Efficient Lossy Compression of Simulated 4d Seismic Wavefields
Description:
State-of-the-art inversion and imaging technologies, such as Full Waveform Inversion (FWI) and Reverse Time Migration (RTM) require the ready availability of the forward simulated wavefields or Green function wavefields during gradient calculation. As storing all the wavefield of all time steps, which a 4D volume, in computer memory is impractical for 3D problems, the most intuitive strategy would be to store the 4D simulated wavefield data on disk. However, this will introduce an important bottleneck because of the significant I/O overhead and decrease in performance. To overcome this bottleneck, we incorporate the efficient lossy compression techniques based on Tucker tensor decomposition into waveform inversion workflow by storing the compressed 4D simulated wavefield onto the disk instead of reconstruction on-the-fly. For Tucker tensor decomposition, the simulated seismic wavefield sequence with three spatial dimensions and a time dimension can be naturally represented as a 4-way tensor, and then detects the global structure of the tensor and removes redundant information, resulting in a low-rank representation of the data. We compared the Tucker-based compression with other common compression schemes on compression ratio, accuracy, and compression/decompression time, and the results show that Tucker-based compression is a good choice for real-time online compression of 4D simulated wavefield data since it can efficiently compress the wavefield by more than 100x with less than 40% time overhead in our tests. To verify the effectiveness of the method in FWI workflow, we use the 3D collocated-grid finite-difference algorithm on curvilinear grids to simulate seismic wavefield in the presence of surface topography (CGFD3D). The efficient lossy compression is applied in the ambient noise waveform inversion of seismic data in the Longmenshan area, which greatly speeds up the computation compared with directly outputting the forward wavefield and strain Green tensor without compression.
Session: Numerical Modeling in Seismology: Developments and Applications - II
Type: Oral
Date: 5/1/2024
Presentation Time: 11:00 AM (local time)
Presenting Author: Wei
Student Presenter: No
Invited Presentation:
Authors
Wei Zhang Presenting Author Corresponding Author zhangwei@sustech.edu.cn Southern University of Science and Technology |
Wu Wang 11849526@mail.sustech.edu.cn Southern University of Science and Technology |
Yongjie Tang tangyj@sustech.edu.cn Southern University of Science and Technology |
Tao Lei leit@sustech.edu.cn Southern University of Science and Technology |
|
|
|
|
|
Efficient Lossy Compression of Simulated 4d Seismic Wavefields
Category
Numerical Modeling in Seismology: Developments and Applications