Master Thesis defense by Peter Andresen

Title: Observing The Universe Using The Antarctic Ice Sheet - Classifying And Reconstructing Neutrinos In Real Data From The IceCube Neutrino Observatory Using Graph Neural Networks


The IceCube Neutrino Observatory (IceCube) uses a cubic kilometre of the Antarctic ice sheet to detect neutrinos. The IceCube neutrino oscillation measurements and potential contributions to multimessenger astronomy heavily depends on accurate and fast classification and reconstruction of neutrinos in real data. GraphNeT, a Graph Neural Network (GNN) open-source Python framework founded at the Niels Bohr Institute, has been shown to improve both classification and reconstruction of low-energy simulated neutrinos, while speeding the process up by orders of magnitude.

This work shows that the method works in actual data as well. The first large neutrino sample is classified using a GNN from "raw" lvl 2 data and compared to a Monte Carlo neutrino selection across a range of reconstructed and calculated variables. The comparison indicates that the sample has a high neutrino purity. The amount of neutrinos in the cleanGNNselection is compared to that of existing methods in the IceCube Oscillation work group using simulated data, suggesting that the GNN method can generate a 50-70% larger sample at a similar purity. Implementing the GNN in IceCube and increasing the neutrino rate would improve existing analyses and potentially open up the possibility of creating early warnings for electromagnetic telescopes using low-energy neutrinos. This work also contributes to a benchmark study trying to improve the GNN performance for high energy track neutrinos. No significant improvements were obtained.