You are Home   »   News   »   View Article

How do we use machine learning on oil and gas subsurface and production data?

Friday, November 22, 2019

We asked Duane Dopkin, Executive Vice President of Geosciences at Emerson E&P Software, about how he sees machine learning's role in oil and gas production and subsurface data - and where it is going.

Machine learning has been used in oil and gas subsurface projects since the mid-1980s, when algorithms like 'back propagation neural networks' were implemented to automate tedious tasks says Duane Dopkin, Executive Vice President Geosciences at Emerson E&P Software, (formerly Paradigm).

Today there is growing interest in machine learning technology, in part stemming from activity around the 'digital transformation' and the promotion of machine learning by big tech companies like Google, Amazon, Microsoft and IBM. Most of the latest wave of interest is built around more sophisticated machine learning engines which these companies are offering, he says.

There are a number of thought leaders in the oil and gas industry who have invested heavily in the promotion and evaluation of these new machine learning engines, he says.

But there is still a gap between the role and goals of machine learning as envisioned by the thought leaders and the practitioners - geoscientists and engineers who have to extract insights from subsurface and surface digital data, and have confidence in the results. 'It is the role of thought leaders to push this technology out. It is the role of practitioners to put the brakes on - proceed with caution, as it takes a while to solve surface and subsurface problems with emerging technologies.

'It is an evolution, it doesn't happen overnight. The gap will be closed with time and continued refinements, not only to machine learning methods, but to preparation of the input data and careful 'crafting' of the training sets - prerequisites for a successful outcome from machine learning applications.

Many successful applications of machine learning operate on consistent and highly sampled data, like consumer data. Oil and gas digital data is more complex and the problems we are attempting to solve with machine learning are also more complex. So different methods need to be applied.

Defining machine learning

Mr Dopkin says that machine learning uses algorithms that are used to solve nonlinear problems, or to put it more simply, to help us understand multi-dimensional data relationships that are difficult to ascertain using deterministic methods.

This relationship may depend on many other factors, including factors you don't know, such as the relationship between the hours you spend teaching a child math and their scores on an exam.

There are some deterministic algorithms that have been used extensively in the oil and gas industry which are very reliable and very predictable; therefore their limitations and strengths are well understood. These methods set a performance bar for machine learning methods.

'Some of the tasks we applied machine learning to in the 1980s and 1990s are pretty mature now,' he says. 'They are institutionalised in our workflows. Geoscientists are comfortable with those methods.'

To use these algorithms, you first select which algorithm is appropriate together with your understanding of the physics of the subsurface. You prepare your input data, train the machine learning method, and then turn it loose on the problem.

Today, many learning algorithms are available as open source, making it much easier for geoscientists to try them out without having to invest heavily in them upfront. These software trials also allow geoscientists to assess the effort and cost in integrating (embedding) them into their proprietary solutions.

Deep learning

More recent developments use 'deep learning', neural networks, that look for patterns in large data sets. The deep learning engines contain multiple layers, multiple nodes, and accept multiple inputs to carry out more advanced image recognition problems or transform one data type to another.

'These deep learning applications are still relatively low on the maturity curve, he says. 'But they are gaining a lot of ground quickly,' particularly for classifying data and image recognition. Deep learning algorithms show a great deal of progress in their application to surface and subsurface oil and gas digital data.

Machine learning lessons

Over the past 30 years in which Emerson E&P Software (Paradigm and Roxar) has been working on machine learning technology, it has learned a great deal.

It has learned that data preparation is critical, if you want machine learning to make useful predictions, data transformations, or data integrations.

It has also learned about the importance of selecting the right machine learning algorithm for the problem. The company has over 15 different machine learning 'engines' or algorithms that can be applied to solve different surface and subsurface challenges.

It can require a great deal of experimentation, with a practitioner exposing different machine learning methods to different input datasets.

It advocates the use of 'physics-based models', where the model reflects in some way the physics of how the parameters actually relate.

All of these algorithms 'absolutely' need domain expertise to use, he says. 'You have to understand what you are doing.'

Surface and subsurface data

Machine learning is more adaptable to surface recorded (engineering) data, such as production data, than to subsurface data, Mr Dopkin says. 'Engineering data carries less complexity than subsurface data.'

For example, machine learning can be used to forecast future production based on observed patterns in historical data. It can also be used to predict equipment behaviour. The algorithms can be run in real time (during operations), enabling them to support real-time decisions, such as when to do workovers, or strategies for optimising production.

'Machine learning is very applicable to that kind of data, looking for trends, forecasting futures and predicting failures,' he says.

But in the subsurface domain, machine learning 'is not as applicable right away,' he says. There are many more complexities, with diverse data types, data resolutions, and data sizes'.

Working out how to apply machine learning to these data sets, and understanding the limitations, takes time, he says. And the large data sets create extra challenges. 'We can be talking about applying ML to tens or hundreds of terabytes of data.'

In the subsurface, machine learning is typically applied to different sets of problems, including automating tedious activities such as picking first breaks, velocities, or seismic horizons as part of the seismic interpretation process. It can also be used to classify rock types or seismic facies, or to integrate seismic data with well logs.

Deep learning algorithms can now be applied to pre-stack seismic data to search for specific geologic features, such as faults, fractures, reefs, edges, channels, or karst features. 'Machine learning holds a lot of promise here,' he says.

Data integration

Machine learning can be used to help integrate subsurface data types together, a tricky task to do manually when the data samples different regions of the subsurface and at different resolution levels - such as cores, well logs and seismic.

Some data is sampled heavily in 3D space, like seismic data - and other data is densely sampled in a vertical domain, such as well logs. In addition, the different data types see the rock itself a bit differently.

There are many disciplines in the industry working with different types of data - for example, geophysicists with seismic data, petrophysicists with well logs, geologists with geologic models, engineers with pressure and saturation data. But it all needs to come together to enable reliable decisions about how to make safe and economic drilling decisions. In that sense, machine learning can be considered a collaboration tool, Mr Dopkin says.

Data managers

Data managers are still necessary in the world of machine learning - keeping track of where data is and what it shows.

Machine learning actually 'creates a new set of data management problems,' he says - storing the data about the machine learning training, so that other people can access it and apply it to their data.

There also need to be standards for how machine learning data is stored.

And since machine learning might be run on multiple data types, including seismic data, core data, well log data, and maybe engineering data, it is important to keep track of the sources of the data.

'Currently, machine learning does not simplify the data management problem; rather, it is forcing another set of data management standards and controls on digital data and applications,' he says. However, machine learning holds tremendous potential to help solve many other types of data management problems.

Associated Companies
» Emerson
comments powered by Disqus


To attend our free events, receive our newsletter, and receive the free colour Digital Energy Journal.


Clustering Considerations in the Machine Learning Workflow – Examples with Exploration Data
Philip Lesslar
from Precision DM


Latest Edition Jul-Aug 2019
Aug 2019

Download latest and back issues


Learn more about supporting Digital Energy Journal