Machine learning for oil & gas exploration

Author: Nick Brown
Posted: 27 Jul 2018 | 15:25

We are working on a machine-learning project with Rock Solid Images (RSI), a geoscience consulting firm that provides borehole characterisation with the goal of reducing exploration drilling risk for oil and gas companies.

RSI is one of the main players in the interpretation of seismic data with well log data and it has built its business on using advanced rock physics methods combined with sophisticated geologic models to deliver highly reliable predictions of where oil and gas might be found.

One of RSI’s key aspirations is to predict the geology of remote and difficult-to-reach locations from well explored areas of known similar geology. This is an open question and likely to raise considerable challenges. We won’t necessarily solve all of them in this project, but as our machine-learning models mature we will explore the portability of them to other areas. For instance, how accurately can the models that we have trained based on the geology of mid-Norway predict rock properties in the Barents Sea?

The oil and gas industry is awash with subsurface data and the investigation of a single well produces very many data fields, with values for each of these recorded at many depths down the well. This raw data then needs to be interpreted before it is useful but crucially interpretation is manually intensive and it takes over a week for an experienced petrophysicist to interpret each well. The resulting interpreted data is then fed into RSI’s rockAVO™ software, which enables its customers to both understand the rock physics of existing wells and also predict geology elsewhere in a region.

The central aim for RSI’s customers is to make informed decisions about where to drill and rockAVO™, along with the interpreted well data, is a key part of this. In this process the more wells you have, the more accurate a prediction can be made. But the manual interpretation time of each well fundamentally limits the number that can be used. It is currently common for RSI’s customers to use between 10 and 100 wells per region, whereas there is raw data available for thousands or even tens of thousands of wells in some locations... but with the current, manual interpretation, process these would take centuries to interpret!

Process optimisation

This project will focus on optimising the process of petrophysical interpretation by using machine learning. Pattern recognition underlies the action performed by the experienced petrophysicist, so a key question is whether one can leverage machine learning approaches to bring down the interpretation time from a week to a matter of a few hours or even minutes. To this end we will develop models that tackle the different steps in their petrophysical workflow.

It is really interesting not only to learn more about this industry but also see how it is becoming very interested in using machine learning techniques to address grand challenges. It is clear to me that, while the application of machine learning in this industry is still at an early stage, there is a real momentum behind obtaining more value from data and an understanding that, if they get it right, it could be a game changer.

This 12-month project, called Streamlined WorkflOws for Optimal Petrophysics (SWOOP), is funded by the Oil and Gas Innovation Centre (OGIC).


Nick Brown, EPCC