Machine learning based background rejection for Baikal-GVD neutrino telescope

Abstract

Baikal-GVD is a gigaton-scale underwater neutrino telescope currently under construction in Lake Baikal. Its principal components are optical modules, registering photons propagating through the telescope’s working volume. Part of the activations of the optical modules are due to the natural luminescence of the water, and thus appear as noise in the data. We present a neural network, which efficiently rejects this background and reaches 97% signal purity (precision) and 99% survival efficiency (recall) on the Monte-Carlo data. The neural network has a U-net like architecture based on the temporal structure of optical modules activations.