SMuK 2023 – wissenschaftliches Programm

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

T: Fachverband Teilchenphysik

T 63: ML Methods III

T 63.5: Vortrag

Mittwoch, 22. März 2023, 16:50–17:05, HSZ/0405

Tau neutrino identification with Graph Neural Networks in KM3NeT/ORCA — •Lukas Hennig for the ANTARES-KM3NET-ERLANGEN collaboration — Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU), Erlangen Centre for Astroparticle Physics, Nikolaus-Fiebiger-Straße 2, 91058 Erlangen, Germany

One of the goals of the KM3NeT collaboration is to constrain the PMNS matrix elements associated with the tau neutrino flavour. The data needed to perform this task is taken with KM3NeT/ORCA, a neutrino detector currently under construction in the Mediterranean deep sea. To constrain the matrix elements, one needs to measure the tau neutrino flux produced by atmospheric muon and electron neutrinos oscillating into tau neutrinos. Selecting the tau neutrino events from the full neutrino event dataset is a notoriously difficult task because the final states of tau neutrino interactions look very similar to the final states of muon or electron neutrino events. This classification problem was tackled in my Master's thesis using Graph Neural Networks (GNNs), a type of neural network architecture that showed promising results, e.g., on the related task of jet tagging. This talk will discuss the different methods used to optimise the GNN's performance on this classification task, including a computation-intensive automated hyperparameter search, and present the performance gains achieved by each of these steps and the final performance of the tau event classifier.