Refueling at the High Flux Isotope Reactor at Oak Ridge National Laboratory.Oak Ridge National Laboratory
The nuclear power industry has long run on slow-moving currency: paperwork.
Building a new reactor or, as in the case of California’s Diablo Canyon, simply extending a license, requires wading through mountains of complex regulations, maintenance logs, and engineering records.
It’s a process that costs plants like Diablo Canyon an estimated 15,000 staff hours a year in document searches alone.
But the paper chase is over. A powerful solution has arrived, thanks to the nation’s most advanced computing.
Tech startup Atomic Canyon has harnessed the power of the Frontier supercomputer — the world’s first exascale machine, located at the Department of Energy’s Oak Ridge National Laboratory (ORNL) — to train the first nuclear-specific Artificial Intelligence models.
The system, called the Neutron platform, can accurately search for and interpret diverse, complex nuclear information.
Frontier has a peak performance of 2 exaflops per second — meaning it can perform more than a billion-billion calculations per second. Credit: Carlos Jones/ORNL, U.S. Dept. of Energy
Overcoming the documentation challenge with AI
The idea originated with Diablo Canyon, California’s only operational nuclear power plant.
It is situated on the Pacific Coast, on a cliffside, and supplies roughly 8% of California’s total electricity, powering over 4 million people.
Although originally scheduled for decommissioning in 2025, California state leaders extended its operating license to 2030 in 2022 to meet the growing energy demands.
“That meant we had to pivot into restarting a lot of things, including a massive application to the NRC that was about 3,000 pages,” said Maureen Zawalick, vice president at Diablo Canyon.
“So, we were doing things in a very time-compressed manner — going through thousands and thousands of documents and records and information to meet all the requirements,” Zawalick added.
The difficulty of navigating Diablo Canyon’s massive database makes the time-consuming process of license changes even more burdensome.
The problem wasn’t a lack of AI; commercial tools exist. The issue was precision. Off-the-shelf AI couldn’t handle the nuclear industry’s highly specific jargon and abbreviations.
To ensure reliability, the Atomic Canyon team chose to build a dedicated AI model from scratch.
“To ensure accuracy and reduce hallucinations, we needed a tremendous amount of data and the ability to run the data many times over to properly train the AI models,” said Trey Lauderdale, the founder and CEO of Atomic Canyon. “For us to start building AI that would work reliably, we needed a supercomputer.”
Frontier’s superpower
Atomic Canyon was allotted 20,000 GPU hours on the Frontier supercomputer to develop the Neutron platform.
This platform uses specialized FERMI sentence-embedding models — a retrieval-focused AI that assigns numerical values to words — to accurately search for and understand complex nuclear data by recognizing both terminology and context.
The models were trained extensively on the Nuclear Regulatory Commission (NRC)’s ADAMS database — over 3 million documents — to prevent the hallucinations common in general LLMs. The database contains the history of every U.S. nuclear reactor since 1980.
Even in the early stages, the new AI tools are showing impressive results, with staff experiencing increased productivity and a very high return on investment.
These models, designed for use across the entire U.S. nuclear fleet, will help facilities comply with NRC requirements, including licensing, construction oversight, reactor design evaluation, and decommissioning review.
As the nation focuses on nuclear energy to meet rising demands, these AI tools will be crucial for accelerating licensing, maintenance, and overall safety procedures.
The Blueprint
