Convolutional Neural Networks for Radio Signals from Cosmic-Ray Air Showers

Abstract

The interaction of high-energy cosmic rays from outer space with the Earth’s atmosphere produces cascades of particles known as extensive air showers (EAS). These EAS emit radio signals, primarily due to the time variation of transverse currents, which result from the deflection of electrons and positrons in the geomagnetic field. These impulse radio signals can be captured by antennas and used to reconstruct EAS properties, such as energy and the depth of shower maximum, X_max. However, natural radio background from galactic sources and man-made radio frequency interference (RFI) significantly limit radio detection capabilities.

To address this challenge, machine learning techniques, specifically convolutional neural networks (CNNs), are utilized in this work to identify and denoise radio signals. The CNNs are trained using simulated radio signals combined with measured radio background noise from antennas at the South Pole. These antennas are part of a surface enhancement planned for IceCube, which involves adding stations with scintillator panels and radio antennas to the existing IceTop array. The results show that CNNs significantly improve the accuracy of pulse power and arrival time reconstruction, particularly at low and intermediate signal-to-noise ratios (SNR). Once trained, the CNNs were applied to search for EAS radio events. The CNNs demonstrated significant improvements over a standard method, which relies on a SNR cut. Using CNNs, approximately five times more events were identified compared to the standard method, and the false event rate was also significantly reduced.