Post on 18-Jan-2016
PFAs – A Critical LookPFAs – A Critical Look
Where Does (my) SiD PFA go Where Does (my) SiD PFA go Wrong?Wrong?
S. R. MagillANL
ALCPG 10/04/07
A Systematic PFA Development
Starting Point :
100% pure calorimeter cell population – 1 and only 1 particle contributes to a cellMore practically, no overlap between charged particles and neutrals-> Defines cell volume – v(dIP,η,B?)-> Start of detector design optimization-> Perfect PFA is really perfect – no confusion to start
100% pure tracker hits (or obvious crossings)-> Defines Si strip size-> Start of design optimization-> Perfect Tracks are really perfect
PFA is an intelligent mixture of high purity and high efficiency objects – not necessarily both together
Occupancy Event DisplayOccupancy Event Display
All hits from all particles
Hits with >1 particle contributing
Defining Perfect PFA (Perfect Reconstructed Particles)
Takes generated and simulated MC objects, applies rules to define what a particular detector should be able to detect, forms a list of the perfect reconstructed particles, perfect tracks, and perfect calorimeter clusters.
Complicated examples :-> charged particle interactions/decays before cal-> photon conversions-> backscattered particles
Critical for comparisons when perfect (cheated) tracks are usedExtremely useful for debugging PFAChecks detector calibration procedure
Perfect TracksPerfect Neutrals (photons, neutral hadrons)Perfect Cal Clusters
SiD (SS/RPC)
Photons from Perfect PFA (ZPole events in ACME0605 W/Scin HCAL)
/mean ~ 18%/E
/mean ~ 18%/E
/mean ~ 24%/E
Detector Calibration Check
24%/E
18%/E
18%/E
Neutral Hadrons from Perfect PFA (ZPole events in ACME0605 W/Scin HCAL)
/mean ~ 47%/E
/mean ~ 49%/E
/mean ~ 46%/E
46%/E
49%/E
45%/E
Track/CAL Shower Matching
This is an example of where high purity is preferred over efficiency
-> will discard calorimeter hits and use track for particle-> better to discard too few hits rather than those from other particles-> use hits or high purity cluster algorithm
Example :
1) Associate mip hits to extrapolated tracks up to interaction point where particle starts to shower.
-> ~100% pure association since no clustering yet-> tune on muons to get extra hits from delta rays
2) Cluster remaining hits using high purity cluster algorithm – Nearest Neighbor with some fine tuning for neighborhood size
-> iterate, adding clusters until ΣEcl/ptr in tunable range (0.65 – 1.5)-> can break up cluster if E/p too large (M. Thomson)-> err on too few clusters – can add later when defining neutral hadrons
ECAL Radius 175 cm (LDC-like)
ESum of Charged Hadron Clusters per event
Decent match of PFA to Perfect
Candidates for re-clustering?
ECAL Radius 175 cm (LDC-like)
ECAL Radius 125 cm (SiD)
PFA slightly larger than Perfect
Candidates for re-clustering?
ECAL Radius 125 cm (SiD)
Now, high efficiency is desired so that all photons are defined – can optimize for both high efficiency and high purity by using multiple clustering.
Example :
1) Cone or DT cluster algorithm (high efficiency) with parameters :
radius = 0.04seed = 0.0minE = 0.0
2) Cluster hits in cones with NN(1111) to define cluster core (high purity for photons)
mincells = 20 (minimum #cells in reclustered object)dTrCl = 0.02 (no tracks within .02)
3) Test with longitudinal H-Matrix and evaluate χ2
Other evaluations are done in PhotonFinderDriver – like layer of first interaction if cluster fails mincells test, cluster E in HCAL, etc.
Photon Finding
Photon Energy (GeV)
0 5 10 15 20 25 30 35 40
0.660.680.700.720.740.760.780.800.820.840.860.880.900.920.940.960.981.001.021.04
DTPhotonAnalysis204.aida - DTEcalClusters
PhotonAnalysis929.aida - NN663EcalClusters
PhotonAnalysis929.aida - NN442EcalClusters
PhotonAnalysis929.aida - NN111EcalClusters
PhotonAnalysis929.aida - MSTEcalClusters
Photons - Found photon cluster efficiency: bin 1
Photon Energy (GeV)
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 300.45
0.50
0.55
0.60
0.65
0.70
0.75
0.80
0.85
0.90
0.95
1.00
DTPhotonAnalysis204.aida - DTEcalClusters
PhotonAnalysis929.aida - NN663EcalClusters
PhotonAnalysis929.aida - NN442EcalClusters
PhotonAnalysis929.aida - NN111EcalClusters
PhotonAnalysis929.aida - MSTEcalClusters
Photons - Found photon cluster purity: bin 2
Photon Clustering
Photon Energy (GeV)
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 300.45
0.50
0.55
0.60
0.65
0.70
0.75
0.80
0.85
0.90
0.95
1.00
DTPhotonAnalysis204.aida - DTEcalClusters
PhotonAnalysis929.aida - NN663EcalClusters
PhotonAnalysis929.aida - NN442EcalClusters
PhotonAnalysis929.aida - NN111EcalClusters
PhotonAnalysis929.aida - MSTEcalClusters
Photons - Found photon cluster purity: bin 2
Photon Comparisons – ZPole events for ACME 0605 (SiD) detector
On the left is the ESum of photons for each event, red is perfect and blue is from PFA – 20.1 GeV for perfect PFA compared to 19.2 GeV for real PFA. On right is the difference between real PFA and perfect PFA for each event, a -1 GeV difference on average per event.
Photon Comparisons – ZZ @ 500 GeV events for ACME 0605 (SiD) detector
On the left is the ESum of photons for each event, red is perfect and blue is from PFA – 49.8 GeV for perfect PFA compared to 30.0 GeV for real PFA. On right is the difference between real PFA and perfect PFA for each event, now a -20 GeV difference on average per event. From the plot on right, you can see that it’s the high energy photon events that are missing – probably consistent with the H-Matrix failing on large cluster (high E) photons or from overlapping clusters.
Photon Comparisons – ZPole events for ACME 0605, ECAL IR = 175 cm (LDC-like)
On the left is the ESum of photons for each event, red is perfect and blue is from PFA – 20.5 GeV for perfect PFA compared to 19.8 GeV for real PFA. On right is the difference between real PFA and perfect PFA for each event, a -0.75 GeV difference on average per event. As expected, this is slightly better than the smaller detector – probably less overlap.
Photon Comparisons – ZZ @ 500 GeV events for ACME 0605, ECAL IR=175 (LDC-like)
On the left is the ESum of photons for each event, red is perfect and blue is from PFA – 49.9 GeV for perfect PFA compared to 31.2 GeV for real PFA. Very similar to the ~60% as before for ZZ events (slight improvement). On right is the difference between real PFA and perfect PFA for each event, now a -19 GeV difference on average per event. From the plot on right, you can see again that it’s the high energy photon events that are missing – probably consistent with the H-Matrix failing on large cluster (high E) photons or from overlapping clusters. Very slight improvement – needs to be investigated further.
Neutral Hadron ID
Here again, high efficiency is desired – if previous algorithms have performed well enough, purity will not be an issue.
Example :
Cluster with Directed Tree (another high efficiency clusterer)-> clean fragments with minimum cells-> check distance to nearest track – if too close, discard-> merge remaining clusters if close
Needs additional ideas, techniques – pointing?
AugustAugust SeptembSeptemberer
OctoberOctober NovembeNovemberr
DecembeDecemberr
JanuaryJanuary
ToolsTools DigiSimDigiSim
CalibrationCalibration
Perfect Perfect PFAPFA
PhotonsPhotons
dM/M vs dM/M vs ESum ESum (PPFA)(PPFA)
PFA Res.PFA Res. ZZ->qqZZ->qq ZZ->qqqqZZ->qqqq
tt->6 jetstt->6 jets
dM/M vs dM/M vs EjetEjet
ZH->4 jetsZH->4 jetsB-field, B-field, ECAL IR, ECAL IR, HCAL Tech., HCAL Tech., Det. Det. variantsvariants
e+e- @ 1 e+e- @ 1 TeVTeV
Understand contribution to dijet mass resolution from energy, angles
Results of PFA performance at 500 GeV CM for various n-jet processes- Performance vs Ejet or Ejetjet?- Detector parameter variations- Results for F’lab ALCPG
Ferm
iLab ALC
PG