Identification and Denoising of Radio Signals from Cosmic-Ray Air Showers using Convolutional Neural Networks

TL;DR

This work addresses the radio-detection threshold problem for cosmic-ray air showers by deploying two convolutional neural networks, a Classifier and a Denoiser, trained on CoREAS-simulated pulses embedded in real South Pole background. The networks are evaluated on independent test data and then applied to measured data from IceTop’s prototype station, demonstrating a fivefold increase in detected events compared with a traditional SNR-based method, while also reducing false positives and improving pulse power and timing reconstruction at low-to-intermediate SNR. The results are cross-validated against IceTop energy estimators and directional reconstructions, confirming reliable pulse identification and improved measurement precision. The study highlights practical considerations for training on environment-specific background, leverages both polarization channels, and discusses implications for current and future radio-based cosmic-ray experiments.

Abstract

Radio pulses generated by cosmic-ray air showers can be used to reconstruct key properties like the energy and depth of the electromagnetic component of cosmic-ray air showers. Radio detection threshold, influenced by natural and anthropogenic radio background, can be reduced through various techniques. In this work, we demonstrate that convolutional neural networks (CNNs) are an effective way to lower the threshold. We developed two CNNs: a classifier to distinguish radio signal waveforms from background noise and a denoiser to clean contaminated radio signals. Following the training and testing phases, we applied the networks to air-shower data triggered by scintillation detectors of the prototype station for the enhancement of IceTop, IceCube's surface array at the South Pole. Over a four-month period, we identified 554 cosmic-ray events in coincidence with IceTop, approximately five times more compared to a reference method based on a cut on the signal-to-noise ratio. Comparisons with IceTop measurements of the same air showers confirmed that the CNNs reliably identified cosmic-ray radio pulses and outperformed the reference method. Additionally, we find that CNNs reduce the false-positive rate of air-shower candidates and effectively denoise radio waveforms, thereby improving the accuracy of the power and arrival time reconstruction of radio pulses.

Figures (13)

• Figure 1: One of the three SKALA antennas of the prototype surface station at the South Pole (photo from December 2022, courtesy of Roxanne Turcotte Turcotte-Tardif:2022smf).
• Figure 2: Schematic of the neural network architectures for the Classifier and Denoiser. Both networks take two-channel input waveforms and begin with an encoder composed of 1D convolutional layers paired with max-pooling layers. In the Classifier, the encoder is followed by a flattening layer and a dense output layer. In contrast, the Denoiser includes a decoder composed of up-sampling and convolutional layers, followed by a final convolutional output layer.
• Figure 3: Distribution of signal-to-noise ratio of the data set used for training and validation (top) and test (bottom). Shown here is only the distribution for the dataset produced for antenna 1, distributions for other antennas exhibit similar structures, differing only in their mean and slight variations in the tails. The mean background SNR is approximately 14 for both datasets.
• Figure 4: Left: classifier output obtained for both the noisy waveform data (blue) and for the noise-only waveform data (green). Right: true positive and false positive rate as a function of SNR, obtained after applying a threshold cut of 0.5 on the classifier output. Note that this threshold is used only to illustrate performance on the test set. The y‑axis is broken to emphasize low false‑positive fractions. The false-positive curve stops at SNR=100, as almost none of the noise-only waveforms in the data sample exceeds that value (cf. Fig.\ref{['fig:snr_dist']}).
• Figure 5: Example of a denoised waveform, together with the corresponding noisy input and true waveform (CoREAS simulation including the antenna response), shown for the two polarization channels of the antenna. The frequency spectra of the waveforms are displayed in the right panels. The lower subpanels present the difference between the denoised and true waveforms in the time domain, and the ratio between the denoised and true spectra in the frequency domain.