Graph Neural Networks for low-energy event classification & reconstruction in IceCube

IceCube, a cubic-kilometer array of optical sensors built to detect atmospheric and astrophysical neutrinos between 1 GeV PeV, is deployed 1.45 km 2.45 below the surface ice sheet at South Pole. The classification reconstruction events from in-ice detectors play central role in analysis data IceCube. Reconstructing classifying challenge due irregular detector geometry, inhomogeneous scattering absorption light and, 100 GeV, relatively low number signal photons produced per event. To address this challenge, it possible represent IceCube as point cloud graphs use Graph Neural Network (GNN) method. GNN capable distinguishing neutrino cosmic-ray backgrounds, different event types, reconstructing deposited energy, direction interaction vertex. Based on simulation, we provide comparison 1-100 energy range current state-of-the-art maximum likelihood techniques used analyses, including effects known systematic uncertainties. For classification, increases efficiency by 18% fixed false positive rate (FPR), compared methods. Alternatively, offers reduction FPR over factor 8 (to half percent) efficiency. direction, vertex, resolution improves an average 13%-20% 1-30 GeV. GNN, when run GPU, processing nearly double median trigger 2.7 kHz, which opens possibility using online searches for transient events.