electronid in atlas run 2€¦ · savannah thais (yale university) on behalf of the atlas...
TRANSCRIPT
Electron ID in ATLAS Run 2Savannah Thais (Yale University) on behalf of the ATLAS Collaboration
HEP 2018 Poster Session - 11.01.2018 Valparaiso
IntroductionEfficient and accurate electron identification (ID) is critical to measuring physics processes with leptons in the final states in the ATLAS detector
K.#Brendlinger WZ#Overlap#Removal#Studies 4
d
F
0
beam axis
beam spot
pixelsSCT
TRT (72 layers)
first layer (strips)
second layer
third layer
φ
η
hadronic calorimeter
electromagnetic calorimeter
R
R
insertable b-layer
HTη
φ
presampler
electromagnetic calorimeter
second layer
first layer (strips)
presampler
third layer hadronic calorimeter
Rφ
Rη
TRT (72 layers)
SCTpixels
insertable b-layer
beam spot
beam axis
d0
eProbabilityHT
η
φ
∆η×∆φ = 0.0031×0.098
∆η×∆φ = 0.025×0.0245
FHT
TRT pe/(pe+pπ)TRT pe/(pe+pπ)
TRT PID
• Electron identification algorithm uses shower shape, track, and track-cluster matching information from inner detector and EM calorimeter as inputs
• Algorithm distinguishes prompt, isolated (separated from other physics objects) electrons from hadronic jets, converted photons, and heavy flavour decays (backgrounds)
• Algorithm is optimized for all pT (4.5 GeV - ~1 TeV) and central η (0-2.47) range• Electron ID working group provides recommendation for 3 different operating
points (OPs) with varying efficiencies for genuine electron and background processes selection to suit the needs of various ATLAS analyses
Electron Identification MethodElectron ID uses a Likelihood Method: 1. PDFs (Probability Density Functions) of discriminating variables for signal
and background are formed from ATLAS data distributions (J/Ψàee for pT≤15GeV and Zàee for pT≥20GeV)
2. PDFs are smoothed with KDE to minimize statistical fluctuations3. LH discriminant is calculated: 𝑑" =
"$"$%"&
, 𝐿) * �⃗� = ∏ 𝑃/ 0 ,2 𝑥23245
4. Cut on discriminant is selected to match desired signal efficiency for specific Operating Points (note: efficiency of cut is measured in MC to ensure sample purity)
LH Background rejection ~200% of previous cut-based ID
Acknowledgements: O.K. Baker, E. Benhar, J. Reichert, L. Flores, and the ATLAS E/Gamma Electron Tag and Probe Working GroupThis material is based upon work supported by the National Science Foundation Graduate Research Fellowship
References: ATLAS-CONF-2016-024, CERN-EP-2016-262
Tag and Probe Method
ID Efficiency Measurements
Upcoming ID Improvements• Low pT region (≤ 20GeV) is important for low mass SUSY searches and hà4l analyses. ID efficiency is significantly lower here, so new variables are being
studied for low pT LH and variable selection is being optimized using n-1 variable studies• Shifts for MC PDFs are being derived from J/Ψàee samples for low pT region; this will improve low pT discriminant cut tuning.• Pile-up dependence of ID variables is being studied for Run 3 and HL LHC development. • New Machine Learning algorithms are being studied to potentially replace the LH tool. This includes Boosted Decision Trees and Neural Networks, as well as
more complex calorimeter based techniques. Working in collaboration with the ATLAS ML Working Group.
Tag-and-probe forms electron pairs from Zàee or J/Ψàee decays:Places strict selection criteria on one electron (tag) and using the unbiased other electron (probe) for efficiency measurements or PDF calculations
Sample Selection:• Signal MC: truth matching to prompt, isolated electrons• Zàee data: cut on the combined electron mass (main method) or
isolation (validation method) and a template background fit • J/Ψàee data: isolation and di-electron mass cuts, and a cut (or fit) on
𝐿67: the distance between the J/Ψ vertex and primary event vertex (to
select prompt J/Ψ) −1 ≤ 𝜏 = "=>?@ABCD/F
GHD/F ≤ 0.2
• Background : all 2à2 QCD processes (can be misidentified as electrons)
• Measuring the electron ID efficiency and how it differs from data to MC is crucial to cross-section measurements and new physics searches.
• Efficiency=ε(probes passing criteria)/(all probes) is calculated using tag-and-probe methods for all Operating Points
• Efficiency ratios used to correct MC samples so εdata= εMC to ensure reliable physics results for all analyses
New in 2017: LH PDFs made from data rather than MC; reduces uncertainty from mis-modeling and the need to shift and scale PDFs to match data