The document summarizes studies on clustering very low energy particles using the ATLAS calorimeter. It discusses using topoclusters with different seed and neighbor cell energy thresholds to better reconstruct particles below 10 GeV. Preliminary conclusions found that a seed threshold of 4 and neighbor threshold of 2 provided the best energy resolution and efficiency for pions, photons, and neutrons compared to other clustering algorithms. Further studies examined the impact of overlapping nearby particles on cluster reconstruction and found the new splitter algorithm in release 8.2.0 did not significantly improve resolution over not using splitting for particles separated by 0.1 or more in deltaR or below 0.1.
Introduction to ArtificiaI Intelligence in Higher Education
Clustering of very low energy particles
1. Clustering of very
low energy particles
Carmen Iglesias
IFIC-University of Valencia
Dpt. Fisica AMN
ATLAS Calorimetry Calibration Workshop, Štrba, Slovakia Dec 1 - 4, 2004
2. INDEX
Summary of the last analysis
Topoclusters in 8.2.0
Study the overlapping
Preliminary Combined TB analysis
Tile Noise = 70 MeV, LArEM Noise = 70 MeV
Tile Noise = 25 MeV, LArEM Noise = 70 MeV
Very preliminary!!!
Seed_cut=4, Neigh_cut=2
CaloNoiseTool aplied
3. Summary of the last
analysis
Seed Cell
phi
Neighbour Cell
eta
4. Summary last analysis (I)
Use samples of single particles at very low Et (VLE), (because this is the range of
energy better to apply Energy Flow Algorithm) at eta=0.3 and phi=1.59:
π0’s of 1, 3, 5, 10 and 30 GeV
Neutrons of 1, 3, 5, 10 and 30 GeV
π+’s of 1, 3, 5, 10 and 30 GeV
in order to generated ntuples with 1000 events.
Start with these thresholds for Seed Cell and Neighbour cells of the Topocluster:
Seed: E_cell/σ_noise=30 ->usual cut for particles of normal ET
Neighbour: E_cell/σ_noise=3
Compare different thresholds of the Noise in EM for the reconstruction of
CaloTopoCluster of VLE particles:
EM Noise=10 MeV (Non realistic case, only to check with VLE particles)
EM Noise=70 MeV (Fix Value by default for EM cal in Sven’s code)
CaloNoiseTool=true (package with a model for the electronic noise)
Repeat the analysis for lower threshold of Seed Cell and Neighbour cells :
Seed: E_cell/σ_noise=6, 5, 4 -> cut for particles of very low ET
Neighbour: E_cell/σ_noise=3, 2.5, 2
5. Summary last analysis (II)
First, calculate the ET deposited in all CELLs of the calorimeter and consider it as
the “reference Energy Flow”, i.e., the best resolution that could be reach for the
most sophisticated algorithm taking into account the whole ET in all the calorimeter.
For π0’s, compare the resolution of “reference Energy Flow” with the resolution of:
Sliding Window Cluster
EGAMMA cluster
TOPOcluster in EM calorim: XXX_topoEM
For π+’s and neutrons, compare the resolution of “reference Energy Flow” with :
TOPOcluster in EM and Tile: XXX_topoEM and XXX_topoTile (7.8.0 release)
PT of TRACKS from XKalman
Finally, study the ET inside a cone with a radius ∆R=√∆η2+∆φ2
Neutral pions
Cone’s centred in η-φ coord of EGAMMA cluster
Cone’s centred in η-φ coord of TOPO cluster in EM cal
Cone’s centred in η-φ coord of TRUTH generated π0
Charged pions
Cone’s centred in η-φ of TRUTH generated π±
Cone’s centred in η-φ of TRACK position at 2nd layer
Neutrons
Cone’s centred in η-φ of TRUTH generated neutrons
6. Summary of the preliminary conclusions
In TOPOCLUSTER using CaloNoiseTool package and applying:
Seed_cut: E_cell/σ_noise = 4
Neighbour_cut: E_cell/σ_noise = 2
obtain the best values of resolution for π+'s, π0's and neutrons (Even better than
EGamma resolution for π0’s)
eliminate the low efficiency of TOPOclusters for single particles at 1, 3, and 5GeV.
obtain the largest amount of ET deposited inside the TOPOclusters (Even larger
than the ET deposited insider EGamma clusters for π0’s)
The best algo for the recon of the clusters from single particles at very low ET is:
For π±’s: Track-cone with ∆R<0.2
TOPO give good results but worse than the two cone algorithms .
For neutrons: Truth-cone with ∆R<0.2
TOPO with Seed_cut=4 and Neigh_cut=2 is very near and it’s better at 1 and 3 GeV.
For π0’s: Truth-cone with ∆R<0.1
but TOPO is very near and it’s better at 1 GeV.
EGAMMA give worse resolution, in general, is a better algo at 1 GeV
TOPO algorithm is very competitive for neutrons and π0’s, for π±’s TOPO is a good
algo but not enough, for now (I must test 8.2.0 release still)
8. Release 8.2.0
Reconstruction Packages
RecExample-00-00-94
RecExCommon-00-02-09
RecExTB-00-00-27
Calorimeter Packages
CaloRec-02-02-19
Main Changes in this Release:
“super-cluster”: one cluster for all calorim (EM+HEC+FCal+TCal=1region)
Splitter: Split clusters based on topological neighboring and E density defined
local maxima.
The cells in a cluster are searched for local maxima by means of energy density.
The so found local maxima are used as seeds for a topological clustering as in
the CaloTopo-ClusterMaker.
Information about cells wich form the TOPOcluster
I can analyse the ET of the cells inside TOPOcluster I can study
the best cut in ET of cell without and with noise.
Can use macro ROOT
9. TopoClusters in 8.2.0
ET Resolution and Mean of ET_Cluster/ET_gen are better than in 7.8.0 Release, even
when the noise is applied. This improvement is larger at very low ET
π+’s ET resolution Mean of ET_Cluster/ET_gen
ET (GeV) 7.8.0 8.2.0 % improv 7.8.0 8.2.0 % improv
1 56.30 44.57 26.3 0.528 1.05 49.7
3 45.65 38.87 17.4 0.5807 0.8708 41,9
5 33.82 30.65 10.3 0.6522 0.8412 22.5
10 22.41 20.84 7.5 0.7613 0.864 11.9
30 13.91 13.36 4.1 0.8128 0.8515 4.9
neu ET resolution Mean of ET_Cluster/ET_gen
ET (GeV) 7.8.0 8.2.0 % improv 7.8.0 8.2.0 % improv
1 64.31 --- 0.357 49.7
3 57.54 46.19 24.6 0.406 0.681 40.4
5 40.41 34.04 18.7 0.498 0.687 27.5
10 23.98 22.43 7.5 0.643 0.749 14.9
30 12.39 11.94 4.1 0.770 0.811 4.1
10. Study the overlapping
Efficiency of algorithm is limited by the overlap between
neutral and charged particles in the cell of the calorimeter
Use samples of π±’s, neutrons
and π0’s
(instead of only single particles)
Estimate the existence of overlapping (a neutral hadron inside the cluster defined
from the track of the charged pion) and its influence in the ET resolution
Particles far away in ∆R space
Particles closer in ∆R space (∆R=0.1)
Particles closer in ∆R space below 0.1
Generate Root-tuples using:
8.2.0 (without Splitter)
8.2.0 (with Splitter) compare the multiplicity and the ET resolution
Do plots with Macro ROOT and see the maximum density points
11. Samples (nearby particles)
Multiparticles samples based on DC1 events simulated with electronic noise (but without
pile-up) in the barrel region (|η|<1.5)
η=0.35 and φ=1.63
In general Particles far away in ∆R space
Pi0pim55: π0=5GeV π+-=5GeV Pi0pimneu10hfar: π0=10GeV π-+=10GeV neu=10 GeV
Pi0pip28: π0= 2GeV π+=8GeV Pi0pimneu10hfareta2
Pi0pip82: π0= 8GeV π+=2GeV Pi0pimneu5hfar: π0=5GeV π+=5GeV neu=5GeV
Pi0pimneu5hfareta2
Particles closer in ∆R space (∆R=0.1)
Pi0pimneu10d1010: π0=10GeV π+-=10GeV neu=10 GeV
Pi0pimneu555d1010: π0=5GeV π+-=5GeV neu=5GeV
Froidevaux’s EFlow results: as long as particles not closer than ∆R=0.1, Eflow performance
is stable…(but still needs improvement!)
Particles closer in ∆R space below 0.1
Pi0pim77d05: π0=7GeV π+-=7GeV in a distance ∆R=0.05
Pi0pimneu555d0505: π0=5GeV π+-=5GeV neu=10 GeV at ∆R=0.05 and ∆R=0.05
Pi0pip1010d05: π0=10GeV π+-=10GeV in a distance ∆R=0.05
Pi0pip55d07: π0=5GeV π+=5GeV in a distance ∆R=0.07
Froidevaux’s EFlow results: as ∆R between particles decrease below 0.1, Eflow response
Degrades and develops significant tails (there is were the full use of ATLAS calor granularity,
lateral and longitudinal, will surely provide improvements)
12. ROOT plots: Particles far away in ∆R space
Pi0pimneu10hfar (N0 Splitter)
Is possible to distinguish 3 cluster from ECAL Middle layer up to Tile 2
13. ROOT plots: Particles closer in ∆R space (∆R=0.1)
Pi0pimneu10d1010
Is possible to distinguish 3 cluster but only at the end of ECAL and in TILE
14. ROOT plots: Particles closer in ∆R space below 0.1
Pi0pip1010d05 (N0 Splitter)
The two cluster are very closevery difficult to distinguish!!
15. ET resolution
Calculate resolution with and without Splitter Cluster applied in Topoclusters:
Particles far away in ∆R space
ET Resolution Mean of ET_Cluster/ET_gen
No Splitter Splitter No Splitter Spliter
pi0pimneu10hfar 10.26 10.21 0.8171 0.8185
pi0pimneu5hfar 14.97 14.70 0.7535 0.7543
Particles closer in ∆R space (∆R=0.1)
No Splitter Splitter No Splitter Spliter
pi0pimneu10d1010 9.34 9.32 0.8282 0.8287
pi0pimneu555d1010 13.96 14.17 0.7715 0.7717
Particles closer in ∆R space below 0.1 (∆R=0.05 and ∆R=0.07)
No Splitter Splitter No Splitter Spliter
pi0pim77d05 11.87 11.89 0.8969 0.8975
pi0pimneu555d0505 14.51 14.55 0.7786 0.7801
pi0pip1010d05 9.04 9.00 0.9093 0.91
pi0pip55d07 14.77 14.74 0.8926 0.8944
16. First conclusions
Very similar results!!, even in the case of particles generated to be very
close (∆R=0.05 and ∆R=0.07)
EtDensityCut: All local maximum candidates must also pass it, specifying the
min Et divided by the volume of the cell in order to be accepted as local max.
The default value corresponds to Et= 500 MeV in a LArEM barrel layer 2 cell
EtDensityCut = 500*MeV/600000*mm3
Very low energetic particles don't produce big local maxima as I I lowered the
Seed_cut for topocluster maker I might to try lowering the seed cut in the splitter
by the same factor
I going to investigate the deposited ET in a LArEM barrel layer 2 cell by
Charged Hadrons at VLE
Neutral Hadrons at VLE
17. ET deposited in EM cell of 2nd layer
Charged pions
Neutral pions
the mean ET deposited in cell of EM on the 2nd layer for π±’s and π0’s at very low
ET is ~280 MeV. So I'm going to change the EtDensityCut from the default value
(500*MeV/600000*mm3) to 250*MeV/600000*mm3.
18. ET resolution with Splitter
BUT the value of ET Resolution using the default EtDensityCut and
250*MeV/600000*mm3 in the 3 cases:
Particles far away in ∆R space
Particles closer in ∆R space (∆R=0.1)
Particles closer in ∆R space below 0.1 (∆R=0.05 and ∆R=0.07)
are identical !!!
Even when I try to repeat the ROOT plots, I obtain the same plots!!
so :
From my results(*), I only can conclude that Splitter cluster can not be used to
study the overlap of particle at very low energy (only useful for energetic particles)
(*) I don’t think that it will be a code fault because I was talking with S. Menke about the way to
change EtDensityCut value in the code (in CaloRec/share/CaloTopoCluster_jobOptions.py)
20. Offline reconstruction root-tuple
CALO/168 contains the eta, phi, energy (in MeV) per LAr cell
nhit: total number of cells ecell: total energy EtaCells: eta
NCells: number of cells in ntuple PhiCells: phi
ECells: energy (MeV) DetCells: pseudo identifier
CALO/169 contains the same for Tile cell
TB/Tree contains the information about clusters
we have, for the time being, 4 times the final number of variables in the ntuple :
em and combined clusters (sliding window),
topo_em and topo_tile clusters.
Emcluster: clusters from the sliding window algorithm
Tbemclusters: clusters from an algorithm that have been used in previous
test beam. It has been added to allow comparison. It is a window of 3 by 3 cells.
Emclusters and tbemclusters use only cells from the Larg calorimeter.
Cmbclusters: sliding window clusters but they are done on towers (larg+tile)
and not anymore on cells. It is not working for the moment because of a coordinate
problem between LAr and Tile, LAr is shifted with respect to Tile by "half module" :
TileCal has just 3 modules -0.15 < eta < +0.15
LAr has -0.2 < eta < 0.2,
i.e. there are 3 slices with ∆φ=0.1 in Tile and 4 slices in LAr, shifted by half of the slice
topo_em & topo_tile clusters:
22. ET Resolution
η= 0.20 run EM_SW TB_EM_SW TOPO_Em TOPO_Tile TOPO_total
9 GeV 2101022 8.91 7.26 9.44 70.49 11.54
8 GeV 2101024 10.09 8.15 10.80 57.34 13.93
7 GeV 2101086 11.73 9.11 12.36 46.03 16.56
6 GeV 2101025 13.47 10.35 14.24 41.87 38.50
5 GeV 2101026 17.01 12.64 18.40 35.17 39.88
3 GeV 2101087 16.19 22.34 71.12 23.80 33.33
2 GeV 2101088 14.15 73.84 70.15 22.84 40.20
1 GeV 2101089 ---- 63.14 59.76 ---------- -------
Sliding Window give an slightly better resolution than Topo
Very strange plots at 1 GeV: may be due to the low number of cluster well-defined
Very bad result at very low energy (3, 2 and 1 GeV) for TOPO_Em
For TOPO_Tile the ET resolution decrease with the energy instead of increasing!!!
There is a big lost of energy in Tile from 9 to 5 GeV, but from 3 GeV it decrease
23. Mean Value of ET_cluster/ET_generated
run EM_SW TB_EM_SW TOPO_Em TOPO_Tile TOPO_total
9 GeV 2101022 1.108 1.015 1.048 0.221 1.061
8 GeV 2101024 1.076 0.9841 1.012 0.2588 1.029
7 GeV 2101086 1.065 0.974 0.9978 0.29 1.011
6 GeV 2101025 1.053 0.9622 0.981 0.3479 0.9106
5 GeV 2101026 1.021 0.9341 0.9393 0.4259 0.7064
3 GeV 2101087 1.187 0.8738 0.5521 0.6884 0.8787
2 GeV 2101088 1.629 0.3937 0.4946 1.029 1.144
1 GeV 2101089 ---- 0.6067 0.7152 ---------- -------
For Sliding Window EM seems as the cluster energy is larger than the generated
energymay be due to double counted??
Electron contamination distort the pions distribution for very low ET pions
There is a big lost of energy in Tile
24. Preliminary Results
(TileNoise=25.0MeV)
From CIS analysis Richard claims that we
have 1.6 ADC counts per channel which
gives:
1.6/1023*800/64/1.1*sqrt(2) = 0.025 GeV -
electronics noise per cell
26. What is happening in TILE?
Changing TileNoise from 70 to 25 MeV increase the bad defined topocluster with
low energy (lower than ET generated) and its distorts the ET distribution. It will be
needed to apply new cuts!!!
Why it happens only at larger ET?
Cell_Cut = |E/σ noise|>Threshold so as σ noise has decrease to 25MeV, at larger E
as 7, 8 and 9 GeV more cluster can pass the threshold.
27. What is happening in LArEM?
In SW TB EM and TOPO_Em the distribution of ET resolution has tails in left side.
These tails decrease with the ET of the particles. WHY?
9 GeV 3 GeV
Tails means cluster with ET lower than generated ET. So clusters has been
reconstructed not completly.
Seed_Cut = |E/σnoise|>Tcell so as E decrease it’s more difficult for the Seed_cell
pass the threshold, but decreasing Tcell we’ll too much not real clusters.
Neighbor_Cut = |E/σnoise|>Tneigh , again as E decrease it’s more difficult for the
Neighbor cells pass the threshold and for this reason the cluster is not reconstructed
So, if we have reduce the tails, we must put:
longer (lightly) the Seed_cut
lower the Neighbor_Cut
28. Electrons TileNoise=25.0MeV, Seed_Cut=4,
Neigh_Cut=2
η=0.40 run SW EM SW TB EM TOPO_Em (*) with a peak near 0
9 GeV 2101204 9.26 8.27(*) 11.10(*)
8 GeV 2101057 10.56 8.90 (*) 12.52 (*)
5 GeV 2101200 15.50 13.56(*) 19.82(*)
- The best resolution from SW TB EM, i.e., clusters from an algorithm used in previous
TB (3x3), but it has peaks near 0 (cluster with very low ET respect to the generated ET)
- Clusters from the sliding window algorithm SW EM has not these peaks.(???)
9 GeV
8 GeV
The peaks near 0 values decrease with the ET of the particles
29. Next Steps:
Sort time plans: reconstruct CombinedTB data using CaloNoiseTool. (For this
Workshop was imposible for me )
In future TestBeam release this tool will be available. (person in charge:
Matthieu LECHOWSKI)
May be Preliminary results about it nex week during SW workshop at
CERN???
Change the Threshold value for Seed_Cut and Neigh_Cut in order to validate my
hipothesis
"Clustering of very low energy particles" ATLAS NOTE in preparation
More Analysis with Combined TB real data
Comparation with Combined TB Simulation data
…Complete Clustering analysis for Rome Physics Workshop