Speaker
Description
Reconstructing the energy fluence of UHE neutrinos and cosmic rays is a crucial challenge for the Giant Radio Array for Neutrino Detection (GRAND) and similar radio arrays. Because radio noise has random phase, signal plus noise can interfere constructively or destructively, so conventional time-domain noise subtraction can yield non-physical negative fluences and underestimates uncertainties. Building on a recently proposed Rice-distribution likelihood for voltage magnitudes, we compute the energy-fluence summaries for each antenna and feed them, together with trigger time, polarization, and geometry, into a physics-informed graph neural network. Our machine-learning model is regularized by a forward model of the radio footprint where we isolate the geomagnetic component and fit a one-dimensional lateral distribution function in shower coordinates, restricting solutions to physically viable patterns and improving generalization. In this poster contribution, we will show (on ZHAireS simulations) that our approach achieves competitive energy-fluence resolution without ad-hoc SNR cuts while leveraging all triggered antennas.