Nuclear power has been proposed as a solution to meet the growing energy demands of artificial intelligence. But what if AI could return the favor and help accelerate the development and deployment of nuclear energy?
Nuclear Engineering PhD student Zavier Ndum Ndum is researching the use of large-scale language models (LLMs), a type of AI used in chatbots such as ChatGPT, to aid nuclear engineering and research. Ndum is building a suite of tools that bring together the ability to quickly gather knowledge and run simulations in one package. One of the tools called frameworks is RADIANT-LLM.
Much knowledge about nuclear engineering and physics is distributed in technical databases and PDFs from various Internet sources. Manually searching these files can be time-consuming, but RADIANT-LLM streamlines this process by quickly retrieving information from relevant documents. The program can also sort out old documents and use the latest versions. These frameworks are lightweight enough to be used on your local computer.
Ndum says RADIANT-LLM could be called ChatGPT for nuclear engineers, but that’s where the similarities end. His framework uses a strategy known as LLM augmentation, which does more than just ask the chatbot questions. Additionally, publicly used AI agents are unable to securely process uploaded private or proprietary documents that may be required by nuclear facilities and must be kept securely. Instead, Ndum’s framework can be retrieved from files stored locally on your computer without risking your private data leaving a secure environment. RADIANT-LLM also allows users to build their own personalized local knowledge base that can be updated with new, trusted information from managed online sources.
Using Ndum’s LLM framework is more reliable than using commercial AI chatbots such as ChatGPT. Such LLM applications are susceptible to “hallucinations” when the program produces text that sounds reasonable, but is based on fabricated information that either does not exist or cannot be verified.
“ChatGPT is great, but I think it’s a jack-of-all-trades,” Ndum says. “It is trained on public data and is general, and as a result, like other public chatbots, it exhibits only superficial fluency in nuclear concepts. It lacks the expertise needed for in-depth analysis. It tends to hallucinate when pressed for technical details such as thermo-hydraulic safety margins and reactor-specific safety recommendations. I’d rather it tell me it can’t do something than give me the wrong answer.”
In contrast, RADIANT-LLM shows how it works. The resulting output backs up the information provided, along with the source name and page number, regardless of whether the original document is on the researcher’s local computer or publicly available through the U.S. government.
“This tool is not meant to replace researchers or engineers,” Ndumu said. “These are designed to be expert-level assistants that significantly reduce the amount of time spent on tedious tasks. If you’re building a general framework for nuclear engineers, it needs to stay up to date with current technology.”
Ndum’s leadership in the College of Engineering recognizes the great potential of RADIANT-LLM. Dr. Yang Liu, a professor of nuclear engineering and an advisor to Mr. Ndum, said this could help reactor builders streamline the licensing process with the Nuclear Regulatory Commission (NRC). Reduce human workload by classifying the treasure trove of NRC information.
“We always want human involvement,” Liu said. “It would take a lot of time to read and retrieve information, but now it’s easier and all you have to do is review it.”
RADIANT-LLM focuses on capturing nuclear-specific knowledge in an efficient manner and builds on Ndum’s previous work using LLMs pre-trained in other areas of nuclear engineering. His model, AutoFLUKA, automates simulation by connecting an AI agent to the FLUKA software that simulates radiation transport. Similarly, AutoSAM automates simulations with software that supports thermohydraulics. His latest framework, AROMA-GPT, is a new LLM agent used to safely monitor and control advanced nuclear reactors. His goal is to integrate these frameworks into an integrated system that uses AI agents to enhance simulations.
Unlike regular chatbots, these modeling tools require specific input files that can be created with Ndum’s enhanced LLM framework. Once the simulation is complete, these frameworks can also analyze the results and display them graphically. Additionally, even if an error stops the simulation run, the model continues. Instead, the AI agent performs diagnostics on this error and attempts to resolve and document it in a knowledge base for future inference.
“The more you use this tool, the more intelligent it becomes,” Ndumu says.
Instead of training a chatbot, Ndum creates a set of instructions that make the chatbot do new things, such as creating a simulation input file or using domain-specific knowledge. To enable developers to use these AI models as the basis for their own projects, ChatGPT and Gemini provide access to pre-trained LLM models through application programming interfaces (APIs).
“You’re extending the model,” Nudam said. “I’m not retraining. I’m not changing the weights.”
RADIANT-LLM is model agnostic, so you can run it on any version of ChatGPT or Gemini once the open API is available.
“Tweaking is good, but only for specific tasks,” Ndum says. “If you’re building a tool for nuclear engineers, it has to stay up to date with current technology.”
According to Liu, undergraduate students were able to learn simulation programs faster using Ndum’s framework.
“The learning curve for new users will be shortened and productivity for experienced users will be increased,” he said.
Ndum announced RADIANT-LLM at the 2025 Institute for Nuclear Materials Management Annual Meeting in August. His paper won the conference’s J.D. Williams Best Student Paper Award in the Nuclear Security category.
“This is a testament to the hard work I have put in and the guidance I have received from my leaders,” Ndumu said. “This also shows the potential of generative artificial intelligence in the nuclear domain.”
Written by Julianne Hodges, Texas A&M University College of Engineering
###
Disclaimer: AAAS and EurekAlert! We are not responsible for the accuracy of news releases posted on EurekAlert! Use of Information by Contributing Institutions or via the EurekAlert System.
