Göttingen 2025 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

T: Fachverband Teilchenphysik

T 77: Data, AI, Computing, Electronics VIII (Fast ML, Triggers)

T 77.3: Talk

Thursday, April 3, 2025, 16:45–17:00, VG 2.102

Using Transformer based Graph Neural Networks to Identify Hadronically Decaying Tau Leptons with the ATLAS trigger — •Athul Dev Sudhakar Ponnu and Stan Lai — II. Physikalisches Institut, Georg-August-Universitaet Goettingen.

The increased luminosity at the LHC poses challenges in efficiently selecting interesting events at the Atlas detector. Identifying events containing tau leptons is particularly difficult due to their predominantly hadronic decay, which often mimics light QCD jet signatures. Therefore, effectively discriminating against background jets during the identification of hadronically decaying tau leptons at the trigger level is crucial.

Building on the success of Transformer-based Graph Neural Networks used for offline Tau ID (GNTau) and b-tagging (GN2), this study explores their application to hadronic tau identification at the High Level Trigger (HLT). The online GNTau algorithm exhibits substantial improvements in background rejection compared to existing Deepset-based algorithms, across a wide phase space and variety of processes. After thorough evaluations, the GNTau is set to be deployed at the HLT for the 2025 data-taking period.

Keywords: Tau-lepton; Trigger; Machine learning; Graph-Neural-Network; ATLAS