Graph convolutional neural networks for neutrino telescopes

While it is currently undergoing rapid developments, the neutrino sector still has many open questions: the neutrino masshierarchy is not known, several parameters of the PMNS matrix are poorly constrained, and the inability to explain the nonzero neutrino masses is a clear indication of physics beyond the Standard Model.

Neutrino oscillation experiments at theIceCube Neutrino Observatory may be able to address these fundamental questions, but the reconstruction of neutrinointeractions in the detector is a challenge which urgently needs to be addressed: the current reconstruction algorithm isprohibitively time-consuming, cannot account for all known optical anisotropies in the ice, and cannot make full use of all ofthe information from new modules in the IceCube Upgrade due to excessive computing time and memory requirements.

Thisproject proposes graph convolutional neural networks (GCN) as a machine learning paradigm excellently suited for neutrinotelescope experiments, with potential to revolutionise reconstruction in IceCube.

GCNs impose no structural requirements ondata, requiring only a concept of adjacency, naturally afforded by the spatial, temporal, and causal separation of hits in thedetector.

With expected improvements in particle identification of a factor of 10 compared to analytical methods and a factor10,000 speed-up in reconstruction, GCN-based reconstruction will be developed and implemented in the IceCube-DeepCoreoscillation analysis, to better measure PMNS parameters by improving the atmospheric muon background rejection andperforming per-flavour event categorisation.

Powerful and fast GCN-based reconstruction will benefit several physicsanalyses in IceCube --- and possibly ANTARES, KM3NeT, and Baikal-GVD --- and help answer the open questions in theneutrino sector.

Finally, the possibility for private and public sector partners to benefit from these high-performance GCNtools will be explored through intersectional partnerships.