Graph neural networks match state-of-the-art performance for positron energy regression and neutron tagging in a Hyper-Kamiokande-like water-Cherenkov detector

The authors present a message-passing graph neural network pipeline that builds graph representations directly from raw photomultiplier hits for low-energy event energy regression and signal classification in water-Cherenkov detectors. Prior deep-learning reconstruction in this domain largely used convolutional neural networks, which require artificially flattening the cylindrical detector geometry and introducing discontinuities; graph neural networks natively accommodate non-Euclidean, irregular, and sparse point-cloud structure without such handling. The method is implemented on WCSim simulations using the HyperK_HybridmPMT_IDonly_Realistic geometry, which includes single PMTs and 19-tube multi-PMT clusters, with dark noise rates set to 8.4 kHz and 1 kHz respectively.

The positron energy regression model achieves a resolution comparable to existing effective-hit-count methods over 3–30 MeV, and its resolution curve does not follow the typical square-root-of-energy behavior. The authors attribute this atypical curve to fully data-driven training: at low energies the model may regress energy by identifying deviations from raw noise, at high energies it may observe clear patterns, and in the middle range it may struggle with irregular Cherenkov-cone signatures at the boundary between noisy and well-defined signals. The comparison is against the effective-number-of-hits method used in Super-Kamiokande and extended to Hyper-Kamiokande; the authors explicitly state this is a reference to published low-energy reconstruction performance rather than a direct benchmark, since detector configurations and reconstruction conditions are not identical.

The neutron tagging classifier correctly rejects 97.47% of background candidates while identifying 93.98% of selected neutron candidates, with asymmetric error rates favorable to a background-dominated search. The authors position this classifier as a rough baseline for neutron tagging in Hyper-Kamiokande-like detectors, intended as a reference for modern deep-learning architectures. Performance is reported via a confusion matrix; during training only candidates containing at least 4 true neutron hits are labeled as true neutron candidates, and candidate pre-selection uses a 10 ns sliding window after time-of-flight correction with a minimum-hit threshold.

The authors argue that GNN-based reconstruction is most naturally integrated with established physics-driven reconstruction methods rather than used as a fully stand-alone approach. This judgment follows from Cherenkov light being a projection of Cherenkov cones onto an irregular and anisotropic geometry, so raw data require further work before a GNN can be fully exploited. The authors suggest using a GNN architecture to validate potential IBD vertices for grid-search algorithms such as BONSAI, or training multiple energy regression models for specific ranges of effective hit count.

The pipeline targets Inverse Beta Decay detection in Hyper-Kamiokande-like water-Cherenkov detectors, covering 3–30 MeV positron energy regression and neutron-capture signal classification. The neutron tagging model is explicitly positioned by the authors as a rough baseline that does not implement fiducial volume filtering or pre-selection algorithms a fully deployed model might include, and its performance applies to candidates satisfying the described selection. The energy regression comparison is framed by the authors as a reference to published low-energy reconstruction performance rather than a direct benchmark. The authors note that GNN-based reconstruction is most naturally integrated with established physics-driven methods, for example to validate potential IBD vertices for grid-search algorithms such as BONSAI, or by training multiple energy regression models for specific ranges of effective hit count.

All results come from WCSim simulations; real detector data, detector response, calibration, and systematic effects are not yet included. The energy regression resolution curve does not follow the typical square-root-of-energy behavior, and the authors offer a speculative, data-driven-training-based interpretation that awaits further verification. Neutron tagging performance depends on the candidate pre-selection threshold and time-of-flight correction; vertex position and timing estimation errors spread true neutron hits by roughly 10 ns in time-of-flight-corrected space, and the impact of this error on classification merits attention. The authors mention incorporating effective hit count as a graph-wide feature, pre-selection criterion, or loss term to inject physics knowledge, but the actual gains of these directions remain to be evaluated. In addition, some formula and figure values are not fully rendered in the fast-parsed text; readers needing to reproduce specific resolution values or candidate counts should consult the original figures and tables.