PG&E deploys nation’s first generative AI model for nuclear at Diablo Canyon

One of the largest utilities in the United States is bringing generative artificial intelligence to its nuclear operations.

Northern California utility giant Pacific Gas & Electric Company today rolled out a model designed for document search and retrieval at California’s last remaining nuclear power plant, Diablo Canyon. The project marks the first commercial deployment of a model developed by AI startup Atomic Canyon and the Department of Energy’s Oak Ridge National Laboratory.

Atomic Canyon, founded late last year, trained sentence embedding algorithms — which convert sentences into numerical representations to capture their meaning — on 53 million pages of publicly available documents from the U.S. Nuclear Regulatory Commission. The company is today debuting its Neutron Enterprise solution, which the company’s founder Trey Lauderdale explained is essentially an AI-powered search engine specifically designed for Diablo Canyon’s 2 billion pages of records.

The app, run on-premise on Nvidia H100 servers, will be used by power plant employees to more efficiently find documents and information in legacy record management systems. The goal is to reduce the time it takes to respond to records requests,such as those from the NRC itself.

The process of getting the tool up and running is already underway at Diablo Canyon, Lauderdale told Latitude Media, which provides nearly 9% of California’s electricity and 17% of the state’s zero-carbon energy. Neutron will be live sometime in spring 2025, he added. 

PG&E’s first-of-a-kind deployment comes as interest in nuclear power’s potential to meet the growing electricity demands of the AI boom is on the rise, especially following announcements from several of the country’s largest data center users in recent weeks. And while California has what amounts to a moratorium on building new nuclear, other utilities around the country have signaled their own interest in nuclear as a clean, firm power source.

Training generative AI for nuclear

When Atomic Canyon first started building AI models for the nuclear sector, they quickly realized that hallucinations were rampant, Lauderdale said, in part because the sector’s terminology is so niche.

To solve that problem, the company partnered with Oak Ridge National Lab, where they used the Frontier supercomputer (the world’s fastest) to train sentence embedding algorithms to understand nuclear terminology. Using those algorithms, the model now more accurately interprets and processes complex nuclear data, like Diablo Canyon’s more than 40 years of documentation on everything from work orders to internal procedures. 

That partnership is in large part what has allowed the startup to remain mostly self-funded, Lauderdale added.

“We would’ve needed to have raised $20 or $30 million to buy a bunch of GPUs to do our own training,” he explained. The partnership was announced in May of this year, and has since moved quickly, he added. The first iteration of the model was released in September, and according to Atomic Canyon, returned the correct search result within the top 10 results 98% of the time, and within the top five results 93% of the time.

Because Atomic Canyon worked with Oak Ridge, a national lab, the model used to create Neutron is open-source. It is already available to independent researchers, nuclear institutions, and other national labs.

That said, Lauderdale expects that utilities and nuclear companies will be wary of working off of any particular open-source tool until a company like PG&E gives it the greenlight. “Over time, we will see more adoption of these open-source models,” he said. “But for the time being, I’m not sure the nuclear power industry knows what to do with open-source models yet.”

Moving forward with Neutron specifically, Atomic Canyon is focused on implementation within the country’s existing fleet of reactors — as opposed to planned-for projects — because training is ongoing, and it’s important to be able to “get in and learn nuclear,” Lauderdale said.

That said, Neutron is really a “foundational layer” on which to build other generative AI use cases for the industry, he added. Using AI to find documents — many of which were only recently scanned from microfiche into often-crooked PDFs — is the lowest-risk application. But in the future, Lauderdale anticipates that generative AI can help actually get plants built more efficiently, like using computer vision for construction workers or automatic speech recognition.

“We are going to see an outsized impact of AI across the entire ecosystem of nuclear power,” he said. “AI is absolutely critical if we’re going to meet these nuclear goals.”