Multi-station Neutrino Vertex Reconstruction with Convolutional Neural Networks for a Simulated RNO-G Sub-array

Abstract

Ultra-high-energy (UHE) neutrinos serve as unique messengers for probing the most energetic environments in the Universe. The search for these elusive particles has driven the development of a new generation of in-ice detectors that utilize coherent radio Cherenkov emission (the Askaryan effect) from neutrino-induced cascades to detect neutrinos. The kilometer-scale attenuation length of radio waves in cold ice allows for the instrumentation of large detector volumes with sparse arrays. Consequently, this extended propagation range significantly increases the likelihood of coincident multi-station detections. Despite this, current reconstruction algorithms predominantly focus on reconstruction with just a single detector’s worth of information, ignoring the rich geometric and temporal interplay of multi-station events. In this work, we present a novel machine learning approach utilizing convolutional neural networks to reconstruct the neutrino interaction vertex from four-station coincident events using simulated Radio Neutrino Observatory in Greenland (RNO-G) stations. The architecture is modular, facilitating the rapid adaptation of other single-station reconstruction techniques. We demonstrate that while the model exhibits systematic under-prediction of the vertex radius, the radial precision achieves a relative 1σ (68% containment) error of 0.32, an improvement of a factor of two over the 0.60 achieved by current state-of-the-art single-station methods. These results highlight the transformative potential of deep learning techniques for exploiting multi-station coincidence in next-generation radio neutrino observatories.