2018 Oil and gas machine learning symposium

Geophysical Insights’ (GI) Paradise use cases. Geophysics on Microsoft Azure. ResNet on the Seismic Atlas. IBM ‘corpus’ for cognitive petrology. SEAM-style ML consortium mooted.

The 2018 Oil and gas Machine Learning symposium, held in Houston, was primarily a vehicle for disseminating Geophysical Insights’ (GI) Paradise use cases but the meet extended beyond the GI brief to encompass ML from third parties including IBM, Microsoft along with presentations from Anadarko, Shell and Repsol although the latter three were not available for this write-up. GI was founded by Tom Smith back in 2010* and was pretty well first out of the blocks with commercial machine learning in seismic interpretation. In 2017 GI added technology from the Assisted Seismic Processing & Interpretation (AASPI) consortium at the University of Oklahoma to its portfolio.

Fabian Rada (Petroleum Oil and Gas Services*) gave an introductory run-through of the use of Paradise, involving selection with principle component analysis of attributes of interest and using neural net-based self organizing maps to relate this to seismic information that is below traditional spatial resolution. Paradise relates reservoir geobodies to a ‘neuron number’ obtained from the classifier. Particular neurons are related to significant facies. The resulting classification volume can be interpreted with conventional 3D interpretation tools and further calibration with logs turn SOM maps into net reservoir estimate.

* GI rep in Mexico.

GI’s Ivan Marroquin introduced Paradise ‘ThoughtFlow’, a new platform to provide a user-friendly GUI around its ML solutions. ThoughtFlow enables unsupervised ML, combining attributes to reveal natural clusters in attribute space and preforming geobody extraction from a seismic facies fingerprint. The technique also performs fault picking while SOM identifies the ‘most performant’ predictive model.

Dania Kodeih presented Microsoft’s Azure ML used across oil ands gas. Of the many applications presented we noted document library analysis with ‘coordinate-bound search’, OCR-based document classification and machine reading of text to visualization with Int’s Ivaap. Microsoft also presented its ersatz Paradise-like DNN ‘featurizer’ running ResNet on labelled image data in the Seismic Atlas*. The system has also been tested on the Dutch F3 block seismic cube.

* An industry-academia partnership led by Aberdeen Uni, Leeds and NERC. Last updated in 2014.

Geophysical Insights president and CEO Tom Smith gave an ML seminar, showing the new tools and ways of thinking taht are now available to geoscientists. The Paradise ML classifier identifies natural clusters in attribute space, providing noisy images of geologic features. A natural cluster is the image of a geobody produced by the selected attributes which has stacked with other geobody images in the same location in attribute space. Some 30 or so attributes are available for classification and analysis. Smith related the technique to seismic sequence stratigraphy which subdivides the seismic section into reflection packages. These are used to interpret environmental settings and lithofacies from seismic data. Simple geobodies are autopicked by machine learning in attribute space. Interpreted geobodies are constructed from these by hand editing and construction rules, ‘all seismologists are in the model building business’. Smith recommends that instead of (subjectively) discarding ‘bad’ models, it is preferable to measure the probability of successfully fitting the data. Interpreters can then present the model and discuss its probability of success.

Sumit Gupta showed how IBM’s Corpus conversion service used to build a ‘cognitive petrology’ from various bibliographic sources (Elsevier, AAPG, CC Reservoir ...). This leverages semantic extraction (with IBM’s SPSS SmartReader) from PDFs, such that the system ‘understands’ geological relationships. A complete ‘PowerAI’ ‘open source-based’ ML stack was developed. Really large workloads can use IBM’s GPU-based SnapML.

Bill Abriel (Orinda Geophysical) discussed ML in seismics and suggested that this would be a good subject for a cross-industry project, perhaps along the lines of the SEG’s SEAM consortia. He also announced an upcoming joint SEG, SPE, AAPG ‘Digital Transformation’ conference to be held in Austin in June 2019. More from the symposium home page.

For more on the founding of GI read our 2010 interview with Tom Smith.

Click here to comment on this article

Click here to view this article in context on a desktop

© Oil IT Journal - all rights reserved.