Event analysis in KM3NeT using machine learning

The KM3NeT neutrino telescopes are already taking data while undergoing incremental construction in two locations in the Mediterranean Sea. KM3NeT/ARCA is a large-scale water Cherenkov detector located south-east off the Sicily coast, optimised for investigating astrophysical high-energy neutrino sources in the universe using on the order of a gigaton of seawater monitored by photo-sensors. KM3NeT/ORCA is the low-energy detector of KM3NeT, located off the French coast and sharing the same technology with a smaller and denser network of photo-sensors. The main goal of KM3NeT/ORCA is the determination of the neutrino mass ordering.
This talk aims at demonstrating the general applicability of convolutional neural networks (CNNs) in the reconstruction and data analysis of neutrino telescopes, using simulated datasets for the KM3NeT/ARCA detector as training data. For this purpose, a Keras-based framework called OrcaNet has been used, originally developed for reconstruction and classification in KM3NeT/ORCA. In this work, CNNs are employed to accomplish reconstruction as well as classification tasks for neutrino events in KM3NeT/ARCA, promising complementary information to the very time-consuming analysis pipeline based on maximum-likelihood methods. Some CNN models will be described, which have proved to provide good performance in event reconstruction e.g. for the estimation of the energy and the direction of the incoming neutrino and event-shape classification (shower-like or track-like). For a thorough overview, a short report will also be given of the tasks and performances of CNNs in ORCA.

Theme – Machine Learning, Statistics, and Algorithms