Accelerating nuclear engineering research with AI

Nuclear power has been proposed as a solution to meet the growing energy needs of artificial intelligence. But what if AI could return the favor and help propel the development and deployment of nuclear energy?

Nuclear engineering Ph.D. student Zavier Ndum Ndum researches the use of large-language models (LLMs), a form of AI used in chatbots like ChatGPT, to assist in nuclear engineering and research. Ndum is building a suite of tools to combine the ability to quickly gather knowledge and conduct simulations all in one package. One of those tools, called frameworks, is RADIANT-LLM.

There is a lot of knowledge circulating about nuclear engineering and physics stored in technical databases and PDFs from various internet sources. Searching for these files manually can take time, but RADIANT-LLM streamlines this process by rapidly pulling information from the relevant documents. The program can also sort through outdated documents to use the most updated versions. These frameworks are lightweight enough to be used on a local computer.

RADIANT-LLM could be described as ChatGPT for nuclear engineers, but that similarity only goes so far, according to Ndum. His framework uses a strategy known as LLM augmentation, which goes beyond just asking a chatbot a question. Furthermore, a publicly used AI agent can’t securely handle uploaded private or proprietary documents that might be necessary for nuclear facilities and must be kept secure. Instead, Ndum’s frameworks can pull from files stored locally on a computer without risking private data leaving that secure environment. RADIANT-LLM also lets users build their own personalized local knowledge bases that can be updated by pulling new reputable information from controlled online sources.

Using Ndum’s LLM frameworks is also more reliable than using commercial AI chatbots like ChatGPT. Such LLM applications are prone to “hallucinations,” when the program generates text that sounds reasonable but is based on made-up information that doesn’t exist or cannot be verified.

“ChatGPT is incredible, but I would say it’s a jack-of-all-trades,” Ndum said. “It’s trained on public data, it’s general, and as a result, it displays only a superficial fluency in nuclear concepts, just like any other public chatbot. It lacks the expertise required for in-depth analyses. When pressed for technical details such as thermal-hydraulic safety margins or reactor-specific security recommendations, they tend to hallucinate. I would rather it tell you that it cannot do something than give you the wrong answers.”

In contrast, RADIANT-LLM shows its work. The resulting output backs up the information it provides with the source name and page number, whether that original document is on a researcher’s local computer or publicly available through the U.S. government.

“The tools are not meant to replace researchers or engineers,” Ndum said. “They are designed to be expert-level assistants that dramatically reduce the time spent on tedious tasks. If we are building a general framework for nuclear engineers, then it needs to stay up-to-date with the current technology.” 

Ndum’s mentors in the College of Engineering see RADIANT-LLM’s vast potential. Dr. Yang Liu, nuclear engineering professor and Ndum’s advisor, said this can help reactor builders streamline the licensing process with the Nuclear Regulatory Commission (NRC). It can sort through troves of NRC information to reduce the human workload.

“We always want a human in the loop,” Liu said. “It would take you a lot of time to read and retrieve information, but now this helps do the work for you, and you just need to verify.”

With a focus on gathering nuclear-specific knowledge in an efficient way, RADIANT-LLM builds on Ndum’s previous work using pre-trained LLMs in other areas in nuclear engineering. His model AutoFLUKA automates simulations by connecting an AI agent to the FLUKA software that simulates radiation transport. Similarly, AutoSAM automates simulations in software that works for thermohydraulics. His most recent framework, AROMA-GPT, is a new LLM agent to be used for safely supervising and controlling advanced nuclear reactors. His goal is to bring these frameworks together in a unified system that uses AI agents to enhance simulations.

These modeling tools require specific input files, which Ndum’s enhanced LLM frameworks can create, unlike normal chatbots. Once the simulation is complete, these frameworks can also analyze the results and present them graphically. Additionally, when a simulation stops running due to errors, the model keeps going. Instead, the AI agent performs a diagnosis on this error and tries to resolve and document it in its knowledge base for future inference.

“The more you use the tool, the more intelligent it becomes,” Ndum said.

Instead of training a chatbot, Ndum writes sets of instructions to get them to do new things, like creating simulation input files and using domain-specific knowledge. To allow developers to use these AI models as a base for their own projects, ChatGPT and Gemini provide access to a pre-trained LLM model through application programming interfaces (API). 

“You’re augmenting the model,” Ndum said. “You’re not retraining it. You’re not changing the weights.”

RADIANT-LLM is model-agnostic, meaning it can run with any version of ChatGPT or Gemini once open APIs are available.

“Fine-tuning is good, but for specific tasks,” Ndum said. “If we are building a tool for nuclear engineers, it needs to stay up-to-date with the current technology.”

According to Liu, undergraduate students were able to use Ndum’s frameworks and learn the simulation programs faster.

“It has a reduced learning curve for new users, and improved productivity for experienced users,” he said.

Ndum presented RADIANT-LLM at the 2025 Institute of Nuclear Materials Management Annual Meeting in August. His paper won the conference’s J.D. Williams Best Student Paper award in the Nuclear Security Division.

“It’s a testament to the hard work that I’ve been doing, as well as the guidance that I’ve received from my mentors,” Ndum said. “It also shows the potential that generative artificial intelligence has in the nuclear domain.”

By Julianne Hodges, Texas A&M University College of Engineering

###