A blog post by Geert Wenes, senior practice leader at Cray lifts the lid on PGS’ use of machine learning in full waveform inversion (FWI) for seismic imaging. PGS’ Abel supercomputer has been used to ‘find hidden and unexpected insights in complex data’ with minimal programming effort. The test was run on the 2004 SEG/BP velocity estimation benchmark model. Previous approaches take the initial velocity model and use ‘some sort of a least squares fit’ to seek convergence to a ‘true’ velocity picture, with ‘some level of success.’
The novel PGS/Cray approach uses ‘constrained minimization of a regularized and steered misfit function’ a.k.a. machine learning, to provide what is claimed to ‘dramatically improve’ the quality of the resulting model. The supercomputer ‘learns’ the velocities for a substantially clearer final image without the artefacts of conventional FWI.
For more on machine learning in seismic processing checkout the Sinbad consortium at the University of British Columbia. Sinbad (seismic imaging by next-generation basis function decomposition) is working to adapt recent developments from compressive sensing and machine learning to seismic imaging. Sinbad members include Chevron, Conoco, Schlumberger and … PGS.
© Oil IT Journal - all rights reserved.