Convolutional Neural Networks For Radio Signals From Cosmic-ray Air Showers

Extensive air showers produced by the interaction of cosmic rays with the atmosphere emit radio signals, primarily due to the deflection of electrons and positrons under the influence of Earth's magnetic field. These impulsive radio emissions have proven useful for studying air-shower properties such as energy and the depth of shower maximum, a parameter related to the mass of the primary cosmic ray. However, the detection threshold for radio emissions is significantly affected by the presence of background noise from both natural sources, such as galactic noise, and anthropogenic sources.

In this work, convolutional neural networks (CNNs) were employed to reduce the impact of background noise on radio measurements. The networks were trained using simulated signals from CoREAS and measured background from radio antennas installed at the South Pole, as part of the enhancement plans for the IceTop detector. The trained networks were then used to identify air-shower events, and their performance was compared to a traditional method based on the signal-to-noise ratio. Over a four-month search period, the CNNs identified five times more events than the traditional method, with a substantially reduced false positive rate. The additional events detected using CNNs were primarily at lower energies, demonstrating that these networks can effectively lower the detection threshold for radio measurements of air showers.