Low-rank representation of subsurface extended image volumes with power iterations

TitleLow-rank representation of subsurface extended image volumes with power iterations
Publication TypeSubmitted
Year of Publication2019
AuthorsMengmeng Yang, Marie Graff, Rajiv Kumar, Felix J. Herrmann
Keywordsfull extended image volumes, power iterations, private, randomized SVD, time domain
Abstract

Full subsurface extended image volumes (EIVs) contain abundant information such as subsurface image gathers used in imaging, interpretation of rock properties or velocity analysis, but are expensive in term of computational cost and storage requirement due to their large size. Yet, due to their redundant information, monochromatic EIVs exhibit a low-rank structure that allows us to get a good approximation at a computational cost proportional to the rank k , while conventional techniques require at least the number of sources. Such low-rank factorized EIVs are computed thanks to the randomized singular values decomposition (SVD) algorithm. However, recent developments on low-rank approximations of EIVs raise two major questions. First, monochromatic EIVs rely on time-harmonic wave equation solvers, which do not scale well to realistic 3D models. Second, the rank of the monochromatic EIVs increases with the frequency, which yields increasing computational costs and storage. Here, we propose a construction approach based on a time-domain finite-difference wave-equation solver with time stepping, which is combined to power iteration schemes in the randomized SVD algorithm to accelerate the decay of the singular values. We compare the performances of simultaneous iterations, block-Krylov iterations and the original randomized SVD on the computation of the EIV for a small section of the Marmousi model. Then, we show further results on the full Marmousi model to validate our approach.

Notes

Submitted to SEG on April 1, 2019

URLhttps://slim.gatech.edu/Publications/Private/Submitted/2019/yang2019SEGlrr/yang2019SEGlrr.html
Citation Keyyang2019SEGlrr