Deep Learning Based Event Reconstruction for the IceCube-Gen2 Radio Detector

TL;DR

This paper demonstrates the use of deep neural networks for precise end-to-end reconstruction of neutrino energy and direction in the IceCube-Gen2 radio detector, including complex event topologies, enhancing UHE neutrino measurements.

Contribution

It introduces a deep learning approach for joint energy and direction reconstruction in a hybrid radio array, including modeling non-Gaussian uncertainties with normalizing flows.

Findings

High-precision neutrino direction and energy predictions achieved.

Effective modeling of complex event topologies, including LPM-affected electron neutrinos.

Potential for detector layout optimization based on reconstruction performance.

Abstract

The planned in-ice radio array of IceCube-Gen2 at the South Pole will provide unprecedented sensitivity to ultra-high-energy (UHE) neutrinos in the EeV range. The ability of the detector to measure the neutrino's energy and direction is of crucial importance. This contribution presents an end-to-end reconstruction of both of these quantities for both detector components of the hybrid radio array ('shallow' and 'deep') using deep neural networks (DNNs). We are able to predict the neutrino's direction and energy precisely for all event topologies, including the electron neutrino charged-current interactions, which are more complex due to the LPM effect. This highlights the advantages of DNNs for modeling the complex correlations in radio detector data, thereby enabling a measurement of the neutrino energy and direction. We discuss how we can use normalizing flows to predict the PDF for each individual event which allows modeling the complex non-Gaussian uncertainty contours of the reconstructed neutrino direction. Finally, we discuss how this work can be used to further optimize the detector layout to improve its reconstruction performance.