The Society of Exploration Geophysicists (SEG) annual meeting is one of the largest gathering of geophysicists in the world. The most important and relevant topics affecting geophysics are always discussed in many different forms. This year, by far the most discussed topic was machine learning. From the plenary session that featured a very well attended talk from Darryl Willis of Google, to the many panels, oral sessions, posters, and workshop talks addressed how machine learning is impacting the oil and gas industry from every angle. Google even hosted a hackathon that gave an overview of TensorFlow and allowed attendees to play with machine learning methods in python.

The oral sessions for machine learning were standing room only most of the time, and featured some great talks. The first ML session on Tuesday was heavily weighted towards fault detection using convolutional neural networks (CNNs) (6 out of 8 of the talks) and the others were about using CNNs for salt body detection. My favorite talk came from Xinming Wu from the BEG in Austin. His talk was called "Convolutional neural networks for fault interpretation in seismic images" and I highly encourage reading his abstract. His talk showed how he successfully used thousands of synthetic seismic images that he generated to train a CNN to identify faults in real 3D seismic data. He showed multiple real-data examples of his technique, and compared to the other talks, he has the most successful application of using CNNs for fault identification.

Tuesday afternoon featured a session on ML applications used for noise reduction in seismic data, but I instead attended the interpretation session, which kicked off with two facies classification workflows. The first was out of OU by David Lubo-Robles and Kurt Marfurt with the paper “Unsupervised seismic-facies classification using independent-component analysis.” They used independent component analysis to identify seismic facies from multiple attributes in an unsupervised fashion. Their technique differentiated from principal component analysis in that it produces independent components of the inputs, while PCA tends to mix components. In their real data example, I was particularly fascinated that their technique separated the different geologic signals and acquisition footprint signal into different components. They didn’t say this, but I thought they could have easily interpreted each volume and selectively stacked the components that had the smallest acquisition footprint to significantly improve their volume.

Wednesday morning contained a ML session that specifically address facies classification and reservoir properties. A very interesting talk was Laura Bandura from Chevron with the talk “Enhancing quantitative interpretation in the prestack domain with machine learning.” Her technique was aimed at nearly automating AVO/AVA amplitude analysis using fuzzy-c clustering. Fuzzy-c clustering is a soft-clustering technique that finds the degree of membership (effectively a likelihood) of a datapoint to be within a cluster. Applied to the prestack SEAM data, she yielded successful results extracting a gas-sand.

Capping off the annual meeting were one and a half days of workshops where operators, service companies, and academics can present and discuss bleeding edge topics without the burden or exposure of an extended abstract. Of course machine learning and data analytics topics were featured both days. The first day was a packed house. Dimitri Bevc from Chevron gave a very enlightening talk about how Chevron integrates machine learning into just about every facet of the company. He showed some great internal examples of automatic channel detection from seismic, document searching and classification, and use of Google’s AutoML as a subsurface search tool. Aria Abubakar from Schlumberger discussed their very ambitious goal to essentially automate subsurface exploration through a number of examples. Detlef Hohl from Shell heavily covered the topic of physics-driven data science and Shell’s work to use physics-based simulations to train and constrain machine learning models. He showed a fascinating example of predicting permeability measurements from flow tests using only thin section images and a hybrid physics-ML model. He also pointed out how time-series based ML methods are still poorly developed, yet time-series data is the largest set of oil and gas data. We saw an overview of the Kaggle-TGS salt identification challenge by Arvind Sharma and some of the initial results. Finally, Alan Cohen of the department of energy gave a plea for government and private sector collaboration on physics-based ML problems and talked heavily about exploration basin models.

Overall, the annual SEG meeting was filled with quality talks and a diverse offering of mediums for capturing the latest advancements in machine learning. With the rapid pace at which artificial intelligence is changing, I would not be surprised if next year’s SEG and AAPG annual meetings are overflowing with new machine learning applications.