Göttingen 2025 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

T: Fachverband Teilchenphysik

T 11: Data, AI, Computing, Electronics I (Statistical Methods, Applications)

T 11.1: Vortrag

Montag, 31. März 2025, 16:45–17:00, VG 2.101

Performance measurements of Tau identification tools in ATLAS — •David Dahiya, Christian Schmidt, Arno Straessner, and Asma Hadef — Technische Universität Dresden

Tau leptons are fundamental in a variety of Standard Model and Beyond Standard Model processes currently being studied at the LHC. Their identification is crucial for exploring new physics, as they often serve as key signatures in searches for novel particles and interactions. This work focuses on improving Tau Lepton Identification (TauID) by conducting performance measurements and comparing different TauID models. Current tau identification approaches utilize Recurrent Neural Networks (RNNs), which are trained on a combination of tracks, clusters, and high-level variables to produce a predictive score for each tau candidate. However, recent advancements in machine learning introduce Graph Neural Networks (GNNs) as a promising alternative. GNNs are trained on jet and track-level variables and exploit graph-based attributes to predict features such as vertex position, jet flavor, and track origin, potentially offering a more robust and detailed analysis. This study provides a comparison of the performance of RNN-based and GNN-based models to evaluate the impact of GNNs' added complexity on tau identification. Additionally, GNNs are used to compare and evaluate tau fake factors based on a control data set using the latest Run 3 data.

Keywords: Tau identification; Graph Neural Networks; RNN; Fake Factor