Speaker
Description
The ATLAS detector was designed to detect prompt particles from the LHC. A pair of long-lived particles, as part of a new Hidden Sector added to the Standard Model, would lead to challenging reconstruction and differentiating from background in $pp$ collisions at $\sqrt{s}=$ 13 TeV with the ATLAS detector. The two main backgrounds to a search for such long-lived particles are QCD multijet and Beam-induced Background (BIB), the latter being muons arising from proton bunch interaction with LHC collimators or beam gas which then deposit energy in the calorimeter. Beam-induced background is non-standard and mimics signal very well, such that an algorithm is used to isolate a sample of BIB jets in ATLAS data which is used to train a neural network. This neural network was designed to take low-level variables from the ATLAS tracker, calorimeters and muon system through 1D convolutions and an LSTM to take advantage of the natural ordering and correlations of those subsystems' constituents, and uses this to discriminate between signal, QCD and BIB jets.
Due to the BIB background requiring the use of data in training an adversarial network was added to reduce the effect of simulation/data differences into the final NN score. The adversary is instrumental in controlling systematic uncertainties, using a novel technique which trains on both signal and background to reduce the effects of simulation/data mis-modelling. To do this, a control region is constructed containing a multijet selection and signal trigger veto of both data and simulation. This way, the same population of jets in both data and simulation could be studied, such that the only difference would be mis-modelling in the input variables. The adversary uses this control region to ignore the mis-modelling while discriminating between signal, QCD and BIB.
| Are you are a member of the APS Division of Particles and Fields? | No |
|---|
