Classification of Electron and Muon Neutrino Events for the ESS ๐ \nu SB Near Water Cherenkov Detector using Graph Neural Networks
O. Zormpa ยท 2025
Read Classification of Electron and Muon Neutrino Events for the ESS ๐ \nu SB Near Water Cherenkov Detector using Graph Neural Networks by O. Zormpa (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-15247.txt.
Introduction
A large collaboration wrote a paper in two thousand twenty-five on sorting neutrino events with graph neural networks. The setting is a planned European experiment that will fire neutrinos from a particle source toward a huge water detector in Sweden. The near detector watches the particles flash in water, and each flash pattern must be read.
The task is to tell electron neutrino events from muon neutrino events, a split the physics goals depend on. The authors treat each event as a graph, a web of linked points, and let a graph neural network classify it. The paper reports how well that approach works for the water Cherenkov near detector.
Catalog record
| CREATOR | O. Zormpa |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 40 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-15247
Browse the full library ยท About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Classification of Electron and Muon Neutrino Events for the ESS ๐ \nu SB Near Water Cherenkov Detector using Graph Neural Networks O. Zormpa Abstract In the effort to obtain a precise measurement of leptonic CP-violation with the ESS ฮฝ \nu SB 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 events. Addition<ally, 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. 1 Introduction The existence of CP violation in the leptonic sector is among the most interesting open questions in particle physics, and can be investigated through the study of neutrino oscillations [ 1 ] . The potential impact of the CP violating phase \delta_{CP} on the neutrino oscillation probability is almost 3 times larger at the second oscillation maximum than at the first [ 2 ] , but requires 9 times more statistics, as the second maximum is located 3 times further away than the first, and the flux is inversely proportional to the square of the distance. This can be achieved by using a more intense neutrino beam (about one order of magnitude), measuring over a longer period (about 9 times longer), or using more efficient event selection algorithms. The European Spallation Source (ESS) linear accelerator (linac) in Lund, Sweden [ 3 ] can be modified to provide such a high intensity neutrino beam, and the ESS \nu SB is the experiment proposed to measure CP violation using this beam [ 4 , 5 ] . ESS \nu SB will measure the flavour composition of the unoscillated neutrino beam immediately after the beam target at the near detector, which consists of a water Cherenkov (WC) detector, a magnetized scintillator detector called the Super Fine-grained Detector (SFGD), and an emulsion-based detector similar to the NINJA detector employed at J-PARC [ 6 ] . The components of the near detector are illustrated in Figure 1 . The flavour composition of the oscillated neutrino beam will be measured at the far detector, composed of two Cherenkov tanks located 340 km from the target such that the flux will contain both the first and second maximum of muon (anti)neutrino to electron (anti)neutrino oscillations [ 4 ] . By measuring the flavour composition of both the neutrino and antineutrino beam the oscillation probability can be calculated and \delta_{CP} can be estimated. Figure 1 : The three detector components of the near detector complex, the emulsion detector, the SFGD and the water Cherenkov detector. The direction of the neutrino beam is left to right, perpendicularly to the face of the WC tank. Illustration from [ 4 ] . To enable these measurements, accurate reconstruction of the neutrino events is a necessity. While accurate reconstruction methods exist and are implemented in neutrino experiments [ 7 , 8 ] , they are often based on likelihood methods and have limits to their performance and reconstruction speed. The latter becomes particularly important in the development stages of the ESS \nu SB, as having fast event reconstruction methods enable faster exploration of different detector designs and PMT layouts. Recently, Machine Learning (ML) methods and particularly methods based on Graph Neural Networks (GNNs) have proven useful for neutrino event reconstruction [ 9 ] . In the following sections, we explore the possibilities of replacing the current likelihood-based method with one based on a GNN for classification of flavour and interaction type. 2 Methods 2.1 Simulation The GNN models described in the following sections were trained and subsequently benchmarked against the likelihood method applied to two data sets of Monte Carlo (MC) simulated events (used in the ESS \nu SB conceptual design report (CDR) [ 4 ] and [ 10 ] ). The events were simulated using the neutrino interaction vertex generator GENIE [ 11 , 12 ] for neutrino interactions, and the water Cherenkov simulation software WCSim [ 13 ] for particle transport, and the detector response. WCSim is based on Geant4 [ 14 ] and developed by the Hyper-Kamiokande collaboration. For these simulations the Near Detector tank is configured to have a volume of 750 m 3 with a photomultiplier tube (PMT) coverage of 30%. The first dataset contains samples of single charged leptons. These events have been generated homogeneously distributed within the detector tank with an isotropic direction distribution, and uniformly covering kinetic energies from 0 GeV to 1.5 GeV. In the absence of neutrinos and neutrino interactions for this dataset, the flavour of the charged lepton was the target variable for classification. The second dataset contains a uniform distribution of interacting neutrino energies from 0 GeV to 1.5 GeV, with pure water as target material. These events were classified by the flavour of the interacting neutrino and later by interaction type. The simulations contain only muon neutrino and electron neutrino events (muons and electrons for the charged lepton simulations), as the tau neutrinos were omitted from this study 1 1 1 With a beam energy of up to 2 GeV and the tau lepton mass m_{\tau}\approx 1.777 GeV, the tau neutrinos are not produced in significant amounts at the target, and will not partake in CC interaction in the detectors. . Both events with charged current (CC) and neutral current (NC) interactions were considered. 2.2 Reconstruction 2.2.1 Graph Neural Networks The approach developed for this study uses GNNs. GNNs are an ML method based on graph theory, in which data is expressed as graphs of nodes connected by edges, and node information is propagated between the layers of the network through the exchange of information with neighbouring nodes (known as message passing) [ 15 ] . In the case of this study, a neutrino interaction event (or an event containing a single charged lepton) is described by a single graph, with each PMT signal being a node with features such as PMT position, signal arrival time and charge recorded. The graph is completed by connecting the nodes with edges according to a K-nearest neighbours scheme (in x , y , z , t ). To train and apply GNN models to the graphs, the GraphNeT [ 16 ] framework is used. Initially developed by the IceCube collaboration [ 17 , 18 ] , GraphNeT facilitates training of ML models (including GNNs) on data structures and tasks commonly associated with Cherenkov detectors and neutrino telescopes, using the PyTorch [ 19 ] , PyTorch Geometric [ 20 ] and Pytorch Lightning [ 21 ] , libraries as backend. GraphNeT takes as input tables of PMT