NuBench: open benchmark datasets for deep-learning event reconstruction in neutrino telescopes
NuBench: 7 simulated neutrino-telescope datasets (~130 M νμ events, 6 geometries) for ML reconstruction
Seven large-scale simulated datasets with nearly 130 million charged- and neutral-current muon-neutrino interactions from 10 GeV to 100 TeV, generated for six detector geometries inspired by existing and proposed neutrino telescopes in water and ice ("Cluster", "Flower S/L/XL", "Hexagon", "Hexagon Ice LE", "Triangle"). Pulse- and event-level information for energy and direction reconstruction, topology classification, vertex and inelasticity estimation, with train/test selections, the predictions of four reference models (ParticleNeT, DynEdge, GRIT, DeepIce) and their trained model artifacts. Compatible with the GraphNeT dataset classes.
Data card 4/11 sections
| Summary | Cross-experiment ML benchmark for neutrino-telescope reconstruction. |
|---|---|
| Intended use | Training and comparing reconstruction / classification models across detector geometries. |
| Out-of-scope use | Physics results: geometries are inspired by, not identical to, real detectors; no detector data. |
| Known limitations | No DOI or licence; files hosted via ERDA share links listed in the GitHub README. |
Get it
| Location | Role | Host | Formats | Size |
|---|---|---|---|---|
| catalogue and download links | documentation | GitHub | ||
| ERDA (University of Copenhagen): per-dataset SQLite / Parquet, predictions and model artifacts (tar.gz) | primary | Institutional / collaboration web page | Parquet, Archive (tar/zip) of mixed content |
Metadata
| Creators | Rasmus F. Ørsøe; Stephan Meighen-Berger; Jeffrey Lazar; Jorge Prado; Ivan Mozun-Mateo; Aske Rosted; Philip Weigel; Arturo Llorente Anaya |
|---|---|
| Owner | graphnet-team |
| Published | 2025-11 |
| Licence | No licence stated |
| Access | Open |
| Channels | Charged current (CC), Neutral current (NC) |
| Flavours | numu, numubar |
| Detector technology | Ice Cherenkov, Open-water Cherenkov (sea/lake) |
| Energy range | 10–100000 GeV (true neutrino energy) |
| Data levels | Monte Carlo |
| Observables | pulses, event-level truth |
| Events | 130000000 |
| Covariance / systematics / MC | / / yes |
Uses
- Software GraphNeT
Know of another paper or code that uses this, or an input it depends on? Tell us.
Papers
- Ørsøe et al., NuBench: An Open Benchmark for Deep Learning-Based Event Reconstruction in Neutrino Telescopes (prepared for JINST) · arXiv:2511.13111 (describes)
How to cite
- arXiv:2511.13111
Related
- derived from Prometheus: An open-source neutrino telescope simulation — simulation cited in the catalogue README
Curation
| Status | Curated (candidate: harvested, not reviewed · curated: reviewed metadata · verified: resource downloaded and checked against its paper) |
|---|---|
| Added | 2026-10-23 |
| Curator notes | • the per-dataset ERDA share links are in the README; n_events is the approximate total from the abstract |