Master Thesis defense by Andreas M. Hermansen

Title: Neutrino Learning: Exploring Capabilities and Limitations of Graph Neural Networks in IceCube

Abstract:

The IceCube Neutrino Observatory is located at the geographic South Pole. The experiment is capable of detecting neutrinos using an array of digital-optical-modules(DOMs) located deep below the surface ice. A particularly dense region of the IceCube array is called DeepCore, which is essential for low energy oscillation studies. However, any neutrino based research, first needs to transform the signals from the DOMs into meaningful physical quantities. It has recently been shown that Graph Neural Networks (GNN) are capable of solving these tasks above the level of more traditional maximum likehood approaches.

This work shows that it is possible to achieve almost an order of magnitude faster training, by simply restricting the number of signals passed to the GNN at no measurable loss in performance. It also shows that previous discrepancies between the true and predicted zenith angle are a result of the choice of the loss function, and can be somewhat circumvented reducing the Earth Mover’s Distance from 0.17 to 0.06. It also shows that the current choice of a Euclidean based k-nearest neighbour (KNN) algorithm to construct the graphs, outperforms a KNN-algorithm based on the Minkowski space-time metric.

Furthermore this work develops a per-pulse importance function, which shows that GNN based reconstructions accurately identify pulses from the DeepCore as the most important. Lastly this thesis develops a self-supervised-learning task, which allows a further increase in training speed of 20%.