We checked into the grande finale of Agile Scientific’s machine learning hackathon hosted along with the EAGE chez Total. Around 60 heard from a dozen teams’ attempts to develop a demonstrator application in two days of coding. Some were plausible applications, like using supervised machine learning to identify mineralogical constituents of thin section microscope imagery. Others were far more ambitious. Seismic modeling is a labor and machine-intensive process. Team GANsters (generative adversarial networks) decided to cut-out the ‘science’ and use ML. The system was trained on 20k images pairs of models and their seismic representation. A trial on the EAGE’s Marmousi model was deemed to have produced ‘amazing’ results. While ‘this is black box mapping,’ the team plans to unpick the GAN results and see how they can inform processing, or at least, speed things up. More from Model2Seismic. One team set out to automate well correlations, combining ML with geology, using a set of images and interpretations. The process worked reasonably well on low dip conformable geologies, but an attempt to ‘ground truth’ a section with a salt dome produced an outcome that was ‘a bit different.’ The ANother team were more successful in developing an ML-based interpretation for mineralogy. Some 200GB of microscope imagery (polarized, unpolarized, fluorescence) were segmented into ‘superpixel’ clusters which were then tagged by a geologist.
A ‘Classy’ team set out to develop a seismic shot gather interpreter. This used a radon transform to pixelized the traces and develop a labeled training data set. An SVM classifier generated a test data set without labels which was run through the classified to identify events as ground roll, NMO, multiples, scoring ‘21/23!’
PickPickLog observed that stratigraphic interpretation and lithology identification from well logs is done by experts. It could be done by ML. Using an Alberta data set of gamma rays from 2,000 wells, lithology was determined using a logistic regression classifier. Comparison of expert and ML-derived lithologies was ‘quite good’ with a 75% match. The plan is now to replace expert supervision of the learning process with clustering.
In a similar vein, the LogFix team analyzed triple combo logs from the Athabasca basin using a Markov chain to calculate the probability of different lithology stacks e.g. a change from sand to silt. The approach could be used to create a resistivity log from nearest neighbor wells. Or to replace log reruns, perform data QC and patch washed-out sections.
The LogsOnTheRocks team set out to identify various E&P objects (lithological column, VSP, observers log…) in a large scanned image dataset from the UK’s OGA. A neural net was trained with tagged images. On the downside there are less tagged images of logs on the internet than there are of cats. But the approach works, ‘at 70% accuracy.’
All in all, we were more impressed by the enthusiasm of the participants than by the results of their efforts. Read the Agile report from the hackathon here.
© Oil IT Journal - all rights reserved.