It’s often referred to as dark magic within the nuclear science community.
Human experts evaluate experimental nuclear physics data to provide codes that model nuclear reactors, nuclear weapons, and nuclear detection systems.
Assistant Professor and Charles P. Postelle Professor in Nuclear Engineering Vladimir Sobes wants to shed more light on the dark magic. His “passion project” for the last decade has been centered around finding ways to use artificial intelligence/machine learning (AI/ML) capabilities to produce faster, more reproducible, and more reliable evaluation of nuclear data.
“Historically, the reason that data production is sort of black magic is because so much of it has been the tied up in the expert judgment,” Sobes said. “These experts would look at the experimental data, and they would say, ‘Ah, it’s this, or it’s this, or it’s this,’ and then they would put down numbers on paper and then give it to the users. The users just take those numbers and trust them.”
Nuclear data represents measured probabilities of various physical interactions involving the nuclei of atoms. Nuclear data collections, or libraries, incorporate data from multiple sources that have been assessed by nuclear data evaluators who review and combine all available data sets, determine the highest quality data, and decide upon a set of standards.
The current practice of data evaluation by human experts is laborious and time consuming. Sobes and his research group are trying to automate the process by capturing and encoding the knowledge of the evaluators and experts to make into an algorithm that is reproducible and archived. The algorithm could potentially do overnight what takes a human expert two years to accomplish.
“We are trying to make the decisions that the human experts make, and we’re trying to make them systematically, as opposed to an opinion,” Sobes said. “What that leads to is not only reliability, but an automated self-awareness of how well we can actually do.”
Laying the Groundwork
Sobes first developed the idea for his nuclear data project while in graduate school at the Massachusetts Institute of Technology (MIT). He gained even more knowledge by connecting with a world expert in the subject once he began working at Oak Ridge National Laboratory.
“I felt like I was an apprentice in a shoe shop and learning the tricks of the trade. I came up with this idea that if he can teach me, I could teach a computer,” Sobes said. “That’s the sort of overarching vision of this. I want to encapsulate that into a computer program, and I want to essentially teach a computer to do this, instead of him having to teach me.”
Sobes didn’t have the bandwidth to dive deeper into the project until he arrived at UT in 2020 and was able to build a team of students that now includes: Noah Walton, Cole Fritsch, Justin Loring, Jake Forbes, Aaron Clark, Amanda Lewis, Oleksii Zivenko, and Jordan Armstrong. Sobes also collaborated with Assistant Professor Hugh Medal (Department of Industrial Systems and Engineering), who is an expert in optimization and AI/ML systems.
This summer, the nuclear data project received a Cross-Institutional Research Engagement Network (CIREN) grant. Graduate students Walton and Fritsch also have been supported by the National Nuclear Security Administration (NNSA) Nuclear Science and Security Consortium. Walton defended his thesis in October with a working prototype of the algorithm.
“The vision would be slowly for the nuclear data community of these artisanal experts to start adopting this software and using it,” Sobes said. “The software is not meant to replace them. The software is meant to free them up to think about the bigger physics questions while giving them a tool to use. Now, they can do this overnight and have two extra years of thinking that you can do on other stuff.”
Sobes is hoping to help solve another problem in the industry with his project. The field of nuclear data evaluation is suffering from an aging population of experts that could hinder future nuclear work if not addressed.
“Being at UT provides an opportunity for me to infuse young talent into this field,” Sobes said. “We’re working to get grants to train young experts in this field that has sort of been behind the shroud. I see part of my job as bringing young people into this community, introducing them to it, and placing them in jobs in this area to replace the people retiring.”
Contact
Rhiannon Potkey (865-974-0683, rpotkey@utk.edu)