Acceleration strategies for elastic full waveform inversion workflows in 2D and 3D
Research output: Contribution to journal › Article › Research › peer-review
Standard
In: Computational Geosciences, Vol. 21.2017, No. February, 22.10.2016, p. 31-45.
Research output: Contribution to journal › Article › Research › peer-review
Harvard
APA
Vancouver
Author
Bibtex - Download
}
RIS (suitable for import to EndNote) - Download
TY - JOUR
T1 - Acceleration strategies for elastic full waveform inversion workflows in 2D and 3D
AU - Kormann, Jean, Antoine
AU - Rodriguez, Juan Esteban
AU - Ferrer, Miguel
AU - Farres, Albert
AU - Gutierrez, Natalia
AU - de la Puente, Josep
AU - Hanzich, Mauricio
AU - Cela, Jose Maria
PY - 2016/10/22
Y1 - 2016/10/22
N2 - Full waveform inversion (FWI) is one of the mostchallenging procedures to obtain quantitative informationof the subsurface. For elastic inversions, when both com-pressional and shear velocities have to be inverted, thealgorithmic issue becomes also a computational challengedue to the high cost related to modelling elastic rather thanacoustic waves. This shortcoming has been moderately mit-igated by using high-performance computing to accelerate3D elastic FWI kernels. Nevertheless, there is room in theFWI workflows for obtaining large speedups at the cost ofproper grid pre-processing and data decimation techniques.In the present work, we show how by making full use offrequency-adapted grids, composite shot lists and a noveldynamic offset control strategy, we can reduce by severalorders of magnitude the compute time while improving theconvergence of the method in the studied cases, regardlessof the forward and adjoint compute kernels used.
AB - Full waveform inversion (FWI) is one of the mostchallenging procedures to obtain quantitative informationof the subsurface. For elastic inversions, when both com-pressional and shear velocities have to be inverted, thealgorithmic issue becomes also a computational challengedue to the high cost related to modelling elastic rather thanacoustic waves. This shortcoming has been moderately mit-igated by using high-performance computing to accelerate3D elastic FWI kernels. Nevertheless, there is room in theFWI workflows for obtaining large speedups at the cost ofproper grid pre-processing and data decimation techniques.In the present work, we show how by making full use offrequency-adapted grids, composite shot lists and a noveldynamic offset control strategy, we can reduce by severalorders of magnitude the compute time while improving theconvergence of the method in the studied cases, regardlessof the forward and adjoint compute kernels used.
U2 - 10.1007/s10596-016-9593-0
DO - 10.1007/s10596-016-9593-0
M3 - Article
VL - 21.2017
SP - 31
EP - 45
JO - Computational Geosciences
JF - Computational Geosciences
IS - February
ER -