As a part of CMS service task for EGM POG, I worked on a NN based photon ID for PF photons in 2021, based on the MC generated in the Run3Summer21 campaign. I am validating the models in the Run3Summer23 campaign for photon candidates in the high-eta (forward) region, with 2.4 < abs(eta) < 3.0.
Four models are trained with the same architecture, two for barrel and two for endcap photons (non-weighted and pt-weighted in each case). The technical details are given below:
Model architecture: 128/64/64/32/16/1
Activation: 'relu' (except for the last layer, where it is 'sigmoid')
kernel_initializer='he_normal' (not applied on the last layer)
optimizer='adam'
loss='binary_crossentropy',
metrics=['accuracy']
These four models are kept in TrainedModels/ in h5 and pb format, with the scaling parameters and training-vs-testing plots.
The following samples were used for picking the photons candidates (all reco photons above 10 GeV) which may or may not pass the existing PF Photon ID).
Signal photons are prompt, and they match with a gen-photon with a dR cone of radius 0.3. The following Gamma-Jet samples are used to pick these signal photons.
/GJet_Pt-10to40_DoubleEMEnriched_TuneCP5_14TeV_Pythia8/*/GJet_Pt-40toInf_DoubleEMEnriched_TuneCP5_14TeV_Pythia8/*
QCD and TauGun samples were used to pick backgrounds as listed below. These are either fakes, or FSR/ISR photons, or in case of the TauGun sample, most likely coming from pion decays. The fakes are not supposed to match any gen-photons within a dR cone of radius 0.3. The candidates which match with any gen-photon are required to be not prompt.
/QCD_Pt-80to120_EMEnriched_TuneCP5_14TeV_pythia8/*/QCD_Pt-120to170_EMEnriched_TuneCP5_14TeV_pythia8/*/QCD_Pt-170to300_EMEnriched_TuneCP5_14TeV_pythia8/*QCD_Pt-300toInf_EMEnriched_TuneCP5_14TeV_pythia8/*/TauGun_Pt-15to500_14TeV-pythia8/*
where *=Run3Summer21MiniAOD-FlatPU0to70_120X_mcRun3_2021_realistic_v5-v2/MINIAODSIM
The photon variables that were fed into the models are listed below. The scaling parameters for these input variables are summarized in the model_info.json file.
- hadTowOverEm
- trkSumPtHollowConeDR03
- ecalRecHitSumEtConeDR03
- SigmaIEtaIEta3x3
- SigmaIEtaIEtaFull5x5
- SigmaIEtaIPhiFull5x5
- EcalPFClusterIso
- HcalPFClusterIso
- hasPixelSeed
- R9Full5x5
- hcalTowerSumEtConeDR03
Note: The same order has to maintained while feeding these variables to the neural network.
Keeping the definition of signal and background the same, I am using the following datasets to extract the photon candidates.
/GJet_PT-10to40_DoubleEMEnriched_TuneCP5_13p6TeV_pythia8/*/GJet_PT-40_DoubleEMEnriched_TuneCP5_13p6TeV_pythia8/*/QCD_PT-80to120_EMEnriched_TuneCP5_13p6TeV_pythia8/*/QCD_PT-120to170_EMEnriched_TuneCP5_13p6TeV_pythia8/*/QCD_PT-170to300_EMEnriched_TuneCP5_13p6TeV_pythia8/*/QCD_PT-300toInf_EMEnriched_TuneCP5_13p6TeV_pythia8/*/TauGun_E-10to100_13p6TeV_pythia8/*/TauGun_E-100to3000_13p6TeV_pythia8/*
where *= Run3Summer23MiniAODv4-130X_mcRun3_2023_realistic_v14-v2/MINIAODSIM
The models, the training variables and their scaling parameters are picked using the model_info.json file.
I am using a mini-AOD to ntuple maker tool to write photon branches that I need into root files. This tool can be found here. After the root files are generated, these are turned into flat csv files (with each entry representing a photon) using root2csv.py. These csv files are read into the jupyter notebook as pandas dataframes.
The details of the validation are available in the notebook.