How Deep is Too Deep? Reflections and Observations of how AI and Machine Learning are Perceived from AAPG ACE 2018 Annual Convention

May 29, 2018 Scotty Salamoff, Chief Geophysicist at GeoTerra Technologies, LLC

Having just returned from a rather somber atmosphere this year at the annual AAPG conference, I have come to a singular, strong conclusion: AI is already well on its way to becoming a buzzword, and the subsurface community is struggling to find out where it belongs in their world.

Traditionally a get-together for rock lovers worldwide, the AAPG ACE 2018 annual meeting resembled more of a conference on data science, geophysics, and artificial intelligence. Of course, this probably left the rock-lovers feeling isolated - were they at AGU or SEG, or were they at the AAPG conference?

This brings me to the title of this article - "How Deep is Too Deep"? It was disheartening to see machine learning and AI being used to "reinvent the wheel" to do things we already know how to do, and most demonstrations arrived at the same conclusion: the use of neural networks in place of certain - not all - traditional seismic exploration methods (such as fault or basement mapping) resulted in a higher degree of error than mathematical or attribute workflows that are currently in commonplace usage in the industry. This led to a realization that the application of Machine Learning and Neural Net Technology has started an "arms race" of sorts, but an unfocused one that could lead the industry down the wrong path.

Below is a short list of 10 observations made at AAPG that made me question the correct application of deep neural networks to the G&G world, and whether or not we're trying to go too far, too fast, and too deep when it comes to the integration of Machine Learning and subsurface Energy Exploration and production. If you were in attendance at AAPG this year too, please share your thoughts about your experience or observations. It's important to open up a much-needed dialogue about Machine Learning and how it relates to subsurface Energy activities now, before "AI" becomes the next "DHI" buzzword.

Ten Observations on AI and Machine Learning made at AAPG ACE 2018:

1) For the most part, the industry is still trying to figure out where AI fits into the long-term goals of subsurface exploration and production.

2) There seems to be a fundamental misunderstanding among some skeptics that the use of AI technology in E&P activities is a magic button to find resources and is destined to automate G&G jobs making our positions obsolete.

Pictures ? Respective Copyright Holders, modified by author.

3) Unfortunately, most applications of AI technology presented fell into one of two categories: the first being "let's use it to automate something we already know how to do very well"; and the second being "well, let's fill up the proverbial bucket, throw it at the wall, and see what sticks.... 1,000 times".

4) Haphazard promises and dubious claims. Computers are calculators and AI/Neural Networks are code. Very little information was presented on how to interpret the output of a Neural Network, but lots of promises were made about the value of the output.

5) There is a limit to how deep we can dig using this technology in our industry before it begins to generate diminishing returns. This ties back with the importance of knowing exactly what information is being read by the network, how to interpret the training model, and whether or not all information is relevant to the desired objective.

6) There seemed to be some misunderstanding regarding the difference between shallow neural networks, deep neural networks, and data mining when it comes to applications of Machine Learning technology in the O&G Industry.

7) AI and Machine Learning seemed to be an inappropriate focus for a Petroleum Geologists meeting, and it showed on the faces of many attendees.

8) There is a “Space Race” underway between competing service and startup companies to be the first at to the door to provide the best Machine Learning product, but without a clear indication of where or how to get there it seems many companies are falling back onto the “fill the bucket and throw it at the wall” tactic.

9) Attributes have exploded in their number, use, and lack of understanding. Where just 10 years ago a small set of seismic attributes were considered during exploration activities, the integration of more powerful computing with exceedingly complex developments in the fields of physics and mathematics mean attributes can be generated that may not mean anything. Caution is advised when using these sorts of attributes in Machine Learning design.

10) This may be controversial, but the conversation must be had: is Machine Learning, Data Science, and associated technologies a new category of subsurface G&G professionals, or is the technology destined to be integrated with existing Geologist and Geophysicist roles?

Despite the observations/questions posed above AAPG ACE 2018 was a great time as usual...it was fantastic to see old friends and meet new ones. The talks as well as the posters were phenomenal and it was encouraging to see all the new technological developments and academic projects – to all the presenters and speakers, a kudos to you and thank you all for demonstrating the usefulness of AI and associated technologies. You are all truly forward thinkers!