Optimisation of passing condition variables of ELOWEN particles with unsupervised machine learning tools

Details

Supervisors

de Wasseige, Gwenhaël

Faculty

Faculté des sciences

Degree label

Master [120] en sciences physiques, à finalité approfondie

Abstract

Neutrinos are a contributor to the recent multimessenger astronomy. Their sources and energy range are presented, and methods that could enhance our understanding of the neutrino energy spectrum are discussed. The observations of neutrinos are done with detectors such as the IceCube neutrino observatory, which is the main source of the data used within this thesis. The data contains noise and the goal of this work is to minimize the background effects of them to retrieve pure data of the astronomical events. A new approach of unsupervised learning is employed within this document, for which dimension reduction techniques are used. The following are employed: PCA, which addresses the eigenvector problem of a sample covariance matrix, and Isomap, t-SNE, and UMAP are nearest neighbour graph-based algorithms. It is shown that PCA emerged as the optimal algorithm, increasing purity by 10.5% in comparison to the three other unsupervised methods. These algorithms help to create an unbiased background and provide an initial signal of neutrinos that can be trained with supervised learning techniques for potentially better results. This new methodology of data processing using cuts after dimension reduction allows one to identify promising paths for future research.