The Si D Particle Flow Algorithm List of
























- Slides: 24

The Si. D Particle Flow Algorithm

List of Contents • • Assume Particle Flow needs no introduction The Si. D 02 used in the PFA An Overview of the algorithm Status at the time of LOI Introspection of the Si. D-PFA The fix-ups and improvements Where we are Where to next (short and longer term)

The Detector (Si. D 02) ECAL: 30+1 layers of (320μm Si + 2. 5/5. 0 mm W), 3. 5 x 3. 5 mm cells. HCAL: 40 layers of (1. 2 mm RPC + 2 cm steel), 1 cm x 1 cm cells. 3

Basic Building Blocks of the (Iowa) PFA • • • MC hits within 100 ns from IP are digitized Photon, Muon and Electron ID Track and Seed Cluster (Directed Tree) Building Charged Hadron Shower Reconstructing Particles (four-vectors) Hits belonging to Photon, Muon and Electron are removed from the hit list for clustering algorithm Next Use Directed. Tree Clustering for classifying the remaining hits into sub-cluster types like MIPs, Clumps, Blocks and `leftover’s Finally, start building Hadron Showers for one charged track at a time 4

Cluster Building • Extrapolate (each) track to the ECAL surface • Find Seed: sub-cluster directly connected to extrapolated track • Each track typically has one seed Special cases: track without seed, or does not reach calorimeter • Now start connecting other sub-clusters to the seed of each track • Start with lowest and then progressively higher momentum tracks • Up to ten iterations until all track-cluster match satisfy (E – p) within tolerance Connecting Clusters Scoring : (a poor man’s) Probability of a link Based on the sub-cluster type and geometric proximity a score between 0 and 1 is assigned between any two sub-clusters starting with the cluster in consideration The higher the score the higher the probability of a link A cut-off threshold is obtained for an energy by tuning with events 5

Energy dependence Performance at LOI Study just how much is contributed because of leakage 6

Leakage study at 500 Ge. V and 1 Te. V Produce data sets a Si. D 02 -like detector MC with 6 HCAL for 1 Te. V, 500 Ge. V, 200 Ge. V • Change Steel for Cu for absorber • Increase to 54 layers from 40 layers in HCAL • 1. 7 more material in HCAL • No gap between HCAL and Muon endcap (instead of 10 cm) Compare sid 02 with sid 02 -Cu at various energies Check leakage by observing # hits in Muon detector : punch thru; a measure of leakage Simultaneously study the corresponding change in Energy resolution The relative measure from the two gives an approximate semi-quantitative measure of leakage vs performance Although substantial leakage is present at 500 Ge. V confusion is clearly important 7

Punch-through muon hits Si. D 02 -Cu Si. D 02 8

Resolution study (Si. D 02 -Cu comparison) real tracking Si. D 02 -Cu Si. D 02 9

Conclusions from Leakage study • At 1 Te. V leakage comparison shows large difference in performance between Si. D-nominal (dashed) and Si. D-Cu detectors (solid) • At 500 Ge. V leakage comparison shows significant difference in performance between Si. D-nominal (dashed) and Si. D-Cu detectors (solid) • Performance of 1 Te. V Si. D-Cu is similar to 500 Ge. V Si. D-nominal in leakage • At 1 Te. V performance in resolution is worse with Si. D-nominal (dashed) and Si. D-Cu detectors (solid) • At 500 Ge. V performance in resolution is worse with Si. D-nominal (dashed) and Si. D-Cu detectors (solid) • However : The difference of performance in resolution between 1 Te. V Si. D-Cu and 500 Ge. V Si. D-nominal is not similar to that in leakage Although substantial leakage is present at 500 Ge. V, algorithm (confusion) has an important part 10

A 500 Ge. V qqbar event from one side jet Raw MC hits are displayed, each color shows an individual shower Contains a low energy 12 Ge. V neutral hadron and several photons in the ECAL; charged hadrons interacts

reconstructed The same as before shown without Now shown without the isolated hits the isolated and unmatched hits : but after reconstruction, alogorithm still no PFA reconstruction, only of charged hadron track-cluster with knowledge of MC match (cone algorithm) p (orange) = 119 Ge. V, E/p match, enough hits (green) = 17 Ge. V , algorithm introduced a cone-like path in the reclustering to pick up secondary neutrals; but ended up being too aggressive in stealing pieces from

has a low energy 12 Ge. V neutral hadron and several photons present in the ECAL; interaction of charged hadron p (orange) = 119 Ge. V, E/p match, enough hits (green) = 17 Ge. V reconstructed Refined. Cheat. Cluster Diagnosis of `A’ problem: an example Refined. Cluster - sharedhits Had introduced a cone-like path in the reclustering to pick up secondary neutrals; but ends up being too aggressive 13

The `Cone’ Algorithm Cluster Dir. Angle DCA IP Seed Pos. Angle Interaction point

A detailed Study

All plots show variables defined for links between a seed and a cluster. If the seed and the cluster belong to the same truth particle, the link is quoted as “Signal” otherwise it is quoted as “Background” Top-Left Plot: Scores just before the First cone algorithm runs. Top-Middle Plot: Scores just after the First cone algorithm runs. Top-Right Plot: Impact Parameter (IP): Distance between the center of the seed and the straight line from the center of the cluster extrapolated along the cluster’s direction Bottom-Left Plot: Distance of closest approach (DCA) between two straight lines taken respectively from the center of the seed and the center of the cluster and along the respective directions. Bottom-Middle Plot: Angle at the interaction point formed by the positions of the seed and the cluster. Bottom-Right Plot: Angular difference between the direction of the seed and the direction of the cluster.

Score disteribution for links when the first cone algorithm modifies the score. Left plot: Scores before the first cone algorithm. Right plot: Scores after the cone algorithm. While Signal/Background discrimination is better after the first cone algorithm, backgrounds now peak in the Signal region.

Correlated Variables now zoom on signal region: look at links when the first cone algorithm gives a high score (>0. 8).

Sharing of hits: Breaking up into smaller clusters Extending to smaller pieces

Next Steps Allow flexibility in assignment of hits in clusters from tracks in the vicinity; Allocate after arbitration Check where exactly the `cone’ is needed, modify this, dump the rest Wait for results from ongoing study here…. Faster turn around time Improved resolution Next major step : Incorporate the PFA with realistic Si. D (Si. D 03) geometry Now progressing in parallel Expect to take a step backward: non-trivial Improve sophisticated modifications for special types of clusters, like backscattering, complex rare occurrences

Particle Flow • Energy of a hadronic jet in a calorimeter Ejet = E ( + 0) + Ehadronic (neglecting ’s and leakage etc) Electromagnetic and hadronic components have different responses Solutions: compensating calorimetry, measure hadronic and EM separately…. . • However: Ejet = Ephotons + Eneutral-hadrons + Echarged-hadrons Obtain Charged hadron energy ( 60 ) from tracking Obtain photon/EM energy ( 30 ) from ECAL with 19 / E resolution Get neutral hadron energy ( 10 ) from E/HCAL with 67 / E resolution • Therefore the jet energy resolution is Ejet Echarged Ephotons 0 19 / (0. 3 x Ejet) 20 / E fantastic ! Eneutral hadrons 67 / (0. 1 x Ejet)

Particle Flow contd • The concept depends on ability to measure particles independently • Charged and neutral particle confusion degrades resolution Ejet Ephotons Eneutral hadrons confusion The confusion should be minimized in a good PFA Need excellent pattern recognition (also high granularity and low occupancy) • sid 02 : ECAL: 30+1 layers of (320μm Si + 2. 5/5. 0 mm W), 3. 5 x 3. 5 mm cells. HCAL: 40 layers of (1. 2 mm RPC + 2 cm steel), 1 cm x 1 cm cells.

Categorizing: Directed. Tree Clustering 23

Final Clustering : a flow for each track 24