Deep learning reconstruction of neutrino direction, energy, and flavor with complete uncertainty predictions

TL;DR

This paper introduces a deep learning method using conditional normalizing-flows for comprehensive neutrino event reconstruction, providing detailed uncertainty estimates for direction, energy, and flavor from raw signals, aiding future neutrino observatories.

Contribution

It presents the first application of deep neural networks with uncertainty quantification for reconstructing neutrino properties from antenna signals, including asymmetric directional uncertainties.

Findings

Achieved accurate reconstruction of neutrino direction, energy, and flavor.

Provided event-by-event posterior distributions with uncertainty estimates.

Quantified the impact of birefringence on neutrino event reconstruction.

Abstract

With the IceCube-Gen2 observatory under development and RNO-G under construction, the first detection of ultra-high-energy neutrinos is on the horizon making event reconstruction a priority. Here, we present a full reconstruction of the neutrino direction, shower energy, and interaction type (and thereby flavor) from raw antenna signals. We use a deep neural network with conditional normalizing-flows for the reconstruction. This, for the first time, allows for event-by-event predictions of the posterior distribution of all reconstructed properties, in particular, the asymmetric uncertainties of the neutrino direction. The algorithm was applied to an extensive MC dataset of 'shallow' and 'deep' detector components in South Pole ice. We present the reconstruction performance and compare the two station components. For the first time, we quantify the effect of birefringence on event reconstruction.