MSc thesis defense by Moust Holmes

Title: Neutrino and antineutrino separation using inelasticity in IceCube

Abstract: This work aims to enhance the inelasticity reconstruction using Machine Learning techniques for the IceCube Neutrino Observatory Upgrade in order to statistically separate neutrinos from antineutrinos. While features such as energy and direction of incoming particles have been reconstructed with reasonable precision in IceCube, the inelasticity of the neutrino interaction has remained elusive.

I propose a new loss function based on the beta distribution, designed to reward the model more for accurate predictions it is confident in and penalize it less for challenging, incorrect predictions. I integrate this loss function into the Dynedge model from Graphnet, previously used for IceCube reconstruction, and a novel Transformer model, inspired by the runner-up in the Kaggle competition "IceCube - Neutrinos in Deep Ice", aimed at improving zenith angle reconstruction. My work has the potential to enhance the sensitivity of the IceCube detector, improving our understanding of neutrinos and their properties.