Cosmic-ray air showers emit radio waves that can be used to measure the properties of cosmic-ray
primary particles. The radio detection technique presents several advantages, such as low cost
and year-round duty cycle as well as the ability to provide high sensitivity to Xmax and energy
estimation with minimal theoretical uncertainties, making it a promising tool for studying cosmic
rays at the highest energies. However, the primary limitation of radio detection is the irreducible
background from various sources that obscure the impulsive signals generated by air showers. To
address this issue, we investigated the use of Convolutional Neural Networks (CNNs), trained on
CoREAS simulations and radio backgrounds measured by a prototype station at the South Pole.
We developed two different CNNs: a Classifier that distinguishes between cosmic ray event radio
signals and pure background waveforms, and a Denoiser that mitigates background noise to recover
the underlying cosmic-ray signal. After training the networks we apply them to the air-shower data
to search for radio events. With two months data, we were able to identify 51 candidate events.
The event’s arrival direction reconstructed using CNN denoised radio waveforms is found to be
in good agreement with the IceTop reconstruction. Finally, our approach demonstrated improved
directional reconstruction compared to traditional methods.