Reconstruction of the energy of neutrinos with neural networks: Event-by-event uncertainty estimation

Abstract: When high-energy neutrinos interact with matter, radio waves will be emitted. Radio detectionallows us to measure UHE(> 1016eV) neutrinos by instrumenting a huge volume with a sparsearray of radio antenna stations at a low cost. The radio signal measured by the antennas can then beanalyzed to estimate the physics quantity of the corresponding interaction. Traditional reconstructionmethods are time-consuming to develop and often do not account for all information in the signal. Onthe other hand, deep learning-based reconstruction is a powerful technique for radio detector data.Promising results predicting the neutrino energy and direction have been already achieved. However,so far, only the nominal value was predicted, but for the interpretation of data, the event-by-eventuncertainty is crucial and almost as important as the reconstruction of the nominal value. In thisthesis, I added an event-by-event uncertainty prediction of neutrino energy to the deep learningreconstruction using two methods: 1) Likelihood Inference or 2) Normalising Flows. NormalizingFlows allows predicting arbitrary PDFs event-by-event, whereas the first method only predicts astandard deviation per event. The predicted PDFs using Normalising Flows are close to a Gaussian.Hence, both methods can be used interchangeably