Speaker
Description
GNTau is a hybrid graph transformer-based neural network algorithm for the identification of the visible decay products of hadronic taus, used by the ATLAS experiment in Run 3 of the LHC. The algorithm is inspired by [Nature Commun. 17 (2026) 541] and supersedes the previous recurrent neural network (RNN) based approach [ATL-PHYS-PUB-2022-044]. Information from reconstructed charged-particle tracks, energy clusters in the calorimeter associated to candidates, and high-level discriminating variables are combined to discriminate hadronic taus from standard jets, originating from heavy-flavour jets (b/c), light quarks (u/d/s) and gluon-initiated jets. This poster presents the performance of the new GNTau algorithm using simulated events, and compares with the previous RNN-based approach. The new algorithm offers substantial gains in background rejection for the same signal selection efficiency, for several classes of analysis including measurements of Higgs boson (pair) production and couplings, and BSM searches involving 3rd generation fermions.
| Track | Performance and Tools |
|---|