Enhancing Events in Neutrino Telescopes through Deep Learning-Driven Super-Resolution

TL;DR

The paper tackles the challenge of sparse photon information in large neutrino telescopes and its impact on angular reconstruction. It introduces a deep-learning-driven super-resolution approach that inserts virtual optical modules and uses a VAE to encode timing, with a UNet++ predicting unseen hits. In an IceCube-like detector, the method yields notable improvements in angular reconstruction (about 0.8 degrees on average, up to 1.3 degrees at lower energies) and runs rapidly in inference. The technique demonstrates robustness to modest optical-property variations and generalizes to water-based detectors, offering a scalable path to enhanced performance for current and future neutrino observatories.

Abstract

Recent discoveries by neutrino telescopes, such as the IceCube Neutrino Observatory, relied extensively on machine learning (ML) tools to infer physical quantities from the raw photon hits detected. Neutrino telescope reconstruction algorithms are limited by the sparse sampling of photons by the optical modules due to the relatively large spacing ($10-100\,{\rm m})$ between them. In this letter, we propose a novel technique that learns photon transport through the detector medium through the use of deep learning-driven super-resolution of data events. These ``improved'' events can then be reconstructed using traditional or ML techniques, resulting in improved resolution. Our strategy arranges additional ``virtual'' optical modules within an existing detector geometry and trains a convolutional neural network to predict the hits on these virtual optical modules. We show that this technique improves the angular reconstruction of muons in a generic ice-based neutrino telescope. Our results readily extend to water-based neutrino telescopes and other event morphologies.

Figures (5)

• Figure 1: Event displays, showing the masked, super-resolved and unmasked event. The unmasked and masked events are obtained from simulation, representing ideal and realistic detector configurations, while the super-resolution network attempts to enhance the masked event into the unmasked. The top plot shows the photon arrival time series from the super-resolution network and the pre-trained VAE on two particular virtual OMs in the super-resolved event.
• Figure 1: Comparison of angular reconstruction performance when varying scattering length. The masked and super-resolved resolutions are obtained from the nominal test dataset, while the dash-dotted median lines are inferred from test datasets with altered scattering lengths.
• Figure 2: Pipeline of the super-resolution framework. The OM timing information is encoded into a 64-parameter latent vector. Neutrino telescope events are arranged into 2D images by string and sensor on a string. For the network inputs, each OM contains 68 features: 3 (3D sensor position) + 1 (number of photon hits) + 64 (timing latent vector).
• Figure 3: Top-down view of a simulated track event, showing both real and virtual OM strings. Color indicates a hit and its timing.
• Figure 4: Log-scale angular resolution. The median lines are drawn in solid color as a function of the true neutrino energy, produced by a baseline SSCNN method. The 20 and 80 percentiles are denoted by the dashed lines and shaded regions.