Neutrino Fingerprints: Image-Based Encodings of IceCube Events for CNN Direction Reconstruction

Abstract

Determining the trajectory of incoming neutrinos within the IceCube Neutrino Observatory remains a critical challenge in astrophysics. To evaluate reconstruction methods, the public IceCube--Neutrinos in Deep Ice Kaggle competition supplied a dataset comprising 140 million simulated events. We propose a novel approach to this problem by introducing "neutrino fingerprints"—compact $72 \times 72 \times 3$ images where each pixel corresponds to an individual detector. These images encode pulse timing and charge statistics into distinct color channels, effectively converting sparse, irregular pulse data into dense formats ideal for convolutional processing. Utilizing a ResNet18 architecture, our model attained a mean angular error of $1.10$ rad. This result demonstrates that convolutional networks trained on fingerprints can compete with more complex architectures, providing a robust and interpretable baseline for IceCube event reconstruction.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC