360 likes | 532 Views
Energy Recovery of Dead Calorimeter Channels using Artificial Neural Networks (ANN). Stilianos Kesisoglou Victoria Giakoumopoulou Georgios Daskalakis. ECAL DPG Meeting Institute of Nuclear Physics National Center for Scientific Research “Demokritos” October 11, 2012.
E N D
Energy Recovery of Dead Calorimeter Channelsusing Artificial Neural Networks (ANN) Stilianos KesisoglouVictoria GiakoumopoulouGeorgios Daskalakis ECAL DPG Meeting Institute of Nuclear Physics National Center for Scientific Research “Demokritos” October 11, 2012
Objective and Approach Methods f h • Our aim is to provide an estimate for the energyof a Calorimeter Dead Channel (DC). • Estimation will be based on a 3x3 crystal matrix aroundthe max containment crystal (either an electron/positron or photon). • Two approach methods were examined: • Usage of a look-up table to match the 3x3pattern that includes the dead crystalagainst a list of 3x3 patterns for whichall energies are known. • Presented in July 19, ECAL DPG meeting. • Usage of an Artificial Neural Net (ANN)to estimate the energy of the dead crystal based on the energies of the live ones. • Today’s presentation 3 x 3 Crystal Matrix LU UU RU LL CC RR LD DD RD
ANN Training Strategies • Two strategies were followed for the information presentedto the Neural Network during training: • Only energy deposits of the live crystals within the 3x3 matrix was used. • Purely Calorimeter-based strategy • Additional information was provided via particle’s impact pointat the crystal’s front face. • Calorimeter + Tracker based strategy • No charge separation was used during training • Ideally we aim for an ANN that can be used also with photons. • Isolated particles were considered only • Dedicated Neural Network’s for Barrel and EndCaps were trained.
“Training / Test” Datasets (Calorimeter-based) • Analysis was performed with 8 TeV datasets: • JSON file used: Cert_190456-195947_8TeV_PromptReco_Collisions12_JSON.txt • Isolated electrons with ET > 35 GeV. • /Photon/Run2012A-PromptReco-v1/AOD • /DoublePhotonHighPt/Run2012B-PromptReco-v1/AOD • Monte Carlo sample was made using Drell-Yan and Z’ samples: • Z’ sample provides the needed high pT electrons (to test the propertiesof NN at high particle energies) • /DYToEE_M_120_TuneZ2star_8TeV_pythia6/Summer12-PU_S7_START52_V9-v1/AODSIM • /DYToEE_M_200_TuneZ2star_8TeV_pythia6/Summer12-PU_S7_START52_V9-v1/AODSIM • /DYToEE_M_500_TuneZ2star_8TeV_pythia6/Summer12-PU_S7_START52_V9-v1/AODSIM • /ZprimePSIToEE_M-750_TuneZ2star_8TeV-pythia6/Summer12-PU_S7_START52_V9-v1/AODSIM • /ZprimePSIToEE_M-1000_TuneZ2star_8TeV-pythia6/Summer12-PU_S7_START52_V9-v1/AODSIM • /ZprimePSIToEE_M-1250_TuneZ2star_8TeV-pythia6/Summer12-PU_S7_START52_V9-v1/AODSIM
“Training / Test” Datasets (Calorimeter + Tracker) • Real Data: • /Photon/Run2012A-PromptReco-v1/AOD • /Photon/Run2012A-23May2012-v2/AOD • /DoublePhotonHighPt/Run2012B-PromptReco-v1/AOD • /DoublePhoton/Run2012C-PromptReco-v1/AOD • /DoublePhoton/Run2012C-PromptReco-v2/AOD • Monte Carlo: • Only Drell-Yan sample was used • /DYToEE_M_20_TuneZ2star_8TeV_pythia6/Summer12-PU_S7_START50_V15-v1/AODSIM • /DYToEE_M_120_TuneZ2star_8TeV_pythia6/Summer12-PU_S7_START52_V9-v1/AODSIM • /DYToEE_M_500_TuneZ2star_8TeV_pythia6/Summer12-PU_S7_START52_V9-v1/AODSIM • /DYToEE_M_800_TuneZ2star_8TeV_pythia6/Summer12-PU_S7_START52_V9-v1/AODSIM • /DYToEE_M_1000_TuneZ2star_8TeV_pythia6/Summer12-PU_S7_START52_V9-v1/AODSIM
Artificial Neural Network (ANN) • ANN’s are information processing structures • Capability to simulate the learning process • Can solve problems involving pattern recognition • Most common implementation of ANN’sis the MultiLayer Perceptrons (MLP) • Learning Process (training): • Expose the NN to an appropriate samplethat includes all desired patterns. • Ensure the “training” and “applied” sampleshave the same characteristics. • Selection of input variables and NN layout. • Select the training method (minimization). • Bias control during training (“test” sample). • Sufficient training (“epochs”) MultiLayer Perceptron
Learning Process Issues • A series of MC studies performed to determine the best choices for the learning process. • NN input variables Logarithm of crystal energies • “Linearizes” the problem. • NN layout 2 hidden layers with 10 and 5 nodes respectively. • Minimization method “BFGS” method (best out of six methods). • Bias control use both a “train” and a “test” sample during training. • Events were shuffled to ensure the “train” and “applied” samples have the same characteristics (only for MC).
Calorimeter-based Strategy • Separate training for Barrel and EndCaps • Separate training for Real Data and Monte Carlo • Train with a portion of the available events (between 1/2 and 2/3’s) • Treat eight out of nine crystals in the 3x3 matrix as “live” • Input information to the Neural Network • Treat the last remaining crystal as “dead” • Output information to the Neural Network • Train Neural Network for these inputs/output • Repeat the above process for each hypothetical dead crystalin the 3x3 matrix • A total of nine Neural Networks is acquired. • Code and weights for each Neural Network is saved as C++ code ready for use. • Use trained Neural Network and apply it to the remaining eventsto display it’s performance.
Summary of Barrel and EndCap results (data) X.XX % = Sum 3x3 (corrected) Sum 3x3 (corrected) X.XX % = Sum 3x3 (RECO) Sum 3x3 (RECO) Dead xtal (corrected) YY.Y % = Dead xtal (RECO) Barrel Results ANN Look-up Table EndCap Results
Performance versus ET and η • In the next slides we examine the performance of theNeural Network versus particle’s ET and η. • Bins in ET (η) were defined. • The mean value and the error bars represent the resultsof the Gaussian fits for each defined bin (μ and σ). • Plots are provided both for the estimated energies of thedead crystal and the Sum 3x3. • Almost flat behavior as well as unbiased performanceis observed across ET (η) bins.
Calorimeter and Tracker-based Strategy • Input information to the Neural Network consists of: • Energy deposits of the “live” crystals within the 3x3 matrix. • Particle’s impact point at the front face of the crystal. • Same steps followed for the Neural Network training • No significant improvement with the addition of this information • Weights assigned by the Neural Network for the positionof impact point are very small (one-two orders of magnitude) • So we decided to stay with Calorimeter-based information only. • Nice because we treat electrons and photons alike andwe don’t care about the tracker sub-detector.
Photon Performance • Even though we used electrons for the training of the Neural Network, we see that the performance in photons is quite good. • Only MC photons tested for the moment • In the future we can mix electron and photon data for theNeural Network training.
Additional tests • We have also performed the following tests: • Data trained NN applied on MC test sample • MC trained NN applied on MC test sample • MC trained NN applied on Data test sample • Data + MC trained NN applied on both Data, MC test samples • Especially this was done to see we can extrapolateto very high ET (exotica searches) • Because of time limitations we don’t show the hundredplots but we quote the conclusions • No significant difference in performance (bias/resolution)for Data or MC trained Neural Network.
Old implementation in CMSSW There is code in CMSSW but is rather old (summer 2007). We now try to update it in order to use it with the current versions of CMSSW. Basic workflow: • Identify Dead Channels (DC). • Dead channels are defined in a file. • The dead channels must be retrieved from a database. DPG proposal? • Correct them if there is significant energy around them(i.e. E24 around dead channel > … GeV ) • Old threshold was E24 > 4.0 GeV. This threshold can be optimized. • Correct only topologies that look like isolated electrons/photons.Protection against non-isolated electron/photon or special training for such cases. • Identify the max containment crystal in the E24 around the Dead Channel. • We have to think how to correct if the dead channel is the central one.In that case the problem is more complicated. • Identify the dead channel position with respect to the max containment crystal. • Use the appropriate Neural Network and estimate the missing energyfor the particular D.C.
Implementation in CMSSW Packages: • RecoCaloTools/EcalChannelKiller Purpose: • “Kill” specific crystals (read crystals CellID from a file or from the Database) • Produce a new RecHit collection by setting dead channel energyat zero or by removing the dead channel’s RecHit from the collection. This package serves only for tests/studies.
Implementation in CMSSW Package: • RecoLocalCalo/EcalDeadChannelRecoveryProducers Purpose: • Reads the Dead Channels. • For each of them, the appropriate correction method is invoked. • Produces the corrected RecHit collection. Package: • RecoLocalCalo/EcalDeadChannelRecoveryAlgos Purpose: • Implementation of the different Dead Channel correction methods. The machinery for the implementation of the corrections is hiddenin this package. • Correction constants (Neural Network weights and code) are stored in C++ files. • Should we define the full correction function in the OfflineDB? DPG proposal?
Next Steps • Update/upgrade the existing packages in CMSSW • Try to improve the characteristics of the Neural Networkby adding photon data and lowering particle’s ET • Check and improve the performance of the Neural Networkversus particle’s ET and η. • Document the work in a CMS note.