Machine learning is a useful tool for identifying radio pulses from cosmic-ray air showers and for cleaning such pulses from radio background.
This can lower the detection threshold and increase the accuracy for the pulse time and amplitude. We have trained Convolutional Neural Networks (CNNs) using CoREAS simulations and background recorded by a prototype station at the IceTop surface array at the South Pole and have applied them to air-shower measurements by this station.
The station consists of 3 SKALA antennas and 8 scintillators, which are used to trigger the readout of the antennas upon a sixfold coincidence. Afterwards, the radio signal is filtered to the band of 70-350 MHz. By applying neural networks to search for radio signals in about four months of data, we find about five times more events than by a traditional method based on a signal-to-noise ratio cut after filtering for radio frequency interferences. Despite the lower threshold, the purity of the selected events seems to improve, and the angular resolution of the radio measurements does not deteriorate, which we have confirmed by a comparison of the reconstructed shower direction with IceTop. This analysis thus provides experimental confirmation that neural networks can indeed be used to clean air-shower radio signals from background and to lower the radio detection threshold of hybrid arrays combining particle and radio detectors.