Maximum likelihood reconstruction of water Cherenkov events with deep generative neural networks

TL;DR

This paper introduces neural networks that generate probability density functions for water Cherenkov detector signals, enabling more accurate event reconstruction with fewer assumptions than traditional maximum-likelihood methods.

Contribution

The authors develop neural networks that produce likelihood functions for detector signals, improving water Cherenkov event reconstruction over existing methods.

Findings

Neural networks can generate probability density functions for detector signals.

Likelihood-based event reconstruction can be performed with fewer assumptions.

Expected to outperform traditional maximum-likelihood approaches.

Abstract

Large water Cherenkov detectors have shaped our current knowledge of neutrino physics and nucleon decay, and will continue to do so in the foreseeable future. These highly capable detectors allow for directional and topological, as well as calorimetric information to be extracted from signals on their photosensors. The current state-of-the-art approach to water Cherenkov reconstruction relies on maximum-likelihood estimation, with several simplifying assumptions employed to make the problem tractable. In this paper, we describe neural networks that produce probability density functions for the signals at each photosensor, given a set of inputs that characterizes a particle in the detector. The neural networks we propose allow for likelihood-based approaches to event reconstruction with significantly fewer assumptions compared to traditional methods, and are thus expected to improve on the current performance of water Cherenkov detectors.