The IceCube Neutrino Observatory is currently the largest active neutrino detector in the world, aiming to detect high-energy emissions from astrophysical sources. As such it is a cornerstone of experimental astroparticle physics. Particularly, it is an important part of the efforts to pursue a multimessenger approach to astronomical detections, being able to send alerts to collaborating observatories when a significant event is observed. With the intention of staying at the forefront of efficiency in its detection, IceCube is constantly updating their techniques for reconstruction of physical parameters from detected events. Taking this into account, the work presented in this thesis seeks to improve on the reconstruction of energy by implementing Graph Neural Networks, machine learning algorithms designed to recognize patterns from data structured in an irregular manner without loss of information. Graph Neural Networks are a good candidate for the reconstruction of parameters. They have historically shown their ability to perform efficient observations in record times. This fact makes them a good candidate for implementation in systems that reconstruct event parameters during online data taking. The greatest challenge for the implementation of these kinds of algorithms then is the training process, where network architectures need to be trained on proper datasets so they can be deployed without constant supervision. The work on this thesis focuses on applying Graph Neural Networks algorithms for energy reconstruction on the full energy spectrum arriving to the IceCube neutrino Observatory. Configuring the network was a challenge because of the need of implementing a graph representation that is efficient in the \~100 GeV range as well as the \~ 100 PeV range. From the results in this work, we observe that even when trained conservatively, the GNN method is comparatively effective to the currently implemented methods in the RealTime Alert System at IceCube. The precision is noticeably improved in the low energy range, while also showing significant accuracy in the rest of the energy spectrum. These results are obtained while maintaining efficient runtimes across the board, with the possibility of increasing the performance by further management of CPU and GPU resources. we conclude that Graph Neural Networks are effective in the reconstruction of energy in the full energy range of the samples produced from the RealTime Alert System in IceCube, highlighting their possible implementation in online systems as well as possible improvements in the training process.
The IceCube Neutrino Observatory is currently the largest active neutrino detector in the world, aiming to detect high-energy emissions from astrophysical sources. As such it is a cornerstone of experimental astroparticle physics. Particularly, it is an important part of the efforts to pursue a multimessenger approach to astronomical detections, being able to send alerts to collaborating observatories when a significant event is observed. With the intention of staying at the forefront of efficiency in its detection, IceCube is constantly updating their techniques for reconstruction of physical parameters from detected events. Taking this into account, the work presented in this thesis seeks to improve on the reconstruction of energy by implementing Graph Neural Networks, machine learning algorithms designed to recognize patterns from data structured in an irregular manner without loss of information. Graph Neural Networks are a good candidate for the reconstruction of parameters. They have historically shown their ability to perform efficient observations in record times. This fact makes them a good candidate for implementation in systems that reconstruct event parameters during online data taking. The greatest challenge for the implementation of these kinds of algorithms then is the training process, where network architectures need to be trained on proper datasets so they can be deployed without constant supervision. The work on this thesis focuses on applying Graph Neural Networks algorithms for energy reconstruction on the full energy spectrum arriving to the IceCube neutrino Observatory. Configuring the network was a challenge because of the need of implementing a graph representation that is efficient in the \~100 GeV range as well as the \~ 100 PeV range. From the results in this work, we observe that even when trained conservatively, the GNN method is comparatively effective to the currently implemented methods in the RealTime Alert System at IceCube. The precision is noticeably improved in the low energy range, while also showing significant accuracy in the rest of the energy spectrum. These results are obtained while maintaining efficient runtimes across the board, with the possibility of increasing the performance by further management of CPU and GPU resources. we conclude that Graph Neural Networks are effective in the reconstruction of energy in the full energy range of the samples produced from the RealTime Alert System in IceCube, highlighting their possible implementation in online systems as well as possible improvements in the training process.
Implementation of Graph Neural Networks for reconstruction of neutrino energy at the IceCube Real-Time Alert system
COLOMA BORJA, DIEGO ALBERTO
2023/2024
Abstract
The IceCube Neutrino Observatory is currently the largest active neutrino detector in the world, aiming to detect high-energy emissions from astrophysical sources. As such it is a cornerstone of experimental astroparticle physics. Particularly, it is an important part of the efforts to pursue a multimessenger approach to astronomical detections, being able to send alerts to collaborating observatories when a significant event is observed. With the intention of staying at the forefront of efficiency in its detection, IceCube is constantly updating their techniques for reconstruction of physical parameters from detected events. Taking this into account, the work presented in this thesis seeks to improve on the reconstruction of energy by implementing Graph Neural Networks, machine learning algorithms designed to recognize patterns from data structured in an irregular manner without loss of information. Graph Neural Networks are a good candidate for the reconstruction of parameters. They have historically shown their ability to perform efficient observations in record times. This fact makes them a good candidate for implementation in systems that reconstruct event parameters during online data taking. The greatest challenge for the implementation of these kinds of algorithms then is the training process, where network architectures need to be trained on proper datasets so they can be deployed without constant supervision. The work on this thesis focuses on applying Graph Neural Networks algorithms for energy reconstruction on the full energy spectrum arriving to the IceCube neutrino Observatory. Configuring the network was a challenge because of the need of implementing a graph representation that is efficient in the \~100 GeV range as well as the \~ 100 PeV range. From the results in this work, we observe that even when trained conservatively, the GNN method is comparatively effective to the currently implemented methods in the RealTime Alert System at IceCube. The precision is noticeably improved in the low energy range, while also showing significant accuracy in the rest of the energy spectrum. These results are obtained while maintaining efficient runtimes across the board, with the possibility of increasing the performance by further management of CPU and GPU resources. we conclude that Graph Neural Networks are effective in the reconstruction of energy in the full energy range of the samples produced from the RealTime Alert System in IceCube, highlighting their possible implementation in online systems as well as possible improvements in the training process.File | Dimensione | Formato | |
---|---|---|---|
ColomaBorja_DiegoAlberto.pdf
accesso aperto
Dimensione
5.5 MB
Formato
Adobe PDF
|
5.5 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/70104