Idaho National Laboratory is leading an ongoing project to develop and demonstrate a methodology that leverages multimodal large language models (LLMs) to generate computer programs for nuclear risk and reliability analysis. The outcome of this effort will provide safety analysis teams with a tool to automate the development of probabilistic risk assessment (PRA) models, which will significantly reduce the time required for safety analyses and the regulatory review of safety cases.

A key challenge of this effort is that most existing safety analysis and PRA documentation focuses on commercial light-water reactors (LWRs), whereas many anticipated users of the LLM-based tool are working with advanced reactors (ARs). This gap presents difficulties in training LLM algorithms on multimodal inputs specific to AR design, safety analysis, and risk modeling.

To address this challenge, a comparative study of PRA approaches for LWRs and ARs was conducted to assess the state of the art in risk and safety analysis for both reactor types. We selected five representative LWRs and eight ARs with which to examine differences in risk assessment scope, methodologies, and modeling practices. Key areas of comparison included initiating events, event trees, fault trees, data sources, human reliability analysis, end states, assumptions (such as mission time), and documentation maturity.

This presentation summarizes the major observations and insights from these comparative efforts. The findings are expected to enhance the performance and applicability of LLM-based tools for nuclear risk and safety analysis.