DOE Completes Nuclear License Document Test Using AI
The Department of Energy announced the completion of a proof-of-concept demonstration using Everstar's AI tool to generate Chapter 5 of an NRC license application. The 208-page document was created in approximately one day-work that typically requires four to six weeks from a team of experts.
A subject-matter expert who reviewed the document described the results as "consistent with what would be expected of a Revision 0 document," according to the DOE's Office of Nuclear Energy.
The project involved the DOE, Everstar, Idaho and Argonne National Laboratories, and Microsoft. It emerged as part of the Genesis Mission, a national effort launched in November 2025 to accelerate AI use in scientific research.
How the Tool Works
Everstar's tool, called Gordian, uses retrieval-augmented generation architecture. This approach anchors the AI's responses to source documents rather than allowing it to generate unsupported claims, which reduces hallucinations common in other generative AI systems.
Experts authored "skill modules"-structured instruction sets that govern how the AI behaves at each step. These modules enforce citation requirements, define regulatory mapping, specify section-by-section requirements, and include quality assurance routines.
The tool drew from multiple sources: the INL site-level documented safety analysis, the Generic High Temperature Gas Reactor preliminary safety analysis documents, and the NRC's Agencywide Documents Access and Management System database.
What the Test Revealed
The AI correctly identified missing, derived, or inconsistent information across source documents-issues that might otherwise persist through multiple human review cycles, according to DOE officials.
The tool showed limitations in distinguishing between a parameter's safety limits and its operational ranges, a distinction that carries both technical and regulatory significance in nuclear licensing.
Because preliminary safety analysis documents are inherently incomplete, the completeness of the AI output directly reflected the inputs rather than representing a tool failure, the DOE said.
Next Steps
The DOE identified several priorities:
- Develop a dedicated review tool to systematically document and resolve discrepancies
- Pursue a formal NRC working group on AI-assisted licensing
- Evolve document standards so source materials can be properly parsed and cited by AI systems
Long-term applications include using the same modular architecture for additional safety analysis report chapters, preliminary safety analysis generation, nuclear quality assurance documentation, and fuel fabrication facility licensing.
A pilot program using approved documented safety analysis documents from willing participants is planned as the next phase, moving the capability toward production-grade regulatory use. This phase would require a secure operating environment, unlike the proof-of-concept project, which used open-source documents.
The DOE intends to use AI tools for draft authorship, shifting expert labor toward refinement and review rather than initial document creation.
Your membership also unlocks:
