Master Thesis Defense by Panagiotis Tzatzagos

Title: Icy Muons

Subtitle: Utilizing Graph Neural Networks to Align Simulation and Real Data in the IceCube Experiment and Determine Antarctic Ice Anisotropy

Abstract:

Using Graph Neural Networks built with the open-source Python library GraphNet,we trained a machine learning model to distinguish between real and Monte Carlo data. Various sources that enhance the discrepancies between them were identified, and using muon particles, we demonstrated the impact of each source. By mitigating these sources of disparity, the model’s performance achieved an Area Under the Curve (AUC) score of 0.709.

This methodology was also applied to a muon neutrino sample. Although a general trend of reduced model performance was observed, the impact was less pronounced. The worst performance, with an AUC score of 0.809, occurred when only merged SLC pulses were considered.

Additionally, we examined the birefringent effect on light propagation through the ice at the particle level. For this purpose, vector reconstructions of emitted photons for every activated Digital Optical Module (DOM) were determined. By considering the charge ratio of Monte Carlo to data as a function of the azimuth angle, we found that the distribution was non-uniform. While an agreement with the other experiments made by the IceCube collaboration with flashers has been observed, which also noted the strong anisotropy in light propagation at macroscopic scales.