CNN Reconstruction of Neutrinos at the Radio Neutrino Observatory in Greenland

Ultra-high energy neutrinos are unique messengers of extreme cosmic events. Their detection is crucial in identifying and pinpointing high-energy events in the cosmos. The most efficient way to detect them is through the radio waves they emit when interacting with dense media like ice. The Radio Neutrino Observatory in Greenland (RNO-G), the first northern-hemisphere in-ice radio array, aims to detect these neutrinos via this coherent radio emission. A critical challenge is reconstructing neutrino direction and energy from recorded waveforms.

This research focuses on developing a convolutional neural network (CNN) trained on simulated neutrinos to infer neutrino properties. “These simulated events are readout by a simulated detector as waveforms that are processed into structured training data and fed into a neural network. The network was first inspired by an image classification model and the CIFAR10 dataset, but was then adapted to a regression CNN. It was then expanded to be able to operate on multi-station incident events.

In this contribution, I will discuss the the data preparation, network architecture, and initial performance of the model. These results will help evaluate the accuracy of multi-station event reconstruction and benchmark machine learning performance against state-of-the-art physical models.