Speaker
Description
The measurement of the trilinear Higgs boson self coupling, accessible through Higgs pair production, is one of the most important goals of the High-Luminosity LHC (HL-LHC), since it allows to probe the shape of the Higgs potential, a crucial prediction of the electroweak symmetry breaking mechanism. However, the low pair production cross-section is a challenge, namely for the online event processing and selection performed by the trigger system. Despite having the largest branching ratio, the fully hadronic HH->4b channel poses significant additional triggering difficulties, since the cross-section for b-quarks is approximately seven orders of magnitude larger than for the Higgs boson. Currently, most triggers employed at the High-Level Trigger (HLT) rely on b-tagging algorithms, which are both CPU expensive and need high thresholds to reduce background rate. This motivates the study of alternative signal selection techniques.
In this work, we aim to employ Energy Flow Polynomials (EFPs) computed at event-level and study their sensitivity to different event topologies, discriminating four-jet HH->4b signal events from dijet background, for an environment with pile-up of 200 proton-proton collisions per bunch crossing. The performance of a supervised linear classifier trained with the EFPs is benchmarked by two alternative techniques: a standard multijet trigger with a transverse momentum cut and a Deep-Sets-based neural network, implemented with Particle Flow Networks (PFN) . Deep Sets algorithms using hadronic jets and tracks as inputs for b-tagging have been employed by the ATLAS experiment in the HLT. In this work, we use only the four-momentum of the jets in the events as input to the Deep Sets. Jets in the simulated events were reconstructed using the anti-kt algorithm, with R=0.4, and their transverse momentum was smeared to reproduce the expected resolution of the HLT. We show that both event-level EFPs and PFN increase the signal selection efficiency by at least 30% for events below the selection threshold of a trigger based on a multijet transverse momentum cut, for the same background acceptance rate. Finally, we also show that, due to the strong correlation between the EFPs, event selection relies on a very limited number of variables per event, which is particularly relevant in the software trigger, where latency and computational resources are limited.