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

Cosmic-ray air showers undergo geomagnetic deflection and Askaryan emission to produce radio signals which can be measured by antennas on ground, e.g., at a prototype surface station at the IceCube Neutrino Observatory. The radio antennas at this station also detect radio signals from human activity, thermal noise, and the continuous Galactic background which makes identifying the cosmic-ray air shower radio signals challenging. This project makes use of convolutional neural networks (CNNs) in order to (1) determine when a radio waveform includes a cosmic-ray radio signal and (2) remove background noise from these waveforms to reveal the pure signal. We create datasets to train these models by combining simulated radio signals from the CoREAS Monte Carlo code with background noise waveforms from the prototype surface station at the South Pole. We experiment with different network structures and datasets in order to improve the detection threshold of radio experiments and the accuracy of the reconstructed amplitude and arrival time of the cosmic-ray radio pulses.