Use of a Normalizing Flow model for generating Drell-Yan events in the ATLAS collaboration at the LHC

The search for the dimuon decay of the Standard Model (SM) Higgs boson represents a typical bump-hunting physics analysis performed at the Large Hadron Collider (LHC). It looks for a tiny peak created by new physics, or the Higgs boson in this case, on top of a smoothly falling SM background in the two-muons invariant mass spectrum ${m_{\mu\mu}}$. The background events are estimated from a data-driven side-band fit with a floating factor for normalization and a pre-determined function for the background spectrum whose parameters are constrained from systematic uncertainties. The criteria for determining the background function are based on the spurious signal, which measures the residual signal events obtained from a signal-plus-background fit to background-only simulated events. Therefore, these simulated events must have enough statistics, an order of billions of events, so that their statistical fluctuations are negligible compared to the expected number of signal events. However, generating Drell-Yan events with high-order QCD calculations and detailed detector simulation is computationally expensive. Our study uses a normalizing flow model trained on simulated events by the ATLAS experiment to generate billions of events with GPUs for the spurious signal study. Preliminary results show that the normalizing flow model accurately describes both the muon kinematic variables that is trained on and the existing correlations among these variables. This procedure can be easily adapted to other LHC bump-hunting analyses requiring high statistics of simulated events.