Classification of electron and muon neutrino events for the ESSνSB near water Cherenkov detector using Graph Neural Networks
(2025) In Journal of Instrumentation 20(8).- Abstract
In the effort to obtain a precise measurement of leptonic CP-violation with the ESSνSB experiment, accurate and fast reconstruction of detector events plays a pivotal role. In this work, we examine the possibility of replacing the currently proposed likelihood-based reconstruction method with an approach based on Graph Neural Networks (GNNs). As the likelihood-based reconstruction method is reasonably accurate but computationally expensive, one of the benefits of a Machine Learning (ML) based method is enabling fast event reconstruction in the detector development phase, allowing for easier investigation of the effects of changes to the detector design. Focusing on classification of flavour and interaction type in muon and electron... (More)
In the effort to obtain a precise measurement of leptonic CP-violation with the ESSνSB experiment, accurate and fast reconstruction of detector events plays a pivotal role. In this work, we examine the possibility of replacing the currently proposed likelihood-based reconstruction method with an approach based on Graph Neural Networks (GNNs). As the likelihood-based reconstruction method is reasonably accurate but computationally expensive, one of the benefits of a Machine Learning (ML) based method is enabling fast event reconstruction in the detector development phase, allowing for easier investigation of the effects of changes to the detector design. Focusing on classification of flavour and interaction type in muon and electron events and muon- and electron neutrino interaction events, we demonstrate that the GNN reconstructs events with greater accuracy than the likelihood method for events with greater complexity, and with increased speed for all types of events. The GNN flavour classification of neutrino interaction events results in a true positive rate of 85.87 % (57.90 %) for muon (electron) neutrinos, compared to 35.55 % (0.21 %) for the likelihood-based method with identical constraints on the false positive rate, while the reconstruction speed is increased by a factor of 104. Additionally, we investigate the key factors impacting reconstruction performance, and demonstrate how separation of events by pion production using another GNN classifier can benefit flavour classification.
(Less)
- author
- author collaboration
- organization
- publishing date
- 2025-08-01
- type
- Contribution to journal
- publication status
- published
- subject
- keywords
- Analysis and statistical methods, Cherenkov detectors, Neutrino detectors, Performance of High Energy Physics Detectors
- in
- Journal of Instrumentation
- volume
- 20
- issue
- 8
- article number
- P08030
- publisher
- IOP Publishing
- external identifiers
-
- scopus:105014315317
- ISSN
- 1748-0221
- DOI
- 10.1088/1748-0221/20/08/P08030
- language
- English
- LU publication?
- yes
- additional info
- Publisher Copyright: © 2025 The Author(s)
- id
- 9ec717e8-8857-4d56-b9ff-8a5390fc52f0
- date added to LUP
- 2025-09-20 15:08:48
- date last changed
- 2025-09-24 15:29:42
@article{9ec717e8-8857-4d56-b9ff-8a5390fc52f0, abstract = {{<p>In the effort to obtain a precise measurement of leptonic CP-violation with the ESSνSB experiment, accurate and fast reconstruction of detector events plays a pivotal role. In this work, we examine the possibility of replacing the currently proposed likelihood-based reconstruction method with an approach based on Graph Neural Networks (GNNs). As the likelihood-based reconstruction method is reasonably accurate but computationally expensive, one of the benefits of a Machine Learning (ML) based method is enabling fast event reconstruction in the detector development phase, allowing for easier investigation of the effects of changes to the detector design. Focusing on classification of flavour and interaction type in muon and electron events and muon- and electron neutrino interaction events, we demonstrate that the GNN reconstructs events with greater accuracy than the likelihood method for events with greater complexity, and with increased speed for all types of events. The GNN flavour classification of neutrino interaction events results in a true positive rate of 85.87 % (57.90 %) for muon (electron) neutrinos, compared to 35.55 % (0.21 %) for the likelihood-based method with identical constraints on the false positive rate, while the reconstruction speed is increased by a factor of 10<sup>4</sup>. Additionally, we investigate the key factors impacting reconstruction performance, and demonstrate how separation of events by pion production using another GNN classifier can benefit flavour classification.</p>}}, author = {{Aguilar, J. and Zormpa, O. and Bolling, B. and Burgman, A. and Carlile, C. J. and Cederkall, J. and Christiansen, P. and Collins, M. and Danared, H. and Eshraqi, M. and Iversen, K. E. and Lindroos, M. and Park, J.}}, issn = {{1748-0221}}, keywords = {{Analysis and statistical methods; Cherenkov detectors; Neutrino detectors; Performance of High Energy Physics Detectors}}, language = {{eng}}, month = {{08}}, number = {{8}}, publisher = {{IOP Publishing}}, series = {{Journal of Instrumentation}}, title = {{Classification of electron and muon neutrino events for the ESSνSB near water Cherenkov detector using Graph Neural Networks}}, url = {{http://dx.doi.org/10.1088/1748-0221/20/08/P08030}}, doi = {{10.1088/1748-0221/20/08/P08030}}, volume = {{20}}, year = {{2025}}, }