Event-Level Classification of Boosted Multi-Tau Signatures in scNMSSM Di-Higgs Production

Abstract

In the semi-constrained NMSSM at large lambda and low tan(beta), di-Higgs production can proceed through h1h1 -> 4a1 -> 8tau when the lightest pseudoscalar lies below the b-bbar threshold. The tau pairs from each a1 decay are collimated and merge within a single jet. Taking the benchmark point m_a1 = 8.2 GeV identified in earlier work, we examine which observables separate such events from Standard Model processes producing similar final states. Signal and background samples were simulated with MadGraph5_aMC@NLO, Pythia 8 and Delphes 3, with large-radius jets reconstructed at R = 0.8. The substructure observables normally used for two-prong tagging prove ineffective here: the jet mass and the N-subjettiness ratio tau_21 have separation powers of 0.02 and 0.09 against boosted Z -> tau tau. Both are ditau jets, and the neutrinos from tau decays displace the reconstructed masses to a common value near 18 GeV despite parent masses differing by an order of magnitude. Discrimination comes instead from the event as a whole. A boosted decision tree trained on single fat jets reaches an AUC of 0.759 +- 0.002, while an event-level classifier reaches 0.941 +- 0.001, driven by the multiplicity of boosted objects and the masses of the subleading ones. Truth matching confirms these jets correspond to genuine a1 decays in 77% of events. Top-quark pair production reproduces the multi-object structure, and against ttbar -> bbbar tau tau the AUC falls to 0.855 +- 0.008. For a cross-section-weighted combination of the two simulated backgrounds the classifier reaches 0.908 +- 0.002, with 71% signal efficiency at 10% background efficiency. These are classification metrics from fast simulation, not discovery significances: reducible backgrounds, trigger acceptance and pile-up are not included.