- This event has passed.
Speaker: Nicole Hartman (Technical University of Munich) – Title: b-tagging unlocks the Higgs Potential — deep learning for jet flavour tagging on ATLAS
July 3 @ 12:00 pm - 1:00 pm
Research Progress Meeting
Date: July 3, 2025
Time: noon – 1:00 pm [note special time]
Location: Sessler Conference Room- 50A-5132 [In-Person and HYBRID]
Speaker: Nicole Hartman (Technical University of Munich)
Title: b-tagging unlocks the Higgs Potential — deep learning for jet flavour tagging on ATLAS
Abstract:
Electroweak symmetry breaking involves a Higgs potential that generates mass for the weak bosons. In the Standard Model, this potential has a quartic form—but this is merely an ansatz that has yet to be experimentally verified. A first experimental probe of the Higgs potential could come from measuring the simultaneous production of two Higgs bosons (“di-Higgs”). This is an exceptionally rare process, expected to require another decade of data collection to discover using current analysis methods. The most probable decay mode of a Higgs boson is into a pair of b-quarks, which in turn produce b-jets in the detector. The channels driving di-Higgs sensitivity therefore all include b-jets in the final state.
Machine learning is transforming many areas of life, and particle physics is no exception. Because of the complexity of jets at the LHC, b-jet classification (or flavour tagging) has become a leading application of deep learning in the field. This seminar presents the state-of-the-art in flavour tagging: a transformer model, the same backbone architecture used in ChatGPT. These modern transformers deliver an impressive factor-of-four improvement in performance compared to earlier b-taggers based on recurrent neural networks. In addition, transformers are more data-efficient, allowing us to benefit from a tenfold increase in training statistics. Importantly, these gains in simulation also translate to actual LHC data.
Transformers are general-purpose architectures that can be integrated at multiple stages of the analysis pipeline. As training datasets continue to grow, we can begin to view jet tagging as a foundation model for LHC physics—one that can be customized or “fine-tuned” for specific physics goals. We conclude by highlighting how an end-to-end optimizable analysis can fine-tune a jet-tagger for an HH physics search, helping to further advance our di-Higgs program and quest to understand the Higgs potential.
Join Zoom Meeting
https://lbnl.zoom.us/j/98881011162?pwd=1kC2cSwdAjvJJUer0ymPMklwe8NFOE.1
Meeting ID: 988 8101 1162
Passcode: 696672