Machine Learning Methods for the Reconstruction of UHECRs and VHE Neutrinos from Space-based Detectors

Larger exposures are increasingly necessary to advance the study of Ultra-High-Energy Cosmic Rays and Very-High-Energy neutrinos, but expanding ground-based observatories poses a major challenge. Because of this, there is a growing interest in space-based or near-space observation methods such as those being designed for POEMMA Balloon with Radio. These instruments, however, also present unique challenges for data reconstruction, as they typically rely on single or dual telescopes, and the distance between the particle cascade and detector(s) can be quite large.

To address these challenges, this research aims to develop four convolutional neural networks (CNNs) to enhance data processing and event reconstruction. First, preprocessing algorithms prepare the data by reducing the 4D problem space to 3D. The top-level CNN will categorize each event as a shower or background event. The 2nd CNN will distinguish pixels with shower photons from those with only night sky or electronics background. The 3rd CNN analyzes this data to extract the geometric axis of the airshower. The 4th CNN integrates inputs from the previous three to reconstruct the primary particle's energy and composition. In this contribution, an overview of this reconstruction method and its first results on simulations will be presented.