Reconstructing Neutrino Energy using CNNs for GeV Scale IceCube Events

TL;DR

This paper introduces CNN-based methods to enhance energy reconstruction of GeV-scale neutrino events in IceCube, significantly improving resolution and processing speed over traditional likelihood-based techniques.

Contribution

The work presents a novel CNN approach for neutrino energy reconstruction in IceCube, achieving faster processing and better resolution compared to existing methods.

Findings

Approximately 2x improvement in energy resolution.

Reconstruction time reduced to milliseconds per event.

Suitable for large-scale neutrino data analysis.

Abstract

Measurements of neutrinos at and below 10 GeV provide unique constraints of neutrino oscillation parameters as well as probes of potential Non-Standard Interactions (NSI). The IceCube Neutrino Observatory's DeepCore array is designed to detect neutrinos down to GeV energies. IceCube has built the world's largest data set of neutrinos >10 GeV, making searches for NSI a computational challenge. This work describes the use of convolutional neural networks (CNNs) to improve the energy reconstruction resolution and speed of reconstructing O(10 GeV) neutrino events in IceCube. Compared to current likelihood-based methods which take seconds to minutes, the CNN is expected to provide approximately a factor of 2 improvement in energy resolution while reducing the reconstruction time per event to milliseconds, which is essential for processing large datasets.