Abstract
Nuclear energy research is characterized by extreme complexity, requiring decades to transition from theoretical concepts to operational reactors. This study evaluates the potential of frontier Large Language Models (LLMs)—including GPT-4.5, Claude 3.5, and Gemini 2.0—to accelerate this process. Through a workshop at Oak Ridge National Laboratory (ORNL), 14 interdisciplinary teams applied these models to diverse fusion and fission challenges. Results indicate that while LLMs cannot yet replace domain expertise, they serve as powerful "scientific accelerants" for cross-disciplinary synthesis, workflow automation, and rapid prototyping.
Introduction
The development of nuclear energy systems involves "multiphysics" expertise, necessitating the simultaneous management of plasma physics, neutronics, materials science, and thermal hydraulics. The inherent complexity often leads to protracted development timelines. This research investigates whether the latest generation of reasoning-capable LLMs can compress these timelines by identifying research gaps, generating simulation code, and designing experimental programs for next-generation reactors.
Methodology
The study utilized a "human-in-the-loop" approach across 14 technical challenges. Teams leveraged frontier models such as ChatGPT-4.5, o1, o3, and Claude 3.5 Sonnet. The methodology focused on structured workflows rather than simple prompting, employing iterative refinement and Retrieval-Augmented Generation (RAG) to interface with scientific literature and simulation frameworks like XGC, NekRS, and OpenMC.
Key Insights and Technical Synthesis
The Knowledge Bridge
LLMs demonstrated a unique capability for cross-disciplinary synthesis, connecting disparate fields such as plasma control and materials science. This "knowledge bridge" allows researchers to identify solutions that might otherwise be overlooked in siloed environments. For example, models successfully connected plasma control problems with materials science solutions during the intensive workshop sessions.
Workflow Automation and Rapid Prototyping
A primary strength of current LLMs lies in designing experimental programs and workflows. For instance, the "Foundation Model for Fusion" team produced a technically sound concept paper and LaTeX-formatted workflow within hours. Similarly, AI-directed experimental designs for Molten Salt Reactors included proposed budgets and timelines that would typically take weeks to compile manually.
Code and Logic Generation
While the models struggled with production-ready code for highly complex solvers, they were effective at generating Mathematica scripts and structured outlines for simulation input decks. In material degradation studies, AI-generated models achieved a predictive accuracy of $R^2 \ge 0.90$ using acoustic and strain data.
Results and Discussion
The workshop outcomes across the 14 domains are summarized in the following table, detailing the specific challenges, outcomes, and limitations encountered.
Limitations: The "Hallucination" and Data Barriers
Despite their strengths, LLMs are prone to physical hallucinations, such as suggesting impossible cooling fluid velocities (e.g., 50,000 m/s). Furthermore, performance is significantly hindered by "data silos"—specifically paywalled literature and export-controlled software (e.g., SCALE or MCNP) that the models cannot access. There is also a notable lack of source transparency in AI-generated images, making it difficult to verify material-specific defects.
Conclusion and Future Directions
The current generation of LLMs acts as a "complement" to expert validation rather than a replacement. Future research should prioritize the development of "Nuclear-Specific LLMs" trained on curated, secure datasets. The ultimate goal is the creation of "Agentic Workflows" where autonomous AI networks coordinate theory, simulation, and experimental validation in a continuous feedback loop to drastically reduce the time-to-market for clean nuclear energy.
Summary updated