⚛️ high-energy experiments

Exploring ESS $\nu$ SB Near Water Cherenkov Detector Designs Through Graph Neural Network Flavour Identification

This study demonstrates that Graph Neural Network-based flavour identification for the ESS $\nu$ SB near detector remains highly effective even with significantly reduced volume or photomultiplier tube coverage, provided critical signal regions are maintained, thereby offering flexibility in detector design without compromising CP violation measurement goals.

Original authors: J. Aguilar, M. Anastasopoulos, D. Barčot, E. Baussan, A. K. Bhattacharyya, A. Bignami, M. Blennow, M. Bogomilov, B. Bolling, E. Bouquerel, F. Bramati, A. Branca, G. Brunetti, A. Burgman, I. Bustinduy

Published 2026-08-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: J. Aguilar, M. Anastasopoulos, D. Barčot, E. Baussan, A. K. Bhattacharyya, A. Bignami, M. Blennow, M. Bogomilov, B. Bolling, E. Bouquerel, F. Bramati, A. Branca, G. Brunetti, A. Burgman, I. Bustinduy, C. J. Carlile, J. Cederkall, T. W. Choi, S. Choubey, P. Christiansen, I. Christodoulou, E. Cristaldo Morales, P. Cupia, D. D'Ago, H. Danared, J. P. A. M. de Andr, M. Dracos, I. Efthymiopoulos, T. Ekel, M. Eshraqi, G. Fanourakis, A. Farricker, E. Fasoula, T. Fukuda, S. Gago, N. Gazis, Th. Geralis, M. Ghosh, A. Giarnetti, G. Gokbulut, C. Hagner, L. Halić, S. G. Hernández, J. Hiegel, M. Hooft, K. E. Iversen, N. Jachowicz, M. Jensen, I. Karakoulias, E. Kasimi, A. Kayis Topaksu, B. Kliček, K. Kordas, A. Leisos, A. Longhin, M. López, C. Maiano, S. Marangoni, J. García-Marcos, C. Marrelli, D. Meloni, M. Mezzetto, N. Milas, J. L. Muñoz, K. Niewczas, M. Oglakci, T. Ohlsson, M. Olveg, A. Opanasenko, M. Pari, J. Park, D. Patrzalek, G. Petkov, Ch. Petridou, P. Poussot, A Psallidas, F. Pupilli, M. L. Reguera, D. Saiang, E. Salehi, D. Sampsonidis, A. Scanu, C. Schwab, F. Sordo, G. Stavropoulos, M. Stipčević, R. Tarkeshian, F. Terranova, T. Tolba, M. Topp-Mugglestone, E. Trachanas, R. Tsenov, A. Tsirigotis, S. E. Tzamarias, M. Vanderpoorten, G. Vankova-Kirilova, N. Vassilopoulos, S. Vihonen, J. Wurtz, V. Zeter, O. Zormpa

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Deep within the subatomic world, particles known as neutrinos ghost through the universe, rarely interacting with anything they pass. For decades, physicists have puzzled over a fundamental question: do these elusive particles violate a symmetry called charge-parity, or CP? If they do, it could help explain why the universe is made of matter rather than being an empty void of equal parts matter and antimatter. To find the answer, scientists must watch neutrinos change their identity, or "flavour," as they travel. This transformation happens in waves, and the chances of a neutrino switching flavours depend on a specific, hidden angle in the laws of physics. To measure this angle with the precision required, researchers need to catch a massive number of these particles at two specific points in their journey: a spot close to where they are created and another far away. The closer spot acts as a control, showing what the beam looks like before it starts changing, while the distant spot reveals how much it has changed.

The European Spallation Source, a powerful machine in Sweden, is being prepared to fire a beam of these particles with unprecedented intensity. To make the most of this beam, scientists are designing a near detector to sit right next to the source. The plan involves a large tank of ultra-pure water, roughly the size of a small swimming pool, lined with thousands of light-sensitive cameras called photomultiplier tubes. When a neutrino strikes a water molecule inside the tank, it creates a flash of light known as Cherenkov radiation, which these cameras record. The challenge is that the tank is crowded with data, and the signals from different types of neutrinos can look very similar. To sort them out, the team is turning to a type of artificial intelligence called a graph neural network. Unlike traditional computer programs that look for patterns in a grid, these networks treat the detector like a web of connections, learning to identify the unique shape of the light flashes produced by electron neutrinos versus muon neutrinos. The goal is to see if this smart software can do its job even if the detector is made smaller or cheaper, potentially saving millions of dollars without losing the ability to solve the mystery.

In this study, the researchers did not build a new physical tank. Instead, they used powerful computers to simulate the entire experiment, creating millions of virtual neutrino collisions inside a digital version of the proposed water tank. They then trained their artificial intelligence to recognize the difference between electron and muon neutrinos based on the patterns of light these collisions produced. Once the network was trained, they put it to the test by shrinking the virtual tank. They reduced the volume of the water to half, one-quarter, and even one-eighth of the original design, which was about 767 cubic meters. The question was whether the AI would fail when it had less space to work with and fewer light signals to analyze. The results were surprisingly encouraging. Even when the virtual tank was reduced to just one-eighth of its intended size, the artificial intelligence could still distinguish between the two types of neutrinos with high accuracy. While the system did miss a few more events than it did in the full-sized tank, the drop in performance was not a disaster. The researchers found that the AI simply became a bit more cautious, labeling some uncertain events as "ambiguous" rather than guessing wrong. This means that if a smaller detector were built, the experiment could simply run for a longer time to collect enough data to make up for the smaller size.

The team also investigated whether the placement of the light-sensitive cameras mattered. In the original design, the cameras are spread evenly throughout the tank. However, because the neutrino beam comes from a single direction, the light hits the front of the tank harder than the back. The researchers simulated removing cameras from different sections of the tank to see which areas were most critical. They discovered that the location of the cameras mattered significantly. Removing cameras from the back of the tank or the middle sections had very little effect on the AI's ability to identify the particles. However, if they removed cameras from the front end, where the beam enters, or the section of the tank just behind it, the performance dropped noticeably. This suggests that the detector does not need to be uniformly expensive. Engineers could potentially save money by placing fewer cameras in the back and center, while concentrating them where the light is brightest.

The study confirms that the proposed water tank does not need to be a massive, fully instrumented monolith to succeed. By using advanced machine learning, the ESSνSB experiment can consider a more compact and cost-effective design for its near detector. The artificial intelligence proved flexible enough to adapt to smaller volumes and uneven camera coverage, provided that the most critical areas remain well-equipped. This finding offers a practical path forward for the project, allowing scientists to balance the high cost of building a massive detector against the time it would take to gather enough data with a smaller one. The work demonstrates that with the right software, the physical constraints of the hardware can be relaxed, opening the door to a more feasible and efficient way to probe the deepest secrets of the universe.

Technical Summary: Exploring ESS $\nu$ SB Near Water Cherenkov Detector Designs Through Graph Neural Network Flavour Identification

Problem Statement
The ESS $\nu$ SB experiment aims to measure leptonic CP violation with high precision by comparing neutrino beam flavor composition at a near detector with that observed at a far detector (340 km away). Achieving the necessary statistical sensitivity, particularly at the second oscillation maximum, requires high event rates and robust reconstruction of neutrino interactions. The standard design for the near detector involves a 750-ton water Cherenkov (WC) tank instrumented with over 22,000 photomultiplier tubes (PMTs) providing 30% optical coverage. The primary challenge addressed in this work is determining whether the detector design can be optimized for cost and feasibility by reducing the detector volume and/or PMT coverage without compromising the critical requirement of accurate flavor identification (distinguishing electron neutrinos from muon neutrinos).

Methodology
The study utilizes detailed Monte Carlo simulations of charged-current (CC) neutrino interactions generated with GENIE v3.0.6 and detector response simulations using WCSim. The analysis focuses on four detector configurations:

Nominal: 10.96 m length, 9.44 m diameter (767.1 m³).
Half Volume: Dimensions reduced by $\sqrt[3]{2}$ .
1/4 Volume: Dimensions reduced by $\sqrt[3]{4}$ .
1/8 Volume: Dimensions reduced by a factor of 2 (5.48 m length, 4.72 m diameter).

Additionally, the study investigates non-uniform PMT coverage by simulating the removal of 75% of PMT signals in specific detector segments (forward/backward end-caps and barrel slices) without re-running the full particle transport simulations.

The core analysis employs Graph Neural Networks (GNNs), specifically the DynEdge architecture within the GraphNeT framework. Events are represented as graphs where nodes correspond to PMT hits (features: coordinates, time, charge) and edges are defined by K-nearest neighbors in $(x, y, z, t)$ space. The GNN is trained to classify events as electron or muon neutrinos. The study emphasizes the necessity of retraining models for each specific detector configuration to isolate the effects of information loss from model generalization issues.

Key Contributions and Results

Impact of Detector Size: The study demonstrates that GNN-based classification remains effective even when the detector volume is reduced by a factor of eight (1/8 of the nominal size).
- While the Area Under the Curve (AUC) decreases slightly with smaller volumes (from 0.988 for nominal to 0.9779 for 1/8 volume), the True Positive Rate (TPR) at the target False Positive Rate (FPR) of $10^{-3}$ for electron neutrinos drops from 53.5% (nominal) to 39.3% (1/8 volume).
- The authors note that this reduction in efficiency is not a failure of reconstruction but a loss of statistics. Since smaller detectors produce fewer fully contained events (where the charged lepton stops within the tank), the GNN effectively identifies these ambiguous, non-contained events (assigning them scores near 0.5) and rejects them, rather than misidentifying them.
- The performance degradation is described as a trade-off: the loss in efficiency can be compensated by increasing the experiment's runtime.
Impact of PMT Coverage: The study analyzed the effect of reducing PMT coverage in six specific segments (forward end-cap, backward end-cap, and four barrel slices).
- A clear hierarchy of importance emerged. Reducing coverage in the forward end-cap (Segment F) and the barrel segment closest to it (Segment 4) had the most significant negative impact on performance. For electron neutrinos, reducing coverage in Segment F lowered the TPR by 0.219.
- Conversely, reducing coverage in the backward end-cap (Segment B) and central barrel regions had a comparatively small impact.
- This hierarchy correlates with the distribution of detected charge, which is highest in the forward direction of the neutrino beam.
Retraining Necessity: The paper highlights that using a model trained on a nominal configuration for a reduced-coverage or reduced-size detector results in significant performance loss due to a lack of generalization. Retraining the model on the specific configuration recovers most of the performance, indicating that the primary limitation is the loss of information (fewer PMTs or smaller volume) rather than the model's inability to adapt.
Antineutrino Performance: Consistently across all configurations, antineutrino events showed slightly better classification performance than neutrino events, attributed to higher signal collection by PMTs in antineutrino interactions.

Significance and Claims
The paper claims that Graph Neural Networks provide a flexible and robust tool for guiding the design of the ESS $\nu$ SB near detector. The key significance lies in demonstrating that:

Cost-Performance Trade-off: Substantially smaller detector volumes (up to 1/8 of the nominal size) can be utilized without catastrophic loss of flavor identification capability. The resulting efficiency loss can be mitigated by extending the data-taking period.
Optimized Instrumentation: Non-uniform PMT coverage schemes are viable. Instrumentation can be concentrated in high-yield regions (forward end-cap and adjacent barrel) while reducing density in less critical areas (backward end-cap and central barrel) without significantly degrading the physics reach.
Event Ambiguity Handling: The GNN approach naturally handles events where the charged lepton is not fully contained by assigning them ambiguous scores, allowing for the selection of pure samples by rejecting these events rather than misclassifying them.

The authors conclude that these findings support the optimization of the near detector for cost and feasibility while preserving sensitivity to the underlying physics, though they note that future studies are required to determine if these techniques extend to regression tasks (vertex, energy, and direction reconstruction) with acceptable resolution.

More like this