Towards the Composition of sub-PeV Cosmic Rays at IceCube

TL;DR

IceCube's hybrid IceTop–InIce detector enables sub-PeV cosmic-ray composition studies by combining surface electromagnetic signals with deep-ice muon information. The authors introduce a background-rejection method based on track speed $beta$ and angle $Psi$ for coincident events and a convolutional neural network that ingests IceTop images and in-ice histograms to jointly estimate energy $E_{pred}$ and discriminate proton vs iron primaries. They report energy resolutions from roughly $42\%$ at 250 TeV to $19\%$ at 7 PeV and binary classification accuracies of $75.7\%$ (sub-PeV) and $84.3\%$ (>$1$ PeV), with biases attributed to training data balance and trigger efficiency. The study demonstrates practical advantages for testing hadronic interactions at sub-PeV energies and outlines steps to mitigate biases, such as balanced MC samples and exploring intermediate masses and models.

Abstract

With the implementation of a low-energy trigger, the surface array of the IceCube Neutrino Observatory is able to record cosmic-ray induced air showers with a primary energy of a few hundred TeV. This extension of the energy range closes the gap between direct and indirect observations of primary cosmic rays and provides the potential to test the validity of hadronic interaction models in the sub-PeV regime. Composition analyses at IceCube highly benefit from its multi-detector design. Combining the measurement of the electromagnetic shower component and low-energy muons at the surface with the response of the in-ice array to the associated high-energy muons improves the directional reconstruction accuracy and opens unique possibilities to extract the primary particle's mass. In this contribution, a new methodical approach for the analysis of these low-energy air showers is presented, including techniques for the identification of coincident background in the in-ice detector and a machine learning model based on convolutional neural networks to determine the elemental composition. The achieved performance in primary mass discrimination and energy reconstruction of air-shower events is discussed.

Figures (5)

• Figure 1: View of a simulated vertical cosmic-ray event with background contamination seen from below the surface - IceTop on top, the in-ice array below. Bubbles represent pulses measured by DOMs. Their color is determined by the time of detection (red $\rightarrow$ blue), their size corresponds to the recorded charge. The in-ice pulses have been split into two independent pulse series (blue and green). Hypothetical tracks connecting IceTop with in-ice (solid lines) are fit to both, as well as pure in-ice fits (dashed lines). The angle $\Psi$ is measured between solid and dashed lines.
• Figure 2: Distributions of hypothetical track speed $\beta$ (left) and the angle $\Psi$ between in-ice track and hypothetical track (right) for 24 hours of data.
• Figure 3: Architecture of the CNN. The top left block shows an example of IceTop input, the $10\times10$ images on top, the 1-dimensional arrays representing the three in-fill stations below. The darkness of each pixel visualizes the charge recorded by the respective DOM. The four image layers are shown in different color shades. The bottom left block gives an example of a 2D-histogram of in-ice pulses with slant depth (20 m binning) on the horizontal and distance from the track (10 m binning) on the vertical axis. All three parts of the network input are fed into separate convolution blocks that are depicted by arrows. After flattening of the intermediate outputs (colored nodes), the two IceTop blocks and the in-ice part get combined in a fully-connected network. The number of hidden nodes is reduced for better visibility. The CNN is trained for the three output nodes to predict the primary energy and to give a proton and iron score - whichever is higher determines the prediction.
• Figure 4: Bias (left, median) and resolution (right, 68-percentile after bias subtraction) of the energy predicted by the CNN. Black points show the entire test dataset, red and blue denote proton and iron primaries, respectively.
• Figure 5: Confusion matrices for the binary classification of proton and iron. The range in $\log_{10}\left(E_\mathrm{MC}/\mathrm{GeV}\right)$ is indicated above. Low statistics at both ends of the energy range and an imbalance in the lowest energies are evident.