NuBench: An Open Benchmark for Deep Learning-Based Event Reconstruction in Neutrino Telescopes

TL;DR

NuBench addresses the need for open benchmarks in deep learning–based event reconstruction for neutrino telescopes. It introduces seven large-scale simulated datasets across six detector geometries, spanning energies from $10~\mathrm{GeV}$ to $100~\mathrm{TeV}$ and including both $\nu_\mu^{\mathrm{CC}}$ and $\nu_\mu^{\mathrm{NC}}$ interactions, with pulse- and event-level ground truth for five reconstruction targets. By benchmarking four architectures—ParticleNeT, DynEdge, GRIT, and DeepIce—on energy, direction, track/cascade classification, vertex, and inelasticity, the study shows cross-geometry generalization is feasible but performance remains task- and energy-dependent, with DeepIce often excelling in direction and DynEdge in vertex reconstruction. The results underscore the importance of detector density for containment-sensitive tasks and illustrate that no single architecture dominates across all settings. Making the datasets, predictions, and model artifacts openly available fosters reproducibility and collaboration in the neutrino-telescope community.

Abstract

Neutrino telescopes are large-scale detectors designed to observe Cherenkov radiation produced from neutrino interactions in water or ice. They exist to identify extraterrestrial neutrino sources and to probe fundamental questions pertaining to the elusive neutrino itself. A central challenge common across neutrino telescopes is to solve a series of inverse problems known as event reconstruction, which seeks to resolve properties of the incident neutrino, based on the detected Cherenkov light. In recent times, significant efforts have been made in adapting advances from deep learning research to event reconstruction, as such techniques provide several benefits over traditional methods. While a large degree of similarity in reconstruction needs and low-level data exists, cross-experimental collaboration has been hindered by a lack of diverse open-source datasets for comparing methods. We present NuBench, an open benchmark for deep learning-based event reconstruction in neutrino telescopes. NuBench comprises seven large-scale simulated datasets containing nearly 130 million charged- and neutral-current muon-neutrino interactions spanning 10 GeV to 100 TeV, generated across six detector geometries inspired by existing and proposed experiments. These datasets provide pulse- and event-level information suitable for developing and comparing machine-learning reconstruction methods in both water and ice environments. Using NuBench, we evaluate four reconstruction algorithms - ParticleNeT and DynEdge, both actively used within the KM3NeT and IceCube collaborations, respectively, along with GRIT and DeepIce - on up to five core tasks: energy and direction reconstruction, topology classification, interaction vertex prediction, and inelasticity estimation.

Figures (21)

• Figure 1: Illustrations of the two neutrino event morphologies: Cascade (left) and Track (right). Each dot represents an optical module, and the color indicates the arrival time of the pulses, ranging from red (early) to cyan (late). The size of the dots is adjusted to be proportional to the observed charge. Grey dots represent modules that did not observe pulses during the interaction.
• Figure 2: Illustrations of starting- and stopping tracks from our datasets. Left) A 80TeV stopping track. Right) A 44TeV starting track.
• Figure 3: Top-down view of the 6 different detector geometries: Flower S, Flower L, Flower XL, Triangle, Cluster, and Hexagon. Approximate length scales are annotated for comparison.
• Figure 4: Side view of the 6 different detector geometries: Flower S, Flower L, Flower XL, Triangle, Cluster, and Hexagon. Approximate length scales are annotated for comparison.
• Figure 5: Illustration of the photon merging procedure applied to transform photons from PROMETHEUS simulation into pulses of Cherenkov radiation. The merging window is applied starting from the first photon and is set to the TTS assigned to each dataset. The associated charge of the pulses is set to the number of photons found within each merging window.