Speaker
Description
We present a simulation-based inference pipeline for arrival-direction reconstruction in radio arrays, demonstrated on GRANDproto300-like ultra-high-energy cosmic-ray simulations generated with ZHAireS. Each event is ingested by a physics-informed graph neural network that uses an analytic plane-wavefront fit as a strong directional prior, producing a directional embedding that conditions a normalizing flow to return the posterior over sky direction. After training on $\sim$8,000 air showers, we achieve a median angular resolution of 0.38 deg on held-out events, and coverage tests show mildly conservative uncertainties. Because our approach relies on antenna positions and trigger times (rather than waveform features), the same pipeline can be applied to UHE neutrinos by retraining on neutrino simulations for the target layout. Code and runnable example notebooks are publicly available on GitHub and archived on Zenodo (DOI: 10.5281/zenodo.16895985).