Human-AI Collaboration in Nuclear Fault Diagnosis: Operator Evaluation of LLM-Based Assistance

Nuclear power plant operators face complex fault diagnosis challenges that require rapid interpretation of information across multiple systems and parameters. This study investigates the potential of large language models (LLMs), augmented with plant-specific documentation, to support operator diagnostic reasoning. A prototype LLM-based diagnostic assistance system was developed and evaluated in a nuclear control room simulator. Six experienced operators from two crews participated in a pilot study involving an intentionally difficult fault scenario. Once the fault had sufficiently manifested, the simulation was paused and the crew was granted access to the LLM. The model had access to operating procedures, technical manuals, and plant terminology, enabling it to generate diagnostic recommendations in response to operator prompts such as "Steam pressure increasing, feedwater flow decreasing." Operators provided qualitative feedback on system usefulness, presentation, and potential integration into operational practice. They viewed the LLM as a supplementary information source rather than an autonomous system. Identified use cases included diagnostic support, document retrieval, access to historical data, and operator training. This study represents an initial step in an iterative development process guided by operator feedback. The results demonstrate the feasibility and acceptance of LLM-based diagnostic assistance in nuclear control environments and outline a path toward transparent, trustworthy, and human-centered AI tools that enhance safety and efficiency while preserving human authority.

Keywords: Large language models, nuclear power plant operations, fault diagnosis, control room simulation, decision support systems, human-centered AI.