5/4/2026 Michael O'Boyle
Written by Michael O'Boyle
Fermilab physicists and an Illinois Grainger Engineering professor of civil and environmental engineering improved an AI model analyzing neutrino experiment data by encoding physics principles in the model.
Hadi Meidani, a civil and environmental engineering professor in University of Illinois Urbana-Champaign’s Grainger College of Engineering, develops AI models that conform to the laws of physics. Ordinarily, he applies them to areas such as structural engineering and computational mechanics, but he has recently worked with high-energy physicists to develop a method for enforcing fundamental physics in data analysis pipelines. In a study recently published in the Journal of High Energy Physics, Meidani and Fermilab scientists proposed a model to improve particle classification models for neutrino experiments.
“Fermilab was part of a multi-institution effort to use machine learning to identify particle types from energy readouts, but machine intelligence by itself could only take them so far,” Meidani said. “My specialization is physics-informed AI, so I was able to help them add physical laws into the model’s reasoning. That led to a substantial improvement in identifying and classifying particles.”
The collaboration was facilitated by the Illinois Discovery Partners Institute. Both Meidani’s team and Fermilab scientists participated in discussions with the institute, and they learned that they were working on a common problem: incorporating physics principles into AI model predictions.
“Unlike general AI tasks such as image analysis or language modeling, problems in physical science and engineering have absolute governing rules originating in physics. For instance, mass and energy must be conserved. These are things that generic AI models are not built to consider, so specialized, physics-informed AI is needed.”
Hadi Meidani, CEE Associate Professor
The researchers sought to improve the NuGraph2 system developed to support liquid argon time projection chambers — a kind of experiment designed to identify the very weak interactions between neutrinos and ordinary matter. Data consists of energy localization readouts from which particle types and locations must be deduced. Prior results from NuGraph2 were acceptable but had room for improvement.
“The sparsity of the data in these kinds of experiments meant we had to turn to a kind of AI called graph neural networks to capture the geometric complexity, but that in turn made predictions harder to interpret by physicists,” explained Fermilab scientist Giuseppe Cerati. “Our work shows that performance was substantially improved by embedding physics context directly at the input level through features encoding geometric relationships, topology, and expected physical behavior — particularly for rare particle topologies for which there may be limited training data available.”
Meidani and the Fermilab team tried several approaches for incorporating physical constraints into NuGraph2’s analysis, and they found the most success by augmenting the input data with additional context-aware features that provide additional information on how data points relate to their surroundings.
“When you look at spatial energy readouts from these kinds of experiments, two qualitative event classes appear: ‘tracks’ in which particles essentially move in straight lines, and ‘showers’ in which particles cascade,” Meidani said. “Each has its own implications for produced particles. After trying several options, we found that engineering a data feature that captures track-like and shower-like tendencies really helped the model enhance its classification.”
The study focused on events known as Michel electrons, which have proven difficult for AI models to identify owing to scarcity and underrepresentation in datasets.
Besides improving model performance, the researchers also sought to improve the explainability of the AI models. This result will be communicated in a future publication.
Vitor Grizzi and Maggie Voetberg of Fermilab and V. Hewes of the University of Cincinnati also contributed to this study.
The article, “NuGraph2 with context-aware inputs: physics-inspired improvements in semantic segmentation,” is available online. DOI: 10.1007/JHEP03(2026)149
This study was funded in part by the University of Illinois Discovery Partners Institute and the U.S. Department of Energy, Office of Science, Office of High Energy Physics
Illinois Grainger Engineering affiliations
Hadi Meidani is an associate professor of civil and environmental engineering in the Department of Civil and Environmental Engineering in the University of Illinois Urbana-Champaign’s Grainger College of Engineering. He is also affiliated with the Siebel School of Computing and Data Science and the Department of Biomedical and Translational Sciences in the Carle Illinois College of Medicine.