Deep-learning denoising of radio signals for ultra-high-energy cosmic-ray detection

Extensive air showers initiated by ultra-high-energy cosmic rays produce broadband radio pulses detectable over large areas, but Galactic and instrumental backgrounds limit sensitivity near threshold. The Giant Radio Array for Neutrino Detection (GRAND) will use large arrays of autonomous radio antennas to detect these air showers, requiring reliable recovery of weak pulses from background fluctuations. We develop a convolutional denoiser for GRAND-like antenna traces that combines the time-domain waveform with the magnitude and phase of its Fourier transform. The model is trained on 4.1 × 10 5 simulated traces constructed from air-shower emission, a reduced GRAND-like radio-frequency (RF) chain response, and broadband noise normalized to the expected Galactic-noise scale. In these simulations, denoising increases the detection probability at fixed false-positive rate and improves pulse-time recovery near threshold. The reconstructed peak amplitude is attenuated in the noise-dominated regime. This bias decreases as the pulse becomes better resolved, but a smaller channel-dependent bias persists at higher signal-to-noise ratio. Near threshold, denoising therefore primarily improves candidate recovery and timing, while the measured amplitude response provides a basis for calibrating amplitude-sensitive observables.