Advancing Jet Physics in Heavy Ion Collisions with Generative AI

by Yeonju Go (Brookhaven National Laboratory)

Tuesday 18 Nov 2025, 11:00 → 12:00 US/Eastern

Machine learning is rapidly influencing high energy nuclear and particle physics, and many recent studies focus on jet physics where modern methods are beginning to improve analysis quality in significant ways. I will briefly review several applications of machine learning to jet physics, then introduce UVCGAN-S [1], an unsupervised and unpaired image to image translation framework designed for jet background subtraction. Unlike supervised regressions that depend on truth labeled targets and can introduce bias, UVCGAN-S learns to separate signal and underlying event directly from unpaired data. This approach achieves strong improvements in background subtraction performance in the most challenging heavy ion environments and shows robust domain generalization, accurately processing quenched jets generated by JEWEL even though it is trained only on vacuum jets from PYTHIA. Because the method is fully unsupervised, it can be trained directly on experimental data, allowing the model to learn the most faithful structure without any generator driven bias. This work demonstrates the first use of an unsupervised unpaired generative model for jet background subtraction and opens a path to precision analysis in imaging-based measurements.
https://arxiv.org/abs/2510.23717

251118_BNL_seminar_yeonju.pdf