Machine Learning usage for MET trigger

ML MET Classification (Charlie)

JIRA, mattarmost

• ATR-27266 CNN FakeMETFilter Studies
• mattermost: Cell-based ML MET

Presentations

• MET trigger meeting on 30 Jan 2023 Identify Fake MET Events Using ML (Kenji Hamano)
• UVic ML MET study meeting on 17 Feb 2023 MET Machine Learning Update (Charlie Chen)
• UVic ML MET study meeting on 01 Mar 2023 MET Machine Learning Update (Charlie Chen): CNN architecture details
• UVic ML MET study meeting on 01 Mar 2023 Additional Input for CNN (Kenji Hamano)
• MET trigger meeting on 6 Mar 2023 Classifying True/Fake MET Events using ML (Charlie Chen)
• UVic MET meeting on 10 Mar 2023 MET Machine Learning Update (Charlie Chen)
• UVic ML MET study meeting on 15 Mar 2023 MET Machine Learning Update (Charlie Chen)
• UVic MET meeting on 24 Mar 2023 NN MET filter considerations (Kenji Hamano)
• UVic ML MET study meeting on 29 Mar 2023 MET Machine Learning Update (Charlie Chen)
• UVic ML MET study meeting on 12 Apr 2023 MET Machine Learning Update (Charlie Chen)
• UVic ML MET study meeting on 12 Apr 2023 To do (Kenji Hamano)
• UVic ML MET study meeting on 26 Apr 2023 FakeMETFilter: A CNN-based Approach to Classifying True/Fake MET Events in the ATLAS Trigger (Charlie Chen)
• UVic ML MET meeting on 3 May 2023 FakeMETFilter: A CNN-based Approach to Classifying True/Fake MET Events in the ATLAS Trigger (Charlie Chen)
• MET trigger meeting on 8 May 2023 FakeMETFilter: A CNN-based Approach to Classifying True/Fake MET Events in the ATLAS Trigger (Charlie Chen)
• UVic ML MET meeting on 10 May 2023 Feedback from the MET Trigger Meeting (Kenji Hamano)
• UVic ML MET meeting on 10 May 2023 MLMET_Classifier (Ben Scheuer)
• UVic ML MET meeting on 17 May 2023] FakeMETFilter: A CNN-based Approach to Classifying True/Fake MET Events in the ATLAS Trigger (Charlie Chen)
• UVic ML MET meeting on 24 May 2023 MLMET_Classifier (Ben Scheuer)
• UVic MET meeting on 26 May 2023 FakeMetFilter status and ideas (Kenji Hamano)
• UVic ML MET meeting on 31 May 2023 MLMET_Classifier (Ben Scheuer), 2d pattern investigation, data vs mc.
• UVic ML MET meeting on 24 May 2023 FakeMETFilter: A CNN-based Approach to Classifying True/Fake MET Events in the ATLAS Trigger (Charlie Chen), tSNE plots.
• UVic ML MET meeting on 24 May 2023 Zmumu events (Kenji Hamano)
• UVic ML MET meeting on 7 Jun 2023 FakeMETFilter: A CNN-based Approach to Classifying True/Fake MET Events in the ATLAS Trigger (Charlie Chen), updated training method.
• UVic ML MET meeting on 7 Jun 2023 MLMET_Classifier (Ben Scheuer), 2d pattern investigation with correct phi binning.
• UVic MET meeting on 9 Jun 2023 FakeMETFilter: A CNN-based Approach to Classifying True/Fake MET Events in the ATLAS Trigger (Charlie Chen), new mc-only training with corrected phi binning, Saliency plots
• UVic ML MET meeting on 14 Jun 2023 MLMET_Classifier (Ben Scheuer), test on signal like EX data.
• UVic ML MET meeting on 14 Jun 2023 data only training (Charlie Chen): EB data + signal like EX data
• UVic ML MET meeting on 14 Jun 2023 data only training (Charlie Chen)
• UVic MET meeting on 16 Jun 2023 FakeMETFilter: A CNN-based Approach to Classifying True/Fake MET Events in the ATLAS Trigger (Charlie Chen), ZeroBias+EX signal training.
• UVic ML MET meeting on 21 Jun 2023 FakeMETFilter: A CNN-based Approach to Classifying True/Fake MET Events in the ATLAS Trigger (Charlie Chen), EX background + EX signal training.
• Pre-ATLAS Week ATLAS-Canada mini-Workshop on 26 Jun 2023 FakeMETFilter: A CNN-based Approach to Classifying True/Fake MET Events in the ATLAS Trigger (Charlie Chen)
• UVic ML MET meeting on 5 Jul 2023 Exploring Different Binning for MLMET (Ben Scheuer)
• UVic ML MET meeting on 19 Jul 2023 Current status and future plans (Kenji Hamano)
• UVic ML MET meeting on 26 Jul 2023 FakeMETFilter: A CNN-based Approach to Classifying True/Fake MET Events in the ATLAS Trigger (Charlie Chen), pixel by pixel comparison between data and mc.
• UVic ML MET meeting on 26 Jul 2023 resulting plots.
• UVIc MET meeting on 28 Jul 2023 FakeMETFilter: A CNN-based Approach to Classifying True/Fake MET Events in the ATLAS Trigger (Charlie Chen), the second bin was merged into the first bin.
• UVIc MET meeting on 28 Jul 2023 resulting plots.
• UVIc MET meeting on 28 Jul 2023 KS Pass/Fail 2d plot added.
• UVic ML MET meeting on 2 Aug 2023 New 2d eta-phi plots (Kenji)
• UVic ML MET meeting on 2 Aug 2023 KS Pass/Fail 2d plot with new binning (Charlie)
• UVic ML MET meeting on 9 Aug 2023 Training with new binned signed inputs (Charlie).
• Contribution to HCW2023 FakeMETFilter: A CNN-based Approach to Classifying True/Fake MET Events in the ATLAS Trigger (Charlie).
• UVic ML MET meeting on 13 Sep 2023 Impact of filters in the convolution layer (Charlie)
• MET Trigger meeting on 2 Oct 2023 FakeMETFilter: A CNN-based Approach to Classifying True/Fake MET Events in the ATLAS Trigger (Charlie)
• UVic MET meeting on 13 Oct 2023 Cell noise and Global variables (Kenji)
• UVic ML MET meeting on 18 Oct 2023 Diagnostic plots. EB data, ttbar probability; tSNE, Saliency maps, Filter maps, Feature maps, PM signal acceptance (Charlie)
• UVic ML MET meeting on 18 Oct 2023 Cell noise and event shape variables (Kenji)
• MET Trigger meeting on 23 Oct 2023 FakeMETFilter: A CNN-based Approach to Classifying True/Fake MET Events in the ATLAS Trigger (Charlie)
• UVic ML MET meeting on 1 Nov 2023 Adding more global variables and significance layer (Charlie)
• UVic ML MET meeting on 15 Nov 2023 Normalization of significance layer and CNN only model are included (Charlie). "mm_gf_original_cn" model is the best.
• UVic ML MET meeting on 13 Dec 2023 EB data, ttbar probability, PM signal acceptance; Saliency maps, Filter maps (Charlie)
• MET Trigger meeting on 22 Jan 2024 FakeMETFilter: A CNN-based Approach to Classifying True/Fake MET Events in the ATLAS Trigger (Charlie)
• MET Trigger meeting on 22 Jan 2024 Event by event Saliency Maps (Charlie)
• UVic ML MET meeting on 24 Jan 2024 Saliency maps with mirrored input (Charlie)
• UVic ML MET meeting on 24 Jan 2024 Event by event Saliency maps with original input
• UVic ML MET meeting on 24 Jan 2024 Event by event Saliency maps with mirrored input
• UVic Met meeting on 2 Feb 2024 ONNX Implementation Steps (Charlie)
• UVic ML MET meeting on 28 Feb 2024 CNN using only E_T sig for layer 1 and 2. The new CNN is as good as before. (Charlie)
• UVic ML MET meeting on 28 Feb 2024 Saliency maps for the new CNN. More symmetric than before. (Charlie)
• ATLAS Machine Learning Forum on 7 Mar 2024 FakeMETFilter: A CNN-based Approach to Classifying True/Fake MET Events in the ATLAS Trigger (Charlie)
• UVic ML MET meeting on 13 Mar 2024 background eff vs signal eff plots (Charlie)
• UVic Met meeting on 22 Mar 2024 background eff vs signal eff plots with muon triggers (Charlie)
• ATLAS Canada Workshop on May 23, 2024 FakeMETFilter: A CNN-based Approach to Classifying True/Fake MET Events in the ATLAS Trigger (Charlie)

Documentation

• MLMETClassifier User Guide by Charlie

Codes

• gitLab: MLMETClassifier

Cell ntuple production

• gitLab: trigMetCellTree

LAr plotting macros

The plots are indeed by a combination of athena (to dump the raw cell info), root macros and a root library I wrote for the endcaps. For the FCal there is a separate standalone root-based program.

I've attached a few files that are neefded for these plots:
dumpCellsInClusters.py: a simple athena jobO file I ran in Athena-22.0.41.9 on a ESD from run 325713 to make ascii files for a few events with the raw cell info

PlotPolar.{hh,cc}: the library code to make the endcap plots
PlotPolarLinkDef.h: needed for the above libary
Makefile.PlotPolar: to compile the library

plotB.C: the macro to plot the barrel
plotEC.C: the macto to plot the endcaps

plotLArSplash2022.C: the master macro to plot one event

For the FCal:
drawFCal-April2011.cc: the root-based program source to make FCal plots
FCalBinNumbersNew-April2011.txt: the bin mapping needed for the FCal plots
Makefile.drawFCal: to compile the FCal plot program
FCalDataNew-April2011-Guillaume.txt: example data file to plot

Typical usage for the barrel and endcap plots on one event would be in root:
gSystem->Load("../PlotPolar/libPlotPolar.so");
.L ../plotB.C++
.L ../plotEC.C++
.L ../plotLArSplash2022.C
plotLArSplash2021("All_cells_in_clusters_4.dat",false,325713,426221175,400,"Cells in Clusters")

Typical usage for the FCal plots (directly on comandline, not in root):
./drawFCal-April2011 -d../data23_comm.00447705_ESD_beam1/FCal_cells_0.dat -i2 -z"Energy [MeV]" -sC

Reference

• ATL-PHYS-PUB-2023-001 Electron Identification with a Convolutional Neural Network in the ATLAS Experiment
• arXiv:2311.17914 (hep-ex) Reconstruction of electromagnetic showers in calorimeters using Deep Learning
• arXiv:2311.05952 (hep-ex) Learning Broken Symmetries with Resimulation and Encouraged Invariance
• ML Forum General Meeting on 07 Mar 2024 ML at Work: Electron Energy regression with CNN

Plotting format:

• LAr Cells in Clusters with Timing

Note the barrel grid in eta-phi, and the transverse plane projective (in depth) projection for the HEC. I am particularly interested in how he does the pie-slice shapes in the HEC figures (probably some polar plot option), and how the FCAL is handled (I believe one per module since not projective in eta).

Global variables

• arXiv:2007.12600 (hep-ex) Measurement of hadronic event shapes in high-pT multijet final states at s√=13 TeV with the ATLAS detector

The quantities are:

a) the transverse thrust and its minor component (equation 1);

b) the sphericity and aplanarity, obtained from the linearized sphericity tensor (equations 2 and 3);

c) the transverse sphericity (equation 4 and previous).

These could be calculated over cells, assuming massless 4-vectors centred in the middle of the ATLAS detector.

a) requires a maximization to find the transverse thrust axis, while b) and c) require to find the eigenvalues of a 3x3 or 2x2 matrix. Perhaps b) and c) are easier to implement. There may already exists implementations somewhere in ATLAS.

It would be interesting to see if their distributions have some sensitivities to real vs fake MET events. My gut feeling is that a) or c) might be the most useful.

Note that the paper quoted above has an erratum: https://link.springer.com/article/10.1007/JHEP12(2021)053 You can follow the details of who did what/when (perhaps to ask for code...) on Glance https://atlas-glance.cern.ch/atlas/analysis/papers/details.php?ref_code=STDM-2019-02

ML for MET trigger (Yiming (Ablet) Abulaiti)

This study id for phase-II upgrade

Presentations

• PPES meeting on 12 Dec 2022 ML learning for MET trigger Use only Calo data
• PPES meeting on 27 Feb 2023 ML for MET trigger

JIRA

Upgrade samples

R21.9

R23.0

• JIRA: ATLMCPROD-10353 Upgrade MC Prod with ATLAS-P2-RUN4-01-01-00

Code

• KHCalo See "Code to access cells" section.

ML MET (Mohamed Belfkir)

Presentations

• MET trigger mreeting 02 Mar 2022 ML based HLT MET Trigger
• MET trigger mreeting 13 Jun 2022 Update on the ML-based HLT

Code

Cell Ntuple:

• gitLab: TriggerNTupleMaker R21
• gitLab: TriggeNTupleMaker R22

Also

• KHCalo See "Code to access cells" section.

ML code:

• gitLab: HLTTriggerML

NN MET (Zhelun)

Presentations

• 4 Jun 2021 MET HLT combination update (Zhelun Li)
• 2 Jul 2021 MET HLT combination update (Zhelun Li)
• 16 Jul 2021 MET HLT combination update (Zhelun Li)
• 13 Aug 2021 MET HLT combination update (Zhelun Li)
• 4 Oct 2021 HLT Neural network studies (Zhelun Li)

Codes

• gitLab: HLT_METtrigger_combinations

nnMET + offline METNet (Malin Elisabeth Horstmann)

QT JIRA

• JIRA: ATR-26920 QT: nnMET + offline METNet

-- KenjiHamano - 2023-02-23