Master Thesis defense by Andreas Mosgaard Jørgensen

Title: 

Enhancing state of the art IceCube neutrino selection - Adjustment of early selection cuts to match Graph Neural Network classification technique

Abstract:

Embedded in the Antarctic ice sheet lies The IceCube Neutrino Observatory spanning one cubic kilometer of ice. Among the many use cases of such an extraordinary experiment, this thesis concerns the IceCube collaborations research of neutrino oscillations and high energy pointing. The reconstruction done in these analyses needs to be fast and accurate, when dealing with a large quantity of data. The machine learning method called Graph Neural Networks (GNNs) have been shown to be a great choice, and thus it is used in this thesis. More specifically the GNN is an open-source framework called GraphNeT developed in Python by Rasmus Ørsoe at the Niels Bohr Institute. Such methods require a lot of data to become accurate both during training and when simulating events using Monte Carlo methods. Therefore, the main analysis is on making a filter that preserves the highest number of neutrino events, while filtering away muon and noise events and ensuring data Monte Carlo agreement.

The result of this was that by slightly loosening the cut, 8 % neutrinos are gained in Monte Carlo, while 13 % are gained in data of which the far majority is expected to be neutrinos. Furthermore, as the IceCube detector will undergo upgrades in the coming years this work can inspire and lay the grounds of new selection algorithms. Additionally, this thesis will also contain an analysis on high energy neutrinos, where the objective is to enhance the resolution of high energy reconstruction of events supporting the IceCube collaborations contributions to multi-messenger astronomy. The data set contains events from the northern direction, which has the advantage of the bulk of the entire Earth stopping everything except neutrinos making the data purity high. Competing with the existing reconstruction algorithm called SplineMPE was difficult and as a result no improvements were obtained and the project moved on to the before mentioned. However, in the outlook newly found promising methods are described, and the future of that field is thus bright.