Neutrino experiments produce rich, high-dimensional datasets that are increasingly valuable for AI/ML research and for the broader physics community. This portal aggregates links to publicly released data from major neutrino experiments and open benchmark simulation datasets — each experiment hosts its own data, and this site provides organized pointers to those original sources.
This portal was seeded by a presentation at the Open (and Closed) Data in the Age of AI Workshop, Chicago, April 2026 (slides).
Experiments
MicroBooNE Completed
Short-baseline LArTPC • 0.1–2 GeV • Fermilab
Liquid argon TPC overlay samples combining simulated neutrino interactions with real cosmic-ray backgrounds. HDF5 and art/ROOT formats.
MINERvA Completed
Fine-grained scintillator • 1–10 GeV • Fermilab
Analysis-level event samples with full systematic uncertainties, reconstructed variables, and the MINERvA Analysis Toolkit.
Daya Bay Completed
Reactor antineutrino • 2–8 MeV • China
Full 2011–2020 dataset: 3158 days, 5.55 million nGd IBD events. Tabular event data in HDF5, NPZ, and ROOT formats on Zenodo.
IceCube Running
Cherenkov telescope • GeV–PeV • South Pole
Event-level data releases covering atmospheric and astrophysical neutrinos. HDF5 and CSV formats via the IceCube data releases page.
NOvA Running
Long-baseline accelerator • 0.5–5 GeV • Fermilab / Ash River
Publication-level data releases including oscillation and cross-section products. Joint analysis with T2K published in Nature (2025).
T2K Running
Long-baseline accelerator • 0.5–5 GeV • J-PARC / Kamioka
Fit outputs, histograms, and ROOT files on Zenodo. Joint oscillation analysis with NOvA uses containerized likelihoods (ARIA / MaCh3).
DUNE Future / In Construction
Long-baseline LArTPC • 1–10 GeV • Fermilab / SURF
Next-generation experiment under construction. Open data policy is under active development within the collaboration.
Benchmark Simulation Datasets
PILArNet
LArTPC simulation • 300k events • numpy / HDF5 (OSF)
Model-independent Geant4 simulation for LArTPC reconstruction and AI/ML benchmarking. Sparse 2D/3D voxel representations; five particle categories; ground-truth labels for supervised learning.
NuBench
Cherenkov telescope simulation • Parquet / SQLite (ERDA)
Benchmark dataset for Cherenkov neutrino telescope reconstruction. Graph, point-cloud, and tabular representations. Reference GraphNeT implementations included.
Data Summary
Based on the survey presented at the Open (and Closed) Data in the Age of AI Workshop, Chicago, April 2026 (slides).
| Experiment | (1) Data | (2) Metadata | (3) Analysis Data | (4) Tools |
|---|---|---|---|---|
| MicroBooNE | 2D wire-time images, event data — HDF5, artroot — Zenodo | dataset docs — Fermilab site | reduced datasets — Zenodo | Python data loaders, example Jupyter notebooks — GitHub (OpenSamples) |
| MINERvA | tabular ntuples — ROOT — Fermilab server | docs — Fermilab site | reco vars + systematics — Fermilab server | ROOT macros, systematics weighting scripts — GitHub (MinervaExpt) |
| Daya Bay | tabular event data — HDF5, NPZ, ROOT — Zenodo | metadata — Zenodo | analysis dataset (IBD + inputs) — Zenodo | Python analysis package (dayabay-model), fit scripts — GitHub |
| NOvA | histograms, tables — ROOT/plots — Fermilab publicdocs | docs — Fermilab site | oscillation / xsec products — Fermilab publicdocs | — |
| T2K | fit outputs, histograms — ROOT — Zenodo | metadata — Zenodo | fit results — Zenodo | ROOT macros (plot extraction, validation) — Zenodo |
| IceCube | event-level data (time, charge, position) — HDF5/CSV — IceCube site | docs — IceCube site | event catalogs — IceCube site | — |
| PILArNet | 2D/3D sparse images, point clouds — numpy/HDF5 — OSF | dataset paper — arXiv | labeled dataset — OSF | dataset interface scripts — project repo |
| NuBench | detector hits (graph/point-cloud, tabular) — Parquet, SQLite — ERDA | docs — arXiv + GitHub | benchmark tasks — GitHub | data loaders, training scripts, GraphNeT-based models — GitHub |