1 Pixel Cluster Splitting Using Templates D. Fehling, G. Giurgiu, P. Maksimovic, S. Rappoccio, M.Swartz Dept of Physics+Astronomy, Johns Hopkins University
1
Pixel Cluster Splitting Using Templates
D. Fehling, G. Giurgiu, P. Maksimovic, S. Rappoccio, M.Swartz
Dept of Physics+Astronomy, Johns Hopkins University
2
• clustering algorithm needs to include corner adjacency- thresholds can create apparently unlikely cluster shapes
•minimum two-track separation in f (local x) is ~3 pixels (300 mm)
•minimum two-track separation in z (local y) varies from ~2 pixels (h=0) to ~12 pixels (h=2.5) or 0.3-18 mm
• standard and template reconstruction will fail when clusters merge- template reco will return bad probabilities when this
happens
Pixel clusters have a characteristic shape caused by Lorentz drift
Two-Track Separation in Pixel System
3
Slides 4-13 summarize the pixel template reconstruction technique. Lots more detail can be found in CMS Note-2007/033
Template Reconstruction
4
• Pixelav transport simulation + E-field modeling w/ TCAD 9.0- data well described by tunable double-junction model
from F =(0.5-6)x1014 neq/cm2
•Use to calculate a priori cluster shapes for improved analysis technique
Sensor ModelingOver the last 4 years, we (VC + MS) have successfully modeled irradiated pixel sensors fabricated on DOFZ substrates at several F and T,
5
•Sum charges on all pixels: Qclus
•Truncate individual pixel signals to cotb-dependent maximum- sum projections: Py/x
i
•Account for thresholds:- add information back by
creating Pseudo-Pixels at the ends of the cluster
- have 50% of threshold height and 100% uncertainties
- pulls fit near cluster edges and improves resolution
Apply fitting procedure to projections Pyi and Px
i: -> scale and translate shape to fit
Fit projected cluster shapes to simulated shapes (templates):
Template-based Reconstruction Algorithm
6
• RMS residuals not Gaussian fit sigma (tails included) • Before irradiation, template algorithm improves the
resolution at all h- for Q/Qavg<1 (~70% of all hits), 10-20% improvement- for 1<Q/Qavg<1.5 (~30% of all hits), 20-100% improvement
Comparison with Standard AlgorithmAfter small corrections for residual effects
high-h deltas
7
•After irradiation, Standard technique is more affected than templates
- z-resolution in both charge bands, 100% improvement- f-resolution at large h, 30-200% improvement
high-h deltas
8
• Template reconstruction has moderate sensitivity to track angles- use Standard technique for first pass track finding/fitting- use Template technique in second pass track fit (angles from 1st
pass)• Study with sample of simulated muon tracks
• Template technique exceeds the Standard technique at all h and Qclus
• x(f) resolution worsens at large h ?- caused by low Qclus “junk” from showering in our not-so-thin detector- ~ 7% of high-h tracks have low-Qclus hits on them
Implementation in CMS Tracking
•Pulls are sensitive to resolution tails➡ template reconstruction kills tails!•Biggest improvements are in d0, f0 pulls in the regions > 3
s➡ expect to see significant S/N improvements in b/t-tagging
d0 + template alg+ standard alg
f0
Effect of 2nd pass on track parameters
10 GeV m’s
10
Goodness-of-fit•A by-product of the
template fitting procedure is a x2 that reflects the consistency between the shapes of the cluster projection and the interpolated template
- template object stores the expected x2 distribution in a simple parameterization that depends upon Qclus
- convert these into x- and y-probabilities
•Suppresses low-Q junk clusters that arise from secondary interactions with 1-2 % inefficiencies
•Can remove low-Q with no inefficiency
No Probability Cut
P>10-3
11
•A. Dominguez has been developing an improved pixel track seeder that compares the lengths of y-clusters (global z) in the pixel barrels
- can significantly reduce the number of trial seeds and therefore the track finding time (dominates reco time)
•Intrinsic y-length resolution of the templates is about twice that of the simple cluster length method
- seeds have local angles, can use templates in 2nd pass
- template probabilities determine consistency with angle hypotheses and are normalized to resolution
- can do both x- and y-projections- can do barrel/FPix seeds•Avoids “junk” hits on tracks
(may be more junk in real LHC environment)
Track Seedingy (global z)
x (global f)
12
•Reduces number of seeds and tracking time by factor of ~2
•Loses 1.6% of tracks- quality of lost tracks is unknown as yet•No attempt to optimize cuts or use low-Q cut yet•New seeding in CMSSW 2_0: improvement smaller but still
significant
First Seeding Results (preliminary)D. Fehling, P. Maksimovic (JHU) have created a template-based seed cleaner that works with pixel-doublet seeds. The first test was done with a sample of 750 simulated t-tbar events:
Seed Generator0.13 s/event
Seed Cleaner
0.06 s/event
Kalman Filter1.80
s/event
1085k seeds
1085k seeds
476k seeds
37.6k tracks1.92 s/event
37.0k tracks1.15 s/event
Kalman Filter0.96
s/event
13
•Use 80-120 GeV PT QCD events
•Track counting doesn’t need re-calibration
- track probability also improves /wo calib
•Improvement in S/N is in range 2-3!
b-Tagging (preliminary)D. Fehling has studied the effect of the 2nd-pass template reco and templated-based seed cleaning+2nd-pass reco on b-tagging:
Standard RecoTemplate Reco Only
Template Seeding+Reco
b-efficiency
ud
sg-e
ffici
ency
14
Templates in Cluster Splitting•Template technique has only modest sensitivity to the
track angles- 1-2 mm shifts in cluster position do not affect resolution•Template probabilities flag unlikely cluster shapes/sizes- should avoid using the probabilities at the seeding level
✴ want to include “bad” hits on tracks (to associate merged clusters to tracks and get angle estimates)
•Current version of Template Technique works in two 1-D projections
- full 2-D templates are possible but don’t exist currently✴ very cpu intensive to generate✴ would be significantly slower (not usable for everyday
seeding)✴ no resolution advantage✴ would improve discrimination of template probability✴ would improve cluster splitting capabilities
➡The following is a sketch of a high pt jet re-tracking algorithm based on current 1-D cluster technology
15
•Step 2:- examine template
probabilities of tracked pixel hits✴ if small, try fitting two
hit hypotheses in both projections
✴ take the angles to be the same for both hits
✴ should improve template probabilities
✴ produces 4 new hits w/ 2-fold ambiguity (2-x X 2-y coords)
•Step 3:- re-track event w/ tighter
cuts
•Step 1:- first pass tracking with “loose” cuts on x2
✴ road search and/or✴ CTF with simple seeder✴ templates in second pass only
hit 2hit 1
How to Begin•Coding of a cluster splitter should be fairly
straightfoward:- 1-3 weeks for initial development/coding
(tuning/iterating could take longer)- initial testing with merged pixelav hits
✴ test code needs to be developed also- 2-hit hypothesis probability needs calibration
✴ add more info to the basic template infrastructure?•Need full re-tracking procedure•Testing splitting as part of a re-tracking procedure- need samples of problematic events- need diagnostics that identify the inefficiency and
resolutions