realistic ground motion and improve seismic hazard estimates. In this study, we utilize data recorded in a vertical borehole array and a neural operator approach to perform upward/downward continuation of seismic waveform through the top-most structure of the Earth. Specifically, we use 3-channel quality-checked accelerograms of over 5000 earthquakes recorded at the Garner Valley Downhole Array (GVDA) in southern California, paired between instruments at the surface and at 50m in depth. The GVDA site has additional boreholes at greater and shallower depths that may be utilized later. We keep a signal content of up to 20 Hz to satisfy the interests of both earthquake engineering in surface ground motion and improved seismicity monitoring. We train two U-shaped neural operators to map the wavefield upward (50 m depth to surface) and downward (surface to 50 m depth), aiming to predict seismograms of high coherency with observed ones and accurate ground motion (peak ground acceleration, PGA, and peak ground velocity, PGV). Our results show that the trained upward continuation model predicts up to 20 Hz accelerograms of good coherency compared to actual recording at the surface (median coherency of 0.66, 0.81, 0.78 for Z, E, N channel). R-square statistics of PGA and PGV prediction versus ground truth reach 0.86 and 0.70. Our model greatly over-performs the baseline algorithm, where a single scaling factor (median peak ground motion ratio of the accelerogram pairs) is applied. Downward continuation results show a similar capacity. The continuing work will develop a similar ability to map seismic wavefields through the complex shallow structures at additional depths using the GVDA data. With future modification towards generalization, this approach can be applied to other sites that have both surface and borehole data for model training, and possibly including transfer approaches for additional locations.

Key Words
Waveform continuation, Ground Motion, Machine Learning

Citation
Huang, S., & Ben-Zion, Y. (2024, 09). Vertical continuation of seismic waveforms through the shallow structure with neural operators. Poster Presentation at 2024 SCEC Annual Meeting.

Related Projects & Working Groups
Ground Motions (GM)