Identifying Di-Lepton Signatures of New Physics at Neutrino Telescopes using Graph Neural Networks

Lepton pairs are a powerful signature in searches for new particles, including axions and heavy neutral leptons (HNLs). While such signatures have been studied extensively in accelerator neutrino experiments, they have yet to be explored in neutrino telescopes. In this talk, we investigate the ability of IceCube and KM3NeT to identify μ⁺μ^- pairs and, for the first time, e⁺e^- pairs. To do this, we use the open source Prometheus simulation tool to generate di-leptons in IceCube DeepCore, KM3NeT/ORCA and the IceCube Upgrade. We then present a graph neural network model for reconstructing these events and distinguishing them from Standard Model backgrounds. Using this model, we estimate the sensitivity of neutrino telescopes to various HNL models, including type I seesaw HNLs as well as HNLs with new dark sector interactions. We comment on the importance of the detector medium—water or ice—in resolving di-leptons and thus maximizing sensitivity to new physics.