Fast Low Energy Reconstruction using Convolutional Neural Networks

TL;DR

This work addresses the challenge of reconstructing sub-100 GeV neutrino interactions in IceCube-DeepCore for atmospheric oscillations. It develops five CNNs with a two-branch input architecture to estimate energy, direction, and vertex, plus two classifiers for track-like vs cascade-like events and for atmospheric muon background, trained on GENIE and MuonGun MC. The CNNs achieve competitive or improved energy and zenith resolution and have GPU-accelerated inference enabling large-scale analyses, with runtimes far faster than traditional RETRO methods while delivering comparable performance. These reconstructions support higher effective event yields and feed into precise $\nu_\mu$ disappearance measurements, with ongoing work exploring azimuth reconstruction, uncertainty estimation, and graph neural network approaches for further gains.

Abstract

IceCube is a Cherenkov detector instrumenting over a cubic kilometer of glacial ice deep under the surface of the South Pole. The DeepCore sub-detector lowers the detection energy threshold to a few GeV, enabling the precise measurements of neutrino oscillation parameters with atmospheric neutrinos. The reconstruction of neutrino interactions inside the detector is essential in studying neutrino oscillations. It is particularly challenging to reconstruct sub-100 GeV events with the IceCube detectors due to the relatively sparse detection units and detection medium. Convolutional neural networks (CNNs) are broadly used in physics experiments for both classification and regression purposes. This paper discusses the CNNs developed and employed for the latest IceCube-DeepCore oscillation measurements. These CNNs estimate various properties of the detected neutrinos, such as their energy, direction of arrival, interaction vertex position, flavor-related signature, and are also used for background classification.

Figures (10)

• Figure 1: The upper panel shows a top view of the IceCube detector. DeepCore strings are indicated by red circles and IceCube strings by green circles. The DeepCore detector consists of both DeepCore strings and IceCube strings located within the dashed green hexagonal circle. In this work, the CNNs use only the strings within the orange hexagon. The lower panel shows a side view of the detector. The green shading highlights the DeepCore DOMs and the red shading highlights the veto DOMs.
• Figure 2: Two-branch architecture of the CNN, including a legend for names of layers on the right. Eight DeepCore and 19 IceCube strings with 60 DOMs on each string and 5 summarized variables from each DOM from the digitized pulses are fed into the two sub-networks correspondingly. The size of the kernel is listed in the graph and spans only along the $z$-depth axis.
• Figure 3: Distribution of $\nu_{\mu}$ CC events of reconstructed neutrino zenith angle on the left and neutrino energy on the right plotted against true neutrino energy on the top and true neutrino arrival direction ($\cos(\theta^{zenith})$) on the bottom, with the median as solid curves, 68% and 90% quintiles as dark shaded and light shaded for CNN, or dashed and dotted orange curves for SANTA/LEERA.
• Figure 4: Distribution of $\nu_{e}$ CC events of reconstructed neutrino zenith angle on the left and neutrino energy on the right plotted against true neutrino energy on the top and true neutrino arrival direction ($\cos(\theta_{zenith})$) on the bottom, with the median as solid curves, 68% and 90% quantiles as dark shaded and light shaded for CNN, or dashed and dotted orange curves for SANTA/LEERA.
• Figure 5: Distributions of events classified by CNN as track-like or cascade-like, where most track-like corresponds to 1, with $\nu_\mu$ CC events in blue and other types of neutrino interactions in red, where filled histograms show CNN reconstruction and stepped histograms show BDT$_{\rm{track}}$.