Enhance nuclear research using text-generating AI

Text generation AI programs like ChatGPT are known for their everyday tasks, such as answering questions, but that’s not the only one. These AI agents can use their textual parallel power to support further nuclear science and health physicists, as demonstrated in the work of graduate student Zavier NDUM in nuclear engineering.

Large-scale language models (LLMs) such as ChatGPT are the form of artificial intelligence that uses written data to generate text based on predictions. LLM is useful for many domains, such as software engineering and research support, but it is less common for use in nuclear science. This is because nuclear engineering research often uses proprietary data that cannot be safely supplied to common LLMs.

“When it comes to nuclear science, there are a lot of unique data, and data sharing has security issues,” says Ndum. “You can’t give this knowledge to ChatGpt or Copilot. If there’s something you can use within your organization to automate your workflow, it’s more productive and efficient.”

Ndum recently wrote a paper on a project using LLM in nuclear science. The AI ​​agents described in NDUM’s paper Autofluka can help automate tasks in nuclear science research, such as running computer simulations on software called Fluka. The application can retrieve, edit the input file, run the simulation, and then analyze the results with a plot graph.

Other researchers can use their own data to securely build databases using NDUM’s applications to use their own data that goes anywhere other than the originating computer.

“Instead of extracting simple answers, this information can be sorted very easily and very quickly,” says Ndum. “If you have these documents and you have application gain access to those documents, you can easily run Q&A. It can encourage your model to suggestions to enhance a particular application.”

To build Autofluka as a proof of concept, Ndum had to work with what he had. He was unable to access work on the Monte Carlo N-Particle (MCNP), a commonly used but regulated computer simulation code in nuclear science. However, Fluka is well-like to MCNP, allowing you to easily replicate NDUM models for IT and other simulation programs.

Another challenge for NDUM was to shift the focus of his research into AI agents. Before arriving at Texas A&M, he was working on health physics and dosimetry, or researching radiation absorbed by living tissue. However, after working with Dr. John Ford, a professor of nuclear engineering researching health physics, NDUM used the work to provide a case study of autofull cappeda. He also cites the guidance of his PhD, Dr. Yang Liu, professor of nuclear engineering. Dr. Jian Tao, a professor at the University of Performance, affiliated with the Faculty of Nuclear Engineering, with an advisor and building applications.

“When you step into a new territory, it’s always challenging, but you keep working and I believe there’s something convenient that you can make,” he said. “When you make that effort, you know that’s good.”

NDUM, president of the Texas Society of Health Physics Editorial (STC-HPS), also hopes to bring this approach back to the original field of health physics. At the annual meeting of the chapter at the University of Texas at Arlington in October, he presented a speech on the use of LLMS in health physics, serving as a virtual assistant for experts in the field. He tested the application from his paper to retrieve information from health physics-related documents and reduce the search for hours to seconds. This is useful for health physicists like Radiation Safety Officers (RSOs) who need to know about annual radiation dose regulation and the operation of various machines.

“Sorting these regulatory guides for RSO can be time-consuming and intense,” says Ndum.

NDUM also offered seminars last fall as part of a semester-wide series. He will give another talk on the topic at the STC-HPS Student Conference held at Texas A&M University in April. Nuclear Science, Health Physics and Engineering students from universities other than Texas A&M will be participating.

“They can see how they can utilize this technology in their own specific research areas,” he said.

Currently, NDUM is researching by developing sophisticated LLM applications designed to answer complex, domain-specific questions in nuclear science research. The application also allows you to sort and analyze documents and integrate information from online sources in real time. It can read and process multiple file types, including PDFs, images, and spreadsheets, to provide comprehensive support for your research tasks. These capabilities allow applications to provide nuanced answers and insights to make nuclear science research more efficient.

“This is a new realm and it’s really important to explore it in nuclear science,” Ndum said. “We’re going to continue working to see what we can build with this knowledge.”

Liu is director of science machine learning for Advanced Reactor Technologies (Smart) Group and believes Zavier’s work will impact the broader nuclear engineering community.

“LLM is the focus of our group’s research and we are fortunate to be able to push the boundaries of nuclear science applications to Zavier,” Liu said. “His innovative approach to integrating AI into nuclear research is exactly the kind of future mindset needed in this field. The ability to leverage AI for safe, domain-specific automation is a game changer, and Zavier’s contribution paves the way for more efficient data-driven advances in reactor modeling, health safety, and nuclear safety.