Tagging Emerging Jets using Graph Neural Networks

Many extensions to the Standard Model predict the existence of strongly interacting dark sectors, which behave similarly to QCD but interact weakly with SM particles. Depending on the parameters of the dark sector, one potential signature could be that of an "emerging jet". In particle detectors such as the ATLAS experiment, this signature could be seen as sprays of many displaced tracks - particle trajectories that do not originate from the main proton-proton interaction point, and many displaced vertices - the common origin of tracks that are away from the primary proton-proton interaction point. Graph Neural Networks (GNN) have shown great promise in capturing complex dependencies and patterns in graph-structured data, making them well-suited for analyzing the intricate topology of emerging jets. A GNN-based flavor tagging algorithm has recently been deployed in ATLAS and significantly outperforms previous taggers. Its architecture is used to tag emerging jets with high accuracy while significantly suppressing the QCD background. The architecture of the GNN also enables the classification of displaced tracks as well as the identification of displaced vertices within the jet cone, providing valuable insight into the topology of the jet.