A Particle Flow Algorithm for SiD Mat Charles
A Particle Flow Algorithm for SiD 1. What is SiD? Mat Charles
A Particle Flow Algorithm for SiD 1. What is SiD? 2. What is a PFA, and why does SiD want one? Mat Charles
A Particle Flow Algorithm for SiD 1. What is SiD? 2. What is a PFA, and why does SiD want one? 3. How does this PFA work? Mat Charles
A Particle Flow Algorithm for SiD 1. What is SiD? 2. What is a PFA, and why does SiD want one? 3. How does this PFA work? 4. Is it any good? Mat Charles
Upgrade to higher energy anticipated. 2 What is SiD? SiD is a detector design for an e + e linear collider. Nobu Toge, ICHEP08 Nominal ILC CM energy of 500 GeV.
What is SiD? SiD is a detector design for an e + e linear collider. Relatively small / compact (total radius ~ 6m, magnet inner radius ~ 2.5m) Silicon-based tracking 5T solenoidal B-field Has the usual general-purpose detector subsystems: muon system solenoid HCAL muon system ECAL Tracker Vertex detector HCAL 3
What is Particle Flow? For hadronic jet (neglecting neutrinos, leptons, leakage, etc): Ejet = Ephoton showers + EEM part of hadronic showers + Ehadronic part of hadronic showers Measuring Ejet is hard because hadronic & EM responses differ. Solution 1: Make their responses the same (compensation) Solution 2: Measure hadronic & EM components of showers separately Solution 3: Use things other than calorimeter Write the jet energy another way: Ejet = Echarged hadrons + Ephotons + Eneutral hadrons Get charged hadron energy (~60%) from tracking system Get photon energy (~30%) from ECAL with 19% E resolution (for SiD) Get neutral hadron energy (~10%) from ECAL+HCAL with 67% E resolution Then jet energy resolution is: σejet = σecharged σephotons σeneutral hadrons 0 19% (0.3 Ejet) 63% (0.1 Ejet) 20% Ejet Looks fantastic on paper! What s the catch? 4
Particle Flow Whole concept relies on measuring particles individually. Confusion of charged & neutral deposits degrades resolution: σejet = σephotons σeneutral hadrons σconfusion Need low occupancy and good pattern recognition (PFA). sid02 ECAL: 30+1 layers of (320μm Si + 2.5/5.0mm W), 3.5 x 3.5mm cells. sid02 HCAL: 40 layers of (1.2mm RPC + 2cm steel), 1cm x 1cm cells. HCAL ECAL This is not your father s calorimeter! 50 GeV s-quark jet, sid02 Showing PFA output: White: reconstructed tracks. Green lines: reconstructed neutrals. Calorimeter hits colour-grouped by reconstructed particle. Shared hits not shown. 5
6 PFA in practice Start with the easiest parts:
PFA in practice Start with the easiest parts: Muons leave a distinctive trail of isolated or semi-isolated hits. HCAL Muon system Tracker 10 GeV μ ECAL HCAL 6
PFA in practice Start with the easiest parts: Muons leave a distinctive trail of isolated or semi-isolated hits... and so do charged hadrons before they shower. 10 GeV π + ECAL MIP segment HCAL 6
PFA in practice Start with the easiest parts: Muons leave a distinctive trail of isolated or semi-isolated hits... and so do charged hadrons before they shower. EM showers start early in ECAL, have tight core & characteristic longitudinal profile. 10 GeV γ 6
PFA in practice Start with the easiest parts: Muons leave a distinctive trail of isolated or semi-isolated hits... and so do charged hadrons before they shower. EM showers start early in ECAL, have tight core & characteristic longitudinal profile. Electrons are EM showers matched to a track with the Eshower = ptrack 10 GeV e 6
PFA in practice Start with the easiest parts: Muons leave a distinctive trail of isolated or semi-isolated hits... and so do charged hadrons before they shower. EM showers start early in ECAL, have tight core & characteristic longitudinal profile. Electrons are EM showers matched to a track with the Eshower = ptrack Even with these easy cases, we have to take care, e.g.: A photon and a MIP (or...) can overlap. An early hadronic shower with a large EM component looks a lot like a photon (hence photon purity only 90%). Not all photons look like photons (e.g. merged π 0 ) -- hence photon ID efficiency only 90%. Bremstrahlung photons from electrons 6
PFA in practice Now comes the hard part: hadronic showers. Large variations in shape & size from shower to shower Internal structure is large and clearly visible If EM showers are like raindrops, hadronic showers are snowflakes. Instead of trying to cluster whole shower in one go, build it up in pieces Meccano-style: MIPs & track segments Dense clumps Other large clusters Leftover hits & small clusters Then use geometrical information to put the pieces together: pointing, proximity, angle, etc. Track (10 GeV π + ) Track segments Clumps Leftover hits 7
PFA in practice Building charged hadronic showers iteratively: 1. Loop over tracks in order of increasing momentum: a) Start from seed cluster (often MIP) connected to track. b) Calculate score for nearby clusters to be connected to seed. c) Add clusters that look well-connected (score > threshold). d) Repeat process, building shower outwards. e) Stop if no good connections or if energy too big (Eshower > ptrack + nσe) 2. Adjust score threshold or (E-p) tolerance if needed. 3. Identify & flag overlapping showers. 4.Go back to step 1 unless converged. 5. Final passes to fix mistakes (shower-cluster matches that failed earlier) Fuzzy clustering used for leftover hits & secondaries whose assignment is not clear. Similar (but simpler) algorithm for neutral hadrons. 8
9 Is the PFA any good? What performance do we need? Real answer will come from full physics benchmark analyses (ongoing)... including jet-finding, jet flavour ID, PID, efficiency, etc etc etc Strawman: dijet mass resolution ~ 3-4% As proxies for real physics benchmarks, use PFA-centric tests: Energy sum residuals in e + e qq (q=u,d,s) (equivalent to dijet mass residuals assuming m12 2 = 2E1E2(1-cosθ12)) Dijet mass residuals in e + e Z(νν) Z(qq) @ 500 GeV (q=u,d,s) Caveat: Resolutions are quoted as rms90. Defined as RMS of 90% of events that minimize RMS (i.e. excluding outliers). This is a weird metric, but it s become the convention now. Be aware that rms90 < full RMS (e.g. ratio ~ 78% for a Gaussian). Caveat: All results shown are pure Monte Carlo.
Performance e + e qq (q=u,d,s) @ s GeV for sid02 including muon endcaps (real tracking; no use of truth information at all; LOI code snapshot) Entries per 1 GeV bin 400 350 300 250 200 150 µ : -4.87 GeV 90 rms 90 : 5.87 GeV Entries per 1 GeV bin s=200 GeV 120 100 80 60 40 µ : -5.04 GeV 90 rms 90 : 6.22 GeV 100 50 20 Entries per 1 GeV bin 160 140 120 100 0-40 -30-20 -10 0 10 20 30 Event energy sum residual (GeV) 80 60 0-40 -30-20 -10 0 10 20 30 Event energy sum residual (GeV) cosθ <0.8: ΔECM/ECM = 3.0% 0.8< cosθ <0.95: ΔECM/ECM = 3.2% µ : -13.61 GeV 90 rms 90 : 16.87 GeV Target: 3-4% Entries per 1 GeV bin s=500 GeV 50 40 30 20 µ : -14.86 GeV 90 rms 90 : 15.99 GeV 40 20 10 0 0-120 -100-80 -60-40 -20 0 20 40 60-120 -100-80 -60-40 -20 0 20 40 60 Event energy sum residual (GeV) Event energy sum residual (GeV) cosθ <0.8: ΔECM/ECM = 3.5% 0.8< cosθ <0.95: ΔECM/ECM = 3.3% 10
Angular dependence Aside from tracking, two main issues for PFA: leakage & acceptance Muon barrel Magnet HCAL Leakage muon system worst cos(θ)=0 (min calorimeter depth) solenoid Dead material HCAL SiD muon system Max calorimeter depth at cos(θ) ~ 0.8 Muon endcap ECAL Tracker HCAL Leakage increases for cos(θ) ~ 1... but can use muon system as tailcatcher At small jet angles, particles are lost down the beampipe 11
Angular dependence ) CM energy resolution ("E CM /E 0.07 0.06 0.05 0.04 0.03 e + e qq (q=u,d,s) @ 200/500 GeV for sid02 (LOI code snapshot) s s = 200 GeV = 500 GeV Lots of leakage in barrel for 250 GeV jets. sid02 Using muon system endcap as tailcatcher helps somewhat. (Dependent on segmentation.) E CM 0.02 0.01 Leakage not dominant problem for 100 GeV jets. At very small angles, acceptance dominates. End of solenoid 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 cos(!) 12
Energy dependence So we have to think about leakage too: σ = σem σneutral hadrons σconfusion σleakage 4 Barrel Forward 3.5 3!Ecm/Ecm (%) 2.5 2 1.5 σecm/ecm decreases as Ejet goes up...... then ramps up when leakage becomes an issue. 1 0.5 0 0 50 100 150 200 250 300 Jet energy (GeV) 13
Dijet mass resolution Working assumption: ECM resolution in e + e qq events same as dijet mass resolution of qq jets -- so expect O(3.0-3.5%). e + e Z(νν) Z(qq) @ 500 GeV, q=u,d,s Entries per 1 GeV bin 200 180 160 140 120 µ : -1.24 GeV 90 rms 90 : 4.25 GeV Entries per 1 GeV bin 100 80 60 µ : -2.53 GeV 90 rms 90 : 3.50 GeV 100 80 40 60 40 20 20 0-40 -30-20 -10 0 10 20 30 40 2 Dijet mass residual (GeV/c ) 0-40 -30-20 -10 0 10 20 30 40 2 Dijet mass residual (GeV/c ) cosθ <0.8: dm/m = 4.7% 0.8< cosθ <0.95: dm/m = 3.9% Not so good! What could be the problem? Energy measurement OK but direction measurement poor? Non linearity + asymmetric jet energies? 14
15 Real vs cheat tracking sid02; LOI code snapshot Real tracking Cheat tracking barrel forward barrel forward qq100 3.7% 3.8% 3.4% 3.5% qq200 3.0% 3.2% 2.8% 3.0% qq360 2.7% 2.7% 2.6% 2.6% qq500 3.5% 3.3% 3.5% 3.4% ZZ 4.7% 3.9% 4.2% 3.7% } ΔECM/ECM ΔM/M Tracking is part of the story... Use of cheat tracking reduces the mass-vs-energy resolution difference Makes sense intuitively -- imperfect tracking degrades direction measurement (needed for mass but not energy) Rest from non-linearity? Jury still out.
Summary SiD PFA is alive and well. Performance much improved w.r.t. 12-18 months ago. Energy resolution in target range (3-4%) for Ejet 250 GeV. PFA is being used for physics benchmarks in upcoming Letter Of Intent. SiD PFA now using real tracking code Impressive & fast work by SiD tracking group. Switch from cheat tracking had fairly mild impact on resolution. Calorimeter-Assisted Tracking in the pipeline. Still more work for PFA group to do: Some post-loi fixes already queued up. Dijet mass resolution not scaling as well as we d like. Leakage becoming an issue for Ejet > 200 GeV. arxiv:0901.4670 16
Backups 17
Comparison with Pandora This is a hard thing to do properly! So far no true apples-toapples comparison has been done: Difficulties in getting the fine details of LDC00Sc etc right -- we haven t been able to simulate them in org.lcsim properly. Mokka geometry description complex -- haven t succeeded to build SiD from scratch. Marcel Stanitzki has made several SiDish detectors by deforming LDC00Sc, but none is a true sid02. And there are other issues too, e.g. Tracking (TPC vs silicon, handling of secondaries, etc) HCAL segmentation Pandora is tuned for LDC00Sc, not SiD. Bottom line: impossible to completely decouple comparing detectors from comparing algorithms. That said... 18
Comparison with Pandora Let s start with a very unfair comparison: sid02 for 0.0< cos(θ) <0.8 vs ILD for 0.0< cos(θ) <0.7 B=3.5T, Z=2x2.4m, R=1.85m, 30 layer ECAL with 5x5mm Si pixels, 48 layer HCAL with 3x3cm scintillator cells. sid01 org.lcsim sid02 Pandora v03-β ILD CLIC08, 15 Oct 2008 qq90 ΔECM/ECM = 2.5% qq100 ΔECM/ECM = 3.7% qq200 ΔECM/ECM = 3.0% ΔECM/ECM = 2.1% qq360 ΔECM/ECM = 2.7% ΔECM/ECM = 2.0% qq500 ΔECM/ECM = 3.5% ΔECM/ECM = 2.0% So our PFA on sid02 is outclassed by Pandora on ILD. 19
Comparison with Pandora What about a fairer comparison: sid02 vs a SiDish detector? Caution: This is an older version of Pandora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any variants to choose from! Let s look at the closest to sid02. 20
Comparison with Pandora What about a fairer comparison: sid02 vs a SiDish detector? Caution: This is an older version of Pandora.!"#$%&'(&)*+ NB: Scintillator HCAL These lines straddle the right number of layers (40) and iron thickness (20mm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any variants to choose from! Let s look at the closest to sid02. 21
Comparison with Pandora org.lcsim sid02 Real tracking org.lcsim sid02 Cheat tracking Pandora SiDish pair A (mean) qq90 ΔECM/ECM = 3.1% qq100 ΔECM/ECM = 3.7% ΔECM/ECM = 3.4% qq200 ΔECM/ECM = 3.0% ΔECM/ECM = 2.8% ΔECM/ECM = 2.8% So numbers are not so far apart for similar detectors. (... but what about qq360/qq500? No SiDish data yet -- CPU time limitations.) 22