Generative AI Speeding Up Seismic Imaging Workflow

Generative artificial intelligence (AI) can draft marriage vows and create pictures of penguins playing soccer. It’s also useful in the energy sector—capable of generating subsurface images using far less data than previously required.

While massive amounts of compute power are still required for subsurface image generation, machine learning, deep neural networks and computer vision have made it possible to significantly speed up the seismic imaging workflow.

For two years, SparkCognition and Shell have worked together to accelerate seismic imaging using computer vision.

“To their credit, Shell realized this was an open-ended research problem,” Bruce Porter, chief science officer at SparkCognition, told Hart Energy. “They brought it to us as outsiders from the oil and gas industry. We're not oil and gas experts. We are machine learning experts. They wanted to see whether this partnership—with our machine learning and their geoscience—whether that could crack the nut.”

According to Porter, they have. The result is the SparkCognition Oil & Gas Exploration Advisor software.

SparkCognition holds seven patents on the technologies developed to accelerate seismic imaging workflow. Most of those patents are from the “de-noising” migration process, which clarifies the seismic phase imagery.

How long the seismic interpretation workflow takes largely depends on how much shot data needs to be processed, and SparkCognition’s new technology uses between 1% and 3% of the shot data that has historically been used.

“We're not oil and gas experts. We are machine learning experts. [Shell] wanted to see whether this partnership—with our machine learning and their geoscience—whether that could crack the nut.”—Bruce Porter, chief science officer, SparkCognition.

“Given a properly trained neural net, if you prime it with some data points, in this case, shot data, the neural net can fill in for all of the missing shot data, the other 99% to 97% of the shot data that goes unseen and unprocessed,” he said. “The result is these neural nets are able to do what's called the inference step, which is to generate the seismic image. It can do that in a matter of seconds to minutes, filling in for all of these unseen shot data.”

The upshot is that the vast majority of data that’s been collected does not have to be processed, he said.

“Whether that leads to a next generation product in which the acquisition of shot data is reduced, that’s another matter,” he added.

But picking the shots to include takes on more importance when you’re using less than 3% of the shots acquired.

As Porter put it, “There are so few of them, the ones you use matter. You can’t just choose randomly.”

SparkCognition developed a solution to enable the neural networks to select the 1%-3% of shot data that carries the most information and will have the greatest impact in generating an accurate subsurface image. While algorithms run the automated shot selection process, the system is not a complete black box, he said.

Being able to see into the process is important, particularly in light of how wildly off-track some generative AI, such as Chat GPT, have reportedly gone.

Porter said the software generates confidence levels alongside its geological subsurface images, and interpreters can add more shot points and allow it to iterate the new subsurface images with corresponding changes in confidence levels of the image.

“You need the right answer. You need to get the geological substructure correct,” he said. “It’s important that they not be black box-ish, it needs to be one that the human has trust in and can understand where the neural net is being creative in elucidating the geological subsurface and when it’s quite certain of its output.”

Combined approaches

Machine learning is a big field, and many techniques were potential solutions for this particular computer vision problem, Porter said.

“We tried probably 10 to 12 different families of approaches, not just individual algorithms, but whole classes of approaches to the problem, before we settled on the one that did the best,” he said.

But on its own, a generative solution wasn’t enough.

“The machine learning, the AI field has learned over the recent decades that if you approach a problem as complicated as this one using only data, you hit a glass ceiling—and the results aren't great,” he said.

Getting through that glass ceiling called for some creativity and finding a way to bring physics—or geoscience—into the solution, he said.

“A hardcore machine learning person is going to say, ‘No, I don't want to have anything to do with physics. I'm just going to use the data. I'm going to focus on the data, and my algorithms will derive the right answer,’” Porter said. “Uh, no, I don't think that works. We have to have a way of marrying, combining the influence of geoscience into the neural net so that the neural net is drawing inferences. It's creating images that are geologically plausible, and not only plausible, but correct.”

During the companies’ collaboration on the de-noising solution, SparkCognition had access to the Texas Advanced Computing Center at the University of Texas.

“Shell has their own supercomputers,” he said. “But for our research phase, we depended on TACC.”

The technology has been proven on real data from Shell, Porter said.

“We've gotten verification from Shell that the results are very promising, and we are now hardening the software so that it can be released to Shell as a product for deployment,” he said.

The SparkCognition Oil & Gas Exploration Advisor will also be made available to other operators as well.