You are Home   »   News   »   View Article

How to use machine learning in exploration

Friday, November 30, 2018

Machine learning can be used in exploration to help identify potential hydrocarbons in seismic data, map out geobodies and pick facies and faults, said Rocky Roden with Geophysical Insights.

Machine learning can be used in exploration to help identify potential hydrocarbons in seismic data, map out geobodies and pick facies and faults, said Rocky Roden President and Chief Geophysicist at Rocky Ridge Resources, also consulting to Geophysical Insights.

Mr Roden said he had been using machine learning himself since experimenting with early facies classification techniques that used neural networks in the 1990s, and 'it got me hooked'.

Definitions

The first challenge is defining machine learning, with many misconceptions around, he said.

Machine learning is software that learns from the data, typically identifying patterns in large volumes of data. It can analyse large amounts of data simultaneously, identity the relationships between different types of data (a task which can't be done by humans), and then render the results back into three dimensions that humans can use. When applied to seismic data, machine learning can reveal geologic features, properties and trends. And 'this is just the beginning', he said.

Data volumes are growing so large that the traditional analysis techniques are not effective.

Today in seismic processing complex mathematical algorithms in the prestack time/depth migration process are utilized and machine learning approaches are starting to be employed. But machine learning will be an indispensable tool in interpretation, and will complement traditional analysis methods.

Machine learning can be considered a sub-category of artificial intelligence, a label which can be given to 'any technique which enables computers to mimic human intelligence,' he said.

Machine learning is often confused with classical statistics. Machine learning is a class of algorithms in software that learn without being explicitly programmed. The capabilities of machine learning are now practical and economical with modern computing hardware and software architectures. Classical statistics is looking for properties and distributions based on certain assumptions about the data and has been around for centuries. But they come together in that machine learning is rooted in statistics, he said.

Modelling the relationship between different variables is not in itself machine learning. But machines can learn something from the results of it, he said.

'Deep learning' is machine learning which typically employs a number of hidden layers in a neural network.

Machine learning is often applied to a data set indexed by a person. For example if a person has classified a number of images or features about an object, the computer can create an algorithm which 'reverse engineers' the classification, for example to understand that an object with certain features (such as wheels) could be a car.

With deep learning, the computer looks for commonalities in the data set and identifies the features of the object itself before going through the classification process. Deep Learning has gained much interest in the last few years.

'It is confusing sometimes,' he said.

Then there is supervised and unsupervised learning.

Supervised learning means starting with a known set of data and known responses (i.e. you know these are pictures of cars and these are pictures of trains) and use that to train a model, or system which would then be used to identify a picture showing a car or a train. In the business of seismic interpretation, known conditions such as well logs (ground truth) are often calibrated with seismic data to determine reservoir properties. The supervised learning process can help identify those properties within close proximity of the well control.

'Unsupervised learning' looks for natural clusters in the data without the use of well logs. This approach has proven to reveal geologic features in seismic data difficult to interpret previously or not seen at all. Both supervised and unsupervised methods have their place

'Seismic interpreters don't necessarily need to be machine learning experts, they just need to recognise when it can give them a better answer', he said. 'The Paradise machine learning software is equipped with user-guided workflows - 'ThoughtFlows' - to enable every interpreter to apply machine learning technology'.

Direct hydrocarbon indicators

Machine learning can be used to help look for direct hydrocarbon indicators in seismic data.

The work starts with a number of seismic attributes. These are various properties calculated from the seismic. Each attribute has a purpose for highlighting different aspects of geology and stratigraphy, but analyzed together, far greater insights can be obtained. The machine learning process identifies patterns among multiple attributes simultaneously, which is not something a human can do beyond three attributes.

Given a set of 20 seismic attributes that may be candidates to a study area, it would be very time consuming to draw graphs showing how each attribute changes with every other attribute. But instead Principal Component Analysis (PCA) is used in the Paradise software by Geophysical Insights to quantify the variance among the whole set of 20 attributes. PCA is a linear mathematic algorithm which determines those attributes that vary the greatest over a given region, thereby identifying those that are most important.

The more variance an attribute has in a region, the more energy it is imparting. This is a way of the interpreter obtaining a sense of which attributes to use in an interpretation project. The attributes that are contributing the most to the region are good candidates for the application of machine learning.

For the identification of direct hydrocarbon indicators, the most prominent seismic attributes determined from PCA are employed in an 'unsupervised learning' method.

The type of unsupervised learning here is 'Self-Organising Maps (SOM).' The application of multiple attributes in Self-organizing maps produces results in classification and probability volumes that can reveal direct hydrocarbon indicators.

The analysis has proven to reveal flat spots (hydrocarbon contacts), attenuation zones, and anomalous zones associated with hydrocarbons of whatever is in that space is different to rock around it.

Picking geobodies

The seismic analysis can be used to pick out 'geobodies'. There is no firm definition of a geobody but it is usually a specific geological feature, such as a channel or karst.

Employing the results from a SOM analysis, the connectivity of neurons from the classification often reveal geobodies and their areal extent which can be quantified. This quantification can significant impact reserve/resource calculations.

Picking seismic facies and faults

Convolutional Neural Networks, deep learning, has shown to be an excellent approach to identify seismic facies and faults in seismic data.

This supervised neural network approach is applied to a series of seismic lines in a volume where the interpreter has identified specific reflection patterns (facies) or faults. The classification process will take this information and identify seismic facies and fault patterns in all the data.

The same process can be applied to 'well control', using well log data to classify lithofacies (rock layers). If a geoscientist identifies the facies in 3 or 4 wells, a computer can apply it to more wells in the same section.

Re-using attribute relationships

An interesting question is whether specific combinations of seismic attributes employed to reveal certain geologic features, be employed in other areas. The answer is a qualified 'yes', as long as the geology and stratigraphy of the two regions being compared are somewhat similar.

For example there is a combination of 6-10 seismic attributes employed in a SOM analysis that routinely reveal thin beds and detailed stratigraphy.

It is important to always keep in mind that data quality, noise, and acquisition and processing issues can impact machine learning results.



Associated Companies
» Digital Energy Journal
comments powered by Disqus

CREATE A FREE MEMBERSHIP

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

FEATURED VIDEO

A modern data platform to improve predictions and company performance
Jane McConnell
from Teradata

DIGITAL ENERGY JOURNAL

Latest Edition Sept-Oct 18
Sep 2018

Download latest and back issues

COMPANIES SUPPORTING ONE OR MORE DIGITAL ENERGY JOURNAL EVENTS INCLUDE

Learn more about supporting Digital Energy Journal