Münster 2017 – wissenschaftliches Programm
T 32.4: Vortrag
Dienstag, 28. März 2017, 11:45–12:00, JUR 498
SCYNET: Parametrizing the LHC Search results for SUSY using a Neural Net Regression — Philip Bechtle1, Matthias Hamer1, Tim Keller2, •Abtin Narimani1, Björn Sarrazin1, Jan Schütte-Engel2, and Jamie Tattersall2 — 1University of Bonn — 2RWTH Aachen
The LHC has already excluded many signatures of New Physics based on searches for various topologies. Each of the individual searches for different topologies measures a background expectation and a measured number of events along with statistical and systematical uncertainties. These published results can be used to set limits on new models of New Physics. A possible tool for such a study is e.g. CheckMATE. For each model it tests against the LHC results, it generates events, uses a fast detector simulation, performs the selection, and then compares the selected number of signal events to the background and data. This is a very general approach, however, it is very slow.
In order to make this approach useful for global fits, the evaluation of each model point must take O(<1) s. In SCYNET, this is realized by training an Artificial Neural Net regression on O (800k) simulated SUSY model points using CheckMATE for 8 TeV and 13 TeV LHC SUSY searches. In a direct approach, the parameters of the pMSSM11 are trained against a χ2 characterizing the agreement of signal and background with the data in all independent searches. In the indirect approach, pseudo-observables such as the number of partons are used to parametrize the net, such that any model of New Physics and not only a specific SUSY model can be used.