Elemental Composition of Cosmic Rays : Analysis of IceCube data using Graph Neural Networks
Cosmic rays (CRs) are high-energy ionized nuclei that emanate from astrophysical sources and regularly bombard Earth's atmosphere. The IceCube Neutrino Observatory (ICNO), a cubic-kilometer observatory embedded in South-Pole Antarctic ice, is very well suited to investigate CR in the energy reg...
| Main Author: | |
|---|---|
| Other Authors: | , |
| Format: | Doctoral or Postdoctoral Thesis |
| Language: | English |
| Published: |
KIT-Bibliothek, Karlsruhe
2024
|
| Subjects: | |
| Online Access: | https://publikationen.bibliothek.kit.edu/1000169558 https://publikationen.bibliothek.kit.edu/1000169558/152525461 https://doi.org/10.5445/IR/1000169558 |
| Summary: | Cosmic rays (CRs) are high-energy ionized nuclei that emanate from astrophysical sources and regularly bombard Earth's atmosphere. The IceCube Neutrino Observatory (ICNO), a cubic-kilometer observatory embedded in South-Pole Antarctic ice, is very well suited to investigate CR in the energy regime where the transition from Galactic to the extra-galactic origin of CR occurs. This work delves into the analysis of the elemental composition at ICNO using the footprint of extensive air showers (EAS) detected at IceTop (IT) (the surface detector component of IceCube) in conjunction with the in-ice component. Using dedicated Monte Carlo simulations, the work leverages energy deposits by TeV-muons (within EASs) in IceCube, to devise various CR composition-sensitive physics observables. In addition to providing composition-sensitivity, the observables also provide the possibility for testing internal consistencies in phenomenological models which describe hadronic interactions of CR with atmospheric-nuclei.\par An extended part of this work focuses on developing a Graph Neural Network (GNN)-based approach, informed by the physics of EASs. This approach utilizes the observed footprint of EAS at both IceTop and IceCube to estimate logarithmic mass for each individual EAS. As a first-of-its-kind endeavor, the GNN architecture incorporates multiple EAS-physics-inspired inductive-biases to obtain the mass estimate in an efficient manner. Moreover, in order to provide enhanced stability to the network the work also adapts ideas from other leading fields in Deep Learning. In addition to the physics-informed application of the GNN, the network also benefits from the composition-sensitive observables developed in this work, along with other shower observables, to provide enhanced composition sensitivity. A gradient-boosted decision tree-based approach is used for the energy estimate. The different Machine Learning (ML) methods are concurrently utilized to estimate the elemental composition of CRs. The methods developed on ... |
|---|