Great tech can’t deliver its full potential without scaling, but artificial intelligence (AI) and generative AI proof-of-concept (POC) projects are struggling to reach full-scale deployment.
During the 26 August Digitalization Pavilion panel discussion at the International Meeting for Applied Geoscience and Energy (IMAGE) event in Houston, leading industry AI experts said these projects may not reach full deployment because of a lack of vision for the next steps, or having the wrong departments lead the projects. There is no argument, though, that AI and generative AI can greatly multiply what humans alone can achieve.
Ben Wilson, director of products and solutions for AWS, said that, alone, the traditional machine learning and AI efforts from the past decade have been extremely helpful but, when generative AI is brought into the mix, the results can be differentiated insights that didn’t exist before.
“The way I like to think about this is: What can one engineer learn and understand about a basin? How much can they understand by themselves? It’s not everything. These new models that we’ve got with generative AI can remember and know everything and draw on that knowledge to help you understand the results from the machine learning,” he said. “I can take all those reports, that unstructured data, marry it with the structured data and gain new insights it would be impossible for an engineer to have because they can’t remember everything, and that’s really the big difference between the two.”
Nefeli Moridis, who heads subsurface energy solutions for Nvidia’s global energy team, noted increasing industry interest in creating foundation models that are trained on domain-specific information, such as a model trained to handle specific tasks related to the subsurface.
“What we’re seeing is that it’s really accelerating the work of geophysicists and geoscientists, and the reason that that is important, not just because we want to do more work, but also we’re seeing a shift in this industry,” she said. “Some people are retiring, maybe not as many people are coming in, [and they] want to be able to keep that knowledge.”
Additionally, she said, the technology could help eliminate “busy work,” allowing experts to focus “on the actual technical expertise that they have and apply that to the work that they are doing.”
Deepak Gala, chief product owner/business opportunity manager at Shell, said the operator is using AI to improve its drilling outcomes and manage subsurface uncertainty.
“The time it sometimes takes to manage uncertainty could be pretty long,” he said, noting AI models might be able to reduce that timeframe. “Hopefully, that will also result in, I think, better decisions.”
And because models are only as strong as the data they’re trained on, some data sharing could help, he noted.
Sathiya Namasivayam, vice president of data at TGS, said the company dedicated much effort to organizing the data before building a foundation model.
“We have 40 years’ worth of data that we have collected,” he said, noting the company needs to collaborate with operators like Shell and partners like AWS and Nvidia because each has its own domain expertise.
Moridis, of Nvidia, said it’s important for companies working on AI POC projects to have a clear understanding of what business problem they’re trying to solve with the project, how to solve it, and how to scale it.
“There has to be a clear plan from the beginning,” she said. “The scaling part of it is the key component of making it a successful solution, and that includes the technical team that you have internally, the compute [power] that you have.”
Wilson, of AWS, said such projects should also be run by the person who “owns” the profit and loss (P&L) for that division, rather than an information technology (IT) department.
“The reason they need to own a P&L is because you’re going to go buy some data from TGS, you’re going to need some compute [power] from Nvidia or from AWS or someplace,” he said. “You’re going to have to go and get your own buy-in from someplace else.”
He said an IT organization is typically not prepared to make $100 million deals for data and then more to build a model.
“A P&L owner will have the courage to go do that because they would see the ability to transform their business. The IT organization, they struggle because they have constraints like security and all these other things,” he said. “People who own a P&L can see past those things and see that opportunity.”
