Trained On 53 Million Pages: US Supercomputer-Backed Tool To Search Nuclear Reactor Data

Authored by Aman Tripathi via Interesting Engineering,

Nuclear power stations across North America have begun using a dedicated virtual system called NIVA. The software was created by Atomic Canyon in collaboration with the Electric Power Research Institute, the Institute of Nuclear Power Operations, and the Nuclear Energy Institute.

It provides engineers and technicians across the commercial reactor fleet with an automated method to query vast archives of technical, operational, and regulatory records.

"The U.S. nuclear sector sits on decades of invaluable technical and operational knowledge, but too much of that knowledge remains difficult to access at the speed modern deployments require," said Trey Lauderdale, Founder & CEO, Atomic Canyon.

"The fleetwide availability of NIVA shows that AI in nuclear power is real, operational, and ready to be deployed responsibly at a fleetwide scale."

Technical architecture for model training

The technological foundation of the platform rests on a specialized model family named FERMI. Standard, general-purpose language systems frequently fail to parse the dense technical terminology, precise abbreviations, and domain-specific syntax common in nuclear engineering. FERMI was built to resolve this issue by interpreting technical meaning rather than simply scanning for exact keyword occurrences.

Atomic Canyon collaborated with Oak Ridge National Laboratory to train the models on the Frontier exascale supercomputer. The training dataset included more than 53 million pages of documentation from the U.S. Nuclear Regulatory Commission. The completed model weights are publicly available on the Hugging Face repository.

FERMI functions within an operational software interface called Neutron. While NIVA applies the retrieval models across broader sector archives, the Neutron environment links those capabilities directly to an individual station's private network. This structure lets local engineering departments run searches across their own licensing bases, routine maintenance logs, technical drawings, and operating procedures.

The current rollout provides plant personnel with two functional modules, alongside a third tool scheduled for future release.

Functional modules with commercial expansion plan

The first component is the Knowledge Assistant. Plant staff can enter queries in conversational phrasing to obtain direct answers extracted from verified technical repositories. The system sources information from NRC Regulatory Guides, NUREG reports, NEI guidelines, INPO standards, and EPRI research publications. Every output includes direct citations in the text, allowing engineers to verify statements against the original source documents.

The second component is the Operating Experience Assistant. This tool reviews historical plant performance logs and previous event reports. It evaluates the underlying context of a query instead of looking for surface-level word matches, and engineers can interact with the returned records through a conversational query window to investigate past equipment behavior. A third tool focused on diagnostic and troubleshooting tasks is undergoing software development, with plant trials planned for later this year.

The commercial deployment follows a six-month pilot evaluation period at several operating utilities, including Constellation Energy, which operates 21 reactors across 12 station sites.

Atomic Canyon previously tested its generative tools on-site at the Diablo Canyon facility managed by Pacific Gas and Electric Company.

To finance the expansion of the software across additional commercial sites, Atomic Canyon secured an investment round. Backers include NVIDIA, Plug and Play Ventures, and former Vanguard Group chairman Mortimer Buckley. The funds will support ongoing software engineering and integration across operating facilities.