The Jiangmen Underground Neutrino Observatory (JUNO) is a scintillation detector, currently under construction, which aims to solve the neutrino mass hierarchy by measuring reactor electron antineutrino energy spectrum with a a resolution of 3%/sqrt(E(MeV)) – the highest ever achieved in a large mass neutrino detector. Several approaches for energy reconstruction are being evaluated on simulated data, and Deep Learning methods have already shown promising results, both in accuracy and in efficiency. In this work, a new Convolutional Neural Network with a rotational invariant architecture is trained on a small dataset of 160k instances, and is fine-tuned to exploit the detector’s spherical symmetry and make use of position and timing data from individual photomultipliers. This approach proves to be insensitive to the presence of dark noise from thermal fluctuations, leading to a (2.45+-0.03)% visual energy resolution at 2 MeV, only slightly higher than the 2.2% expected from theory, with a reconstruction bias well below 1%. However, a simpler Fully Connected Neural Network, replicated from previous work, which uses only integral data and is trained on a larger dataset (750k instances), leads to a slightly better resolution of (2.26+-0.05)% at 2 MeV, while being more sensitive to added noise – proving that there could still be some margin of improvement for more complex methods.
Deep Neural Networks for energy reconstruction of Inverse Beta Decay events in JUNO
Manzali, Francesco
2019/2020
Abstract
The Jiangmen Underground Neutrino Observatory (JUNO) is a scintillation detector, currently under construction, which aims to solve the neutrino mass hierarchy by measuring reactor electron antineutrino energy spectrum with a a resolution of 3%/sqrt(E(MeV)) – the highest ever achieved in a large mass neutrino detector. Several approaches for energy reconstruction are being evaluated on simulated data, and Deep Learning methods have already shown promising results, both in accuracy and in efficiency. In this work, a new Convolutional Neural Network with a rotational invariant architecture is trained on a small dataset of 160k instances, and is fine-tuned to exploit the detector’s spherical symmetry and make use of position and timing data from individual photomultipliers. This approach proves to be insensitive to the presence of dark noise from thermal fluctuations, leading to a (2.45+-0.03)% visual energy resolution at 2 MeV, only slightly higher than the 2.2% expected from theory, with a reconstruction bias well below 1%. However, a simpler Fully Connected Neural Network, replicated from previous work, which uses only integral data and is trained on a larger dataset (750k instances), leads to a slightly better resolution of (2.26+-0.05)% at 2 MeV, while being more sensitive to added noise – proving that there could still be some margin of improvement for more complex methods.File | Dimensione | Formato | |
---|---|---|---|
Manzali_Francesco.pdf
accesso aperto
Dimensione
1.8 MB
Formato
Adobe PDF
|
1.8 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/20.500.12608/24275