Göttingen 2025 – wissenschaftliches Programm

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

EP: Fachverband Extraterrestrische Physik

EP 15: Astrophysics III

EP 15.8: Vortrag

Freitag, 4. April 2025, 15:30–15:45, ZHG101

Precise Reconstruction of Neutrino Event Energy Using Deep Learning — •Severin Magel, Chiara Bellenghi, Elena Manao, and Rasmus Ørsøe for the IceCube collaboration — Technical University of Munich, TUM School of Natural Sciences, Department of Physics, James-Franck-Straße 1, D-85748 Garching bei München, Germany

The first ever 5σ detection of an astrophysical neutrino source has long been chased by neutrino telescopes like IceCube and KM3NeT. Achieving a high statistical significance in detecting these sources is partially limited by the precision of variable reconstructions for the incoming neutrino direction and energy. We investigate the potential of state-of-the-art deep learning architectures like Graph Neural Networks (GNN) and transformers to improve classical algorithms and obtain a more precise neutrino energy prediction. We force the model to recognise general patterns in the detector response by training it on all signatures left in the detector by the different neutrino interaction channels. This pre-trained architecture is then fine-tuned for the reconstruction of specific neutrino events that are eventually used in various analyses not limited to the search for an astrophysical neutrino sources. In this presentation, I will outline the technical challenges and the physics-oriented results from these efforts.

Keywords: Neutrino Astronomy; Event Reconstruction; Deep Learning