Physics versus machine learning models

AAPG/SEG/SPE Energy in Data webinar hears from Hess on data-driven models in shale exploration. Corva on ROP drilling prediction. Schlumberger – use both ML and physics! Xecta don’t use ML on small data! Data-driven reserves reporting for anyone?

Energy in Data, a joint venture between the SEG, AAPG and SPE, aims to ‘lead the digital journey of the energy sector’ and holds free monthly webinars and other events. We attended a recent webinar on ‘Physics-based vs data-driven models’, subtitled ‘Engineering in a virtual sub-surface’.

Invited by compere Siddharth Misra (Texas A&M) to ‘give us your best shot’ at outlining the data-driven/physical model conundrum, Sebastian Matringe (Hess) cited the ‘huge reliance’ on data-driven methods in shale exploration. Classical reservoir models are unadapted or too slow. So, ‘just take well data and use machine learning to decipher the relationships’. Most shale operators use data-driven today. Kriti Singh (Corva) cited successes in optimizing and predicting drilling rate of penetration. Modeling ROP has been studied since the 1930s using physics-based models of weight on bit and other parameters. These later involved statistical models and most recently, ML. Now there are tools to use real-time data. But it should be noted that these are not data-driven alone. Most are hybrid models, you need to understand the physics. There is a balance between how hard it is to acquire all the data needed for a physical model and how expensive a mistake would be. For space exploration, mistakes are expensive so physics is used*. Ravinath Kausik (Schlumberger) observed that physics should be the ‘default that we trust’, data science comes in after. ‘They can aid one another’.

* Wells are expensive too!

The discussion turned to the reliability of data-driven models. While conventional models are studied and validated by experts, this is not usually done in oil and gas for data-driven models. These need to be made more robust and make them explainable. ‘You don’t want black boxes’. In unconventional exploration, companies are finding it hard to take the results from core acreage to outlying areas. ‘There was some hype earlier on…’.‘Neural nets are not for all circumstances. It might be better to use multi-variate regression that people can understand’. ‘Data-driven is influencing oil and gas hiring’. ‘Data science is the sexiest job of the 21st Century’.

On the subject of convergence of DD and physical models, Kausik suggested using both on the same problem, data science to segment data and then and then apply physics. Data science has a long way to go for widespread application. We still don’t appreciate the uncertainty in our models, when will they fail? If you don’t have big data don’t address the problem with a big neural net, use a simple model! ‘We don’t have as much data as Google’. Satish Sankaran (Xecta Digital Labs) observed that the industry has changed a lot over the years. Oils more or less abandoned R&D some 20 years ago. Today it is coming back. There are also many startups working in this space.

Are data-driven results appropriate/accepted for reserves booking? Matringe sees data-driven as an extension of the accepted SPEE practice of using type curves. These are ‘derived from data in order to justify proved undeveloped results’. However, ‘it’s not a simple answer, it depends on the application and there are a lot of details that have to be thought about when using these practices for something as important as reserves’.

There was pushback on the risk of ‘turning engineers into data scientists’, ‘we still need to honor physics and our engineering degrees’. Also on the real novelty of ‘data-driven’, which are ‘the same tools we've used in the past, just with new names and better hardware’. Not according to Matringe, ‘neural nets and regression look similar but there are some significant differences, today’s NNs are a different class of methods’.

The Q&A threw up some interesting ideas, for more on constraining ML with physics, read Raissi et al. and their Github.

The Energy in Data event is well run with a slick interface that allows for considerable audience interaction. There was only one regret, the posted question ‘we talk about success stories of data-driven methods, where are the outstanding failure stories that we can learn from?’ was left unanswered. Maybe a good topic for a future webinar?
Watch the webinar recording.

Click here to comment on this article

Click here to view this article in context on a desktop

© Oil IT Journal - all rights reserved.