In Episode 94 of the SEG’s Seismic Soundoff podcast, Andrew Geary (51 Features) interviewed Mehdi Aharchaou (ExxonMobil) on ‘The case to adjust quicker to machine learning for geophysics’. Aharchaou reported that outside of our industry, use of supervised ML is now ‘more reasonable’. It does not apply to all problems and should not be considered a ‘silver bullet’. In geophysics, popular applications of ML include data QC, currently a labor intensive area that could be (more) automated.
Siamese neural networks are useful at the common seismic task of comparing two objects. ‘Similarity is at the heart of what we do’, comparing two or more results from a processing sequence and providing an objective assessment of improvement. ‘Edge-aware’ filtering is an objective, streamlining and simplifying the imaging workflow allowing for comparison of observed and simulated results or of near and far trace AVO.
Deep learning can be used to extend seismic bandwidth, learning from an ocean bottom data set and reconstructing the low-frequency content of conventional recording. Likewise, high-frequency data from shallow acquisition can be used to train regular data and make for synthetic broadband. AI trained on Kirchhoff and reverse time images can be used to avoid prohibitively expensive reverse time migration.
But progress to date has been disappointing. We need better adoption and more impactful use. All this work has been done in the last five years. More time is needed to catch up on developments in ML/AI, especially computer vision. ML should not be seen as a hammer looking for nails. Not all problems are candidates and some geophysical processes already automated and have no need for ML. There is also the challenge of moving from proof of concept to at-scale. We are currently just scratching the surface.
More in The Leading Edge Special section: Machine learning and AI (October 2020).
© Oil IT Journal - all rights reserved.