Constructing misfit function for full waveform inversion based on sliced Wasserstein distance
KAUST DepartmentComputer, Electrical and Mathematical Sciences and Engineering (CEMSE) Division
Earth Science and Engineering Program
Extreme Computing Research Center
Physical Science and Engineering (PSE) Division
Permanent link to this recordhttp://hdl.handle.net/10754/663456
MetadataShow full item record
AbstractIt has been proven that full waveform inversion (FWI) based on optimal transport (OT) is robust against noise and contains less local minima. The accurate solution to OT problem on the real line (1D) is analytical with the selected distance function which can be effectively and non-iteratively solved by numerical interpolation. Currently finding an accurate optimal transport plan in higher dimension is known to be expensive even if it exists. The seismic data are naturally sampled in higher dimension and typically considered to be large scale. Given shot gathers from 2D survey, when calculating the misfit along one direction only, the information in other directions would be under-determined. In this paper, we introduce sliced Wasserstein distance to calculate the transport plan function in the lower dimensions of transformed domain such that other directions are considered implicitly. The extra cost caused by radon transform as compared to its counterpart based on $L_2$ norm is negligible. Numerical examples indicate that misfit based on sliced Wasserstein distance can improve the efficiency of FWI and shows the capacity for salt velocity building.
CitationChen, F., & Peter, D. (2018). Constructing Misfit Function for Full Waveform Inversion Based on Sliced Wasserstein Distance. 80th EAGE Conference and Exhibition 2018. doi:10.3997/2214-4609.201801030
SponsorsThe partial work of this paper was done when the first author was visiting the University of Texas at Austin. It is greatly appreciated that KAUST financially supported this visiting. We express our thanks to Sergey Fomel for helpful discussion about radon transform. We also thank Zhiguang Xue for useful discussion about the second numerical example. Special thanks to KAUST for generous HPC support which has facilitated the experiments.
Conference/Event name80th EAGE Conference and Exhibition 2018: Opportunities Presented by the Energy Transition