Traditional, data-hungry AI algorithms face unique challenges in searches searching for rare physics events: rare events are, by definition, exceptionally scarce. Moreover, since potential discoveries could represent Nobel Prize-level breakthroughs, the AI models must maintain rigorous statistical validity and complete transparency. Our group aims to Rare AI algorithms for specific challenges in particle physics: Bayesian Rare Event Surrogate Models that accelerate expensive simulation, specialized foundation models that enable autonomous discovery, and counterfactual fairness techniques that ensure robust validation of potential discoveries.
Different Rare Event Search Detectors generate distinct data signatures requiring specialized neural architectures: LEGEND's germanium detectors produce sharp, brief pulses; ABRACADABRA's SQUIDs generate massive time series with rich frequency content; KamLAND-Zen and XENONnT create multidimensional spatiotemporal patterns; and LIGO detects unique chirp-like gravitational wave signals. Our group develops neural networks that leverage the inherent structure and symmetries of each data type. We create targeted benchmarking datasets that connect algorithm performance directly to each experiment's specific challenges, ensuring our innovations translate into meaningful scientific discoveries.
Our research focuses on deploying ML models for real-time data acquisition in particle physics experiments. We're collaborating with experts on FPGA and heterogeneous system to integrate advanced neural networks onto RFSoC (Radio Frequency System-on-Chip) platforms, enabling KamLAND-Zen to reconstruct particle positions and control detector functions in real-time.
Releasing real neutrino and dark matter detector data to the public, allowing core AI algorithms to extract the signal and produce real physics results thereby advancing fundamental science!
The Germanium Machine Learning (GeM) Group within the LEGEND collaboration, leverages efficient and interpretable AI to aid all aspects of LEGEND analysis while educating domestic and international collaborators to gain AI experiences.
"For the invention of a novel machine learning algorithm that broke down significant technological barriers with monolithic liquid scintillator detectors and, in turn, delivered the world’s most sensitive search for neutrinoless double beta decay." link
"Aobo Li is a postdoctoral research associate and COSMS Fellow in the College’s Department of Physics and Astronomy. He is described by his professors as a “rising star in the neutrino and nuclear physics community.” He leverages artificial intelligence to facilitate the experimental search of neutrino-less double-beta decay – part of an effort to explain why there was matter left after the Big Bang instead of only pure energy. His efforts to stay informed about new tools in machine learning (and immediately learn how to best apply them to experiments) has already led to several notable achievements. Li’s expertise and leadership provide a unique educational opportunity for UNC graduate and undergraduate students."link
"The contribution Ad-hoc Pulse Shape Simulation using Cyclic Positional U-Net by Aobo Li, Julieta Gruszko, Brady Bos, Thomas Caldwell, Esteban León, and John Wilkerson was selected for the MLST Paper Award. The award is sponsored by Machine Learning: Science and Technology and acknowledges an outstanding submission to the workshop."link paper link
This course aims at teaching machine learning to experimental physicists, it discussed current AI frontiers in Computer Vision and Natural Language Processing and how they can be applied to comtemporary particle physics experiments. Course website
Advisor: Chritopher Grant
Thesis: The Tao and Zen of neutrinos: neutrinoless double beta decay in KamLAND-Zen 800