Identification of tau leptons using machine learning at the scouting dataset

Here we present a highly efficient machine-learning algorithm that is designed to classify tau leptons amidst challenging scenarios involving jets mimicking taus. With an impressive average accuracy of 87%, the algorithm outperforms traditional cut-based isolation methods. Leveraging a neural network with a permutation invariant and deep-sets architecture, the model accommodates various tau decay patterns. Implemented on the scouting dataset, our algorithm facilitates the observation of taus at remarkably low masses. Notably, it unveils the upsilon to tau tau decay, a discovery marking the first instance of such observation at any hadron collider. This advancement in tau lepton identification holds significant promise for enhancing the precision and scope of third-generation particle physics research.