If there was ever a time when working smarter clearly outshone working harder, it was during the pandemic. One bit of evidence? The steady adoption of artificial intelligence (AI) technology to automate routine tasks and boost productivity.
With companies determined to make stay-at-home employees as efficient and effective as they’d been on-site, AI became something of an enterprise golden boy, a must-have business asset as essential as hand sanitizer or masks.
According to Gartner, for example, investment in AI continued unabated throughout the pandemic. But the sector proved to be more than just crisis-proof. A full 75% of the business leaders Gartner talked to were piloting or exploring new AI initiatives. The results of a McKinsey survey were equally upbeat: a majority of the respondents planned to increase their AI investments by at least 10% over the next three years.
One concern, though, was how to connect the dots between AI investments and business value. As COVID-19 put pressure on operating budgets, there was no room for wasted or unproductive expenses. Every dollar spent had to generate a meaningful return.
This was especially true in the hard-hit oil and gas industry. Between historically low prices at the beginning of the pandemic and the continuing struggle between oversupply and lagging demand, the argument for investing in AI had to be incredibly compelling — if it didn’t directly produce revenues, at the very least it had to reduce costs.
Now a Houston energy consultant, Gerchard Pfau was an exploration manager when he took the case for advanced AI to his employer, a multinational energy giant. He knew it could help the company’s geoscience division overcome an efficiency challenge related to seismic interpretation. That’s the part of the exploration process where geoscientists parse data and build 2D and 3D models to help them identify oil- and natural gas-rich zones underground.
The issue, it seemed, was that instead of being able to devote most of their time to the art of geology, geoscientists were tediously interpreting line after line of seismic data to analyze examples of subsurface features, a process that would eventually yield accurate models and maps.
As a geoscientist himself, Pfau had experienced the same thing.
“It can take weeks of agonizing work to try and repeatedly interpret in most software to identify various geologic features in seismic,” he said.
To speed up the mechanics of model-building, improve geoscience decision-making, and potentially reduce both drilling risk and time to oil, the company needed a different interpretation system — one that would enable its geoscientists to spend less time clicking and more time thinking. And if it allowed them to work remotely, without requiring data scientists or data engineers to manage the seismic data, well, that would be a bonus.
The answer came in the form of an application called InteractivAI, a product of Houston’s Bluware. Because it learns “on the fly” without data preparation, cloud based InteractivAI cuts seismic data interpretation time from weeks to just minutes.
Training the Brain
At its most basic, AI’s goal is to teach computers to imitate the way people think and behave.
Just like humans, the computer brain starts out with a pretty blank slate. We have to be taught; they need to be fed training data. The more humans learn, the better we become at recognizing patterns and applying correct assumptions to other information we encounter, and the same goes for computers: If it walks like a duck and quacks like a duck and looks like a duck, both the human brain and computer brain will identify it as a duck. And, usually, that’s exactly what it is. Unlike humans, though, it can take the computer a lot longer to learn something and it requires many examples to refer to.
Of course, geoscientists aren’t looking for ducks. What they want to do is visualize every horizon, every fault, on every inline and cross line that could possibly exist within a subsurface formation and be holding hydrocarbons. Considering a seismic shoot may contain millions of images, there’s no practical way geoscientists could examine every single one, so they turned to interpretation software.
With traditional seismic AI, training the brain goes something like this. The geoscientist labels a set of horizons, faults, usually on inlines and/or crosslines to create the training data for the computer, then lets the machine find correlations between what it “knows” and similar features in all the other datasets. The intended result is an accurate interpretation of what lies underground.
Interactive vs Pre-canned AI
The problem with conventional AI based seismic interpretation software, at least according to Pfau, is this: not only is it tough to train it to recognize the patterns geoscientists want them to, but it can also be days or weeks before it generates results.
Perhaps more important, if the geoscientist, who is, after all, only human, makes a mistake in labeling, the machine will replicate that mistake. If it fails to recognize important features or suggests patterns exist where they really don’t, that can be an incredibly costly issue.
Pfau credits the Bluware engineers who developed InteractivAI with overcoming those obstacles. How they did it is reflected in its name.
The application is, truly, interactive.
To explain the difference between InteractivAI and other AI products, Pfau says to consider a geoscientist who sees a tear between lines in a seismic image, highlights it, and labels it as a fault. Using the traditional deep learning approach, the geoscientist trains the software by labeling the faults over and over and hopes it will return an accurate interpretation.
InteractivAI, on the other hand, learns instantaneously as the geoscientist interprets in real-time.
The geoscientist labels a feature, and the network almost immediately detects features similar to it everywhere else in the entire seismic volume. Because InteractivAI confirms what the computer has learned, the geoscientist can see how well training has been applied by the application, if it’s highlighting something that was missed during labeling, and if fine-tuning is required, either before running additional scenarios and to eliminate the propagation of errors.
“As the computer identifies the geobodies in the seismic, for example, the geoscientist can either approve what was labeled is correct, or confirm it is wrong and correct it,” Pfau said. “This is where their expertise comes in. But instead of doing the manual labeling, typically they only interpret less than 1% of the data and the machine does the rest.”
InteractivAI is intended to be a deep learning solution that builds on the geoscientists’ knowledge of a given region. It can illuminate the subsurface like a flashlight and test different theories interactively with the computer and the data. The tool aims to create a virtuous circle of knowledge, data, and the result is a solution that amplifies a geoscientist’s intimate knowledge of the heterogeneity and anisotropy of the earth.
Better Decisions in Less Time
To some degree, the proliferation of AI has created a nagging fear amongst some geoscientists that machines will make their jobs obsolete.
That’s hardly accurate, Pfau said. When Bluware created InteractivAI, they set out to reduce the time-consuming mechanics of manual data interpretations. That would increase geoscientist efficiency and help them make better decisions fast. In turn, they’d be more valuable to the company, not less.
As Piette notes, AI is not intended to replace geoscientists. There’s no substitute for their education or experience. After all, they’re the ones teaching the software by example.
It’s likely, though, that geoscientists that embrace AI will replace those who aren’t using it.
Pfau believes that day may come sooner rather than later.
“The goal is to enable energy companies to get to business decisions faster, whether that’s to drill a well or sell their stake in a reservoir,” he said. “By making the mechanics of interpretation faster, these tools let geoscientists spend more time thinking and building more models. And when you’re talking about drilling a $100 million oil well, you want all the thought possible behind the decision.”
As for Pfau’s former employer, geoscientists report that they’ve already tightened the feedback loop between what they know and what InteractivAI infers. The result is higher-quality interpretation faster than with conventional approaches.
And that is what working smarter is all about.