Master defense by Yasaswy Vishnubhatla Sri Vahn

Title: Looking at the Universe with a New Pair of Eyes

Exploring the use of GNN’s to veto atmospheric neutrinos in IceCube

Abstract: The IceCube Neutrino Observatory is designed to study neutrinos of energies different by several orders of magnitude. The predominance of atmospheric neutrino flux over that of astrophysical neutrinos below the energies of 10 TeV has been a problem in the context of our efforts of wanting to use the astrophysical neutrinos as messengers from the distant universe. After introducing some of the relevant background, this work explores the use of Graph Neural Networks (GNN’s) in trying to identify atmospheric neutrinos (to veto them), on the 1% burnsample, and on a few Monte Carlo (MC) simulations, using the DynEdge model from the GNN package GraphNet, developed specifically for neutrino detectors.

This work introduces the novel idea of "Atmospheric Tagging" (which involves identifying atmospheric neutrinos based on the presence of their associated atmospheric showers), and employs it first on the burnsample without success due to low data, and then on the MC, where attempts to isolate the neutrino signal (from that of the shower) are made, followed by a comparison with a new atmospheric neutrino (CORSIKA showers + NuGen neutrino) simulation, where the idea is shown to work well. Then, using this "veto" simulation separately to train binary classification models shows promising results with both electron and muon atmospheric neutrino events being identified well (ROC-AUC >0.94 and >0.88 respectively) under 100 GeV.