Machine learning and automation can diminish the drudgery of interpreting large seismic volumes and allow more time for experts to focus on quality and value. Early adopters are taking advantage with a head start.
Technology has always been a key driver in the success of the oil and gas industry. For the last couple of decades, digitalization has been on the forefront, driving the industry to a whole new level.
For some, digitalization has crossed a tipping point by changing business models that have provided new revenue and produced amounts of opportunities of value. For others, there are still some significant technical and organizational barriers to overcome to successfully scale up.
Buzzwords like artificial intelligence (AI), machine learning and automation are being thrown around in the industry as it is becoming more mainstream. But there is also a lot of confusion: AI, machine learning, and automation – what is really the difference?
Machine Learning vs. Automation
Machine learning is a subset of AI that works by using neural-network computer algorithms on multi-attribute seismic data attempting to discover whatever information is there to help geologists understand the patterns and relate them to meaningful geological interpretation. The computer algorithm is trained from input data and then adapts independently to produce repeatable and reliable results that can be used for seismic interpretation.
Just like with machine learning, automation is designed to conduct tasks and speed up workflows. But there is an essential difference – automation is solely fixed on repetitive, instructive tasks. In other words, automation performs a job, then thinks no further, while machine learning identifies data signals relevant for the future. With it, an advanced question like “Show me in this 3D Survey all relevant Miocene channels” might become a practical reality one day.
Automation and machine learning are useful in seismic interpretation because it solves two significant problems that interpreters face simultaneously:
- Interpreting large volumes of data in less time
- Understanding the relationships of various types of data at once
Automation and machine learning have already reinvented economics and productivity levels across many industries, including pharmaceuticals, banking, and airlines. So why is not the oil and gas industry embracing this data refinement process by allowing computers to do more of the leg work in seismic interpretation?
The answer to this question lies probably in the fact that the language of multi-attribute seismic data is not yet properly understood. One thing is to search for word patterns in the web, another to properly relate peta-byte of seismic data libraries to meaningful geological formations. And as is the case with most raw materials, value often increases with improvement. Raw data, like crude oil, also must be refined for its real value to shine brightly.
I would say that for industrial practical terms, automation, and to some degree, machine learning are still early concepts. But I also believe that there might already be some advanced prototypes out there. The missing piece is an established software powerful enough to combine the two – so that the machine can learn, and with that, assist the interpreter providing better and more accurate results while saving amounts of time.
With that said, there are some success stories out there, moving the industry forward with next-generation seismic interpretation.
Using Deep Learning in Salt Classification
Conventional algorithms for salt picking typically rely on seismic attributes to define a feature space, which requires expert insight and judgment and is prone to error and bias. Researchers have demonstrated how Convolutional Neural Networks (CNN) can classify salt bodies in seismic datasets without using attributes. The CNN algorithm works directly on the seismic data, learning the most discriminative filters on its own, without using attributes whatsoever.
Training the Intuition
Machine learning is not just useful for computing in the inverse direction, such as with inversion and seismic interpretation. Researchers have also shown that machine learning can help rank the performance of different clustering methods using forward models.
Does this mean that it is time to build forward models for machines to hone their seismic intuition?
Automated 3D Horizon Picking
Traditionally, 3D seismic interpretation methods are interactive labor-intensive, time-consuming, and often limited to regions with relatively simple geology. Their main weakness is that they use only a small fraction of the data at the time, solving just a series of local problems. Essentially, interpreters today are still picking seismic amplitudes the same way they did on paper just a few decades ago. This becomes even more evident for challenging areas where horizons are deep, faulted, and affected by seismic noise. In those cases, interpretation becomes a tedious manual task on the computer monitor.
Recently, French software company Eliis commercialized a truly innovative seismic interpretation platform which uses an algorithm based on the global optimization method. Unlike other methods, this algorithm can exploit the full dimensionality of the data to interpret multiple horizons in parallel, using the whole dataset to find a minimum misfit solution, potentially offering a more accurate geological interpretation. Interestingly, Eliis has not managed to disrupt the seismic interpretation market yet.
Old Habits are Hard to Break
There is a lot of interesting and exciting research proving that implementing automation and machine learning can help predict and monitor technologies for the oil and gas industry. As mentioned previously, these buzzwords are frequently used amongst experts. However, the question lies in whether companies are ready to trust and use them. Tightening R&D budgets are not helping, and it is still safer to rely on old, but proven and established methods using human intelligence.
Humans are Still Very Much Needed
A common misconception is that automation and machine learning will replace human interpreters. But adopting new technology does not necessarily mean it is time for seismic interpreters to start looking for a new career. Let us not forget that the most commonly used decision tree-based machine Autopickers have been in use for 20 years – yet each horizon still needs to be picked, examined, and usually modified by a human interpreter.
What machine learning and automation can do, however, is diminish the drudgery of interpreting large seismic volumes and allow more time for experts to focus on quality and value. Ideally, the complexity of the machine learning systems should be hidden to the user by the software and assist the interpreter in delivering results in much less time that are consistent and clearly better than current manual methods familiar to the interpreter.
The oil and gas industry are most likely to benefit if they proactively and intelligently embrace what this new technology has to offer, and early adopters are taking advantage with a head start.
Written by Gaston Bejarano
Gaston is a geophysicist with over 30 years of experience in the industry. He has specialized in the processes that create geoscience software for reservoir characterization. This includes the definition of technical specifications together with end users and the supervision of the software development steps together with the programmer’s team.