/FastSimNN

Fast simulation of detector response using ML

Primary LanguageC++GNU General Public License v3.0GPL-3.0

FastSimNN

Instant simulation of detector responses using ML. Unlike the traditional Geant4 and fast detector simulation programs, this approach makes almost instant transformation (1-2 ms/event) of true-level Monte Carlo record to a record that closely follows the full/fast detector simulation. It uses an array of neural networks that need to be trained beforehand using true-level and detector-level records.

This is how it is organized:

  • builder - read Delphes files from HepSim and train NN
  • validate - validate NN
  • compare - compare performance by reconstracting resolutions
  • detector - read txt files from Delphes/toy detectors, train NN and make predictions
  • makedata - create txt files from Delphes for NN training, or from toy detectors

Now to run

The program is designed and compiled to run on the ANL ATLAS cluster. You need to link it against Delphes.

Input of data

The inputs for this program are data from the HepSim repository (http://atlaswww.hep.anl.gov/hepsim/). The training is done using ttbar+jets and gamma+jet samples (weighted). This allows tests of jets in the range (25 GeV -3 TeV), photons, muons and electrons.

History

  • Version 1.0 December 1, 2018: Initial version
  • Version 2.0 December 14, 2018: Corrected Eta resolution
  • Version 3.0 December 16, 2018: Added muons, electrons, photons
  • Version 4.0 May 3, 2019: changing the strategy. Use toy detectors

Contact

Send comments to S.Chekanov (ANL)