SMuK 2023 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
T: Fachverband Teilchenphysik
T 9: DAQ NN/ML – HW
T 9.1: Talk
Monday, March 20, 2023, 16:30–16:45, HSZ/0301
Implementation of an improved Neural Network for identification of hadronically decaying τ leptons in the ATLAS trigger system for the LHC Run 3 — •Naman Kumar Bhalla, Ö. Oğul Öncel, and Markus Schumacher — Albert-Ludwigs-Universität Freiburg
The ATLAS detector employs a trigger system to reduce the large event rate by saving only interesting events on mass storage for further analyses. This is done via dedicated triggers for each observable physics object. Being the heaviest lepton in the Standard Model of particle physics, the τ lepton is highly unstable, allowing only its decay products to be directly observed. While the electron and muon triggers can be used for the leptonic decays of the τ lepton, separate triggers are necessary to differentiate between hadronically decaying τ leptons (τhad) and jets, which are produced with significantly higher abundance. ATLAS uses a recurrent neural network (RNN) for τhad identification, which exploits various track, cluster and high-level variables as inputs, and returns a single classifier as output. However, it needed to be retuned for operations in the ongoing Run 3 phase of the Large Hadron Collider (LHC) due to upgrades in the detector and the accelerator. Furthermore, new input variables were added to improve the performance of the RNN. An alternative architecture based on Deep Sets was tested in order to have a more efficient usage of computing resources. This talk presents the results of performance studies of the retuned RNN, and a comparison between the two network architectures in terms of efficacy and resource consumption.