Machine Learning for Classification and Denoising of Cosmic-Ray Radio Signals

Geomagnetic deflection of oppositely charged particles in cosmic-ray air showers produces radio signals that can be detected by antennas of a prototype surface station at the IceCube Neutrino Observatory. However, these antennas also measure thermal noise, human-made radio frequency interferences, and the continuous Galactic background which makes it challenging to separate cosmic-ray radio signals from this background. In this work, we employ convolutional neural networks (CNNs) with two goals in mind: (1) the identification of waveforms that contain cosmic-ray radio signals as opposed to waveforms that contain only background and (2) removing backgrounds from the waveform to reproduce the pure signal. The datasets required to train these models include signal simulations from the CoREAS Monte Carlo code as well as noise waveforms measured with the antennas at the IceCube Neutrino Observatory. Both signal and background traces are filtered to a frequency band of 100-350 MHz before training and analysis. We aim to use these machine-learning models to improve the detection threshold of radio experiments and to improve the accuracy of the arrival time and amplitude of detected cosmic-ray radio pulses.