1. Neural tracking in ALICE Alberto Pulvirenti – University and I.N.F.N. of Catania ACAT ’02 conference Moscow, June 26 2002

2. Outline The ALICE experiment The ALICE experiment Tracking in ALICE Tracking in ALICE Why an ITS stand-alone tracking? Why an ITS stand-alone tracking? Implementation Implementation Results Results Work in progress and outlook Work in progress and outlook

3. The Large Hadron Collider http://www.cern.ch ~9 km LHC SPS CERN

4. The 4 LHC experiments

5. ALICE’s objective: QGP study Pb+Pb @ LHC (5.5 A TeV) The Big Bang The Little Bang

6. ALICE track multiplicity A sketch…

7. ALICE track multiplicity Simulation and reconstruction of a “full” (central) Pb+Pb collision at LHC (about 84000 primary tracks!) takes about 24 hours of a top-PC and produces an output bigger than 2 GB. A sketch… of 1/100 of a typical ALICE event

8. The ALICE detector

9. Tracking in ALICE Time Projection Chamber. Time Projection Chamber.  ~180 points per track  main contribution. Inner Tracking System. Inner Tracking System.  6 points close to primary vertex  improves resolution near to the production vertex. Standard procedure: Standard procedure: 1.Points in the TPC outermost pad-rows are arranged into suitable track seeds. 2.the seeds are propagated through the TPC towards its innermost pad-row, according to a Kalman filter algorithm for both recognition and reconstruction. 3.each track found in the TPC is propagated in the ITS and its parameters are refined with the aid of the six best matched ITS points.

10. Why an ITS stand-alone tracking? …because the TPC is a “slow” detector  some events could be produced in a “high-rate acquisition mode”, by turning on only the fastest ALICE modules (ITS, Muon Spectrometer), to produce large amounts of data useful for all analyses needing high statistics.  in this case, we need at least a satisfactory efficiency for high transverse momentum (p t >1 GeV/c). …because some particles decay within the TPC barrel volume, and the standard TPC tracking doesn’t manage to create seeds for them.  in this case, the tracking is performed after completing the standard Kalman procedure, and working only on the points which the Kalman method didn’t use.

11. Implementation: 1 – definitions

12. Neuron: oriented track segment  2 indexes: [s ij ] links two consecutive points in the particle’s path according to a well-defined direction Implementation: 1 – definitions

13. Weight: geometrical relations between neurons  4 idxs: [w ijkl ] Geometrical constraint: only neurons which share a point have a non zero weight

14. Implementation: 1 – definitions Weight: geometrical relations between neurons  4 idxs: [w ijkl ] Geometrical constraint: only neurons which share a point have a non zero weight Case 1: sequence guess for a track segment good alignment requested

15. Implementation: 1 – definitions Weight: geometrical relations between neurons  4 idxs: [w ijkl ] Geometrical constraint: only neurons which share a point have a non zero weight Case 1: sequence guess for a track segment, good alignment requested Case 2: crossing negative weight leads to a competition between units

16. Weight: geometrical relations between neurons  4 idxs: [w ijkl ] Geometrical constraint: only neurons which share a point have a non zero weight Case 1: sequence guess for a track segment, good alignment requested Case 2: crossing negative weight leads to a competition between units Implementation: 1 – definitions

17. Neural Network Simulation Specifics Associative memory topology (single layer of fully connected units). Associative memory topology (single layer of fully connected units). Real valued (“sigmoidal”) activation function, limited between 0 and 1. Real valued (“sigmoidal”) activation function, limited between 0 and 1. Random initialization. Random initialization. Asynchronous updating cycle (one unit at a time). Asynchronous updating cycle (one unit at a time). Stabilization threshold on the average activation variation after a complete updating cycle. Stabilization threshold on the average activation variation after a complete updating cycle. Resolution of competitions to the advantage of the unit with the greatest real activation. Resolution of competitions to the advantage of the unit with the greatest real activation. Binary mapping of “on” and “off” units with a threshold of 0.6 on the final real neural activation. Binary mapping of “on” and “off” units with a threshold of 0.6 on the final real neural activation.

18. Implementation: 2 – cuts Needed to limit the number of point pairs used to create neurons 1. 1. Check only couples on adjacent layers 2. 2. Cut on the difference in polar angle (  ) 3. 3. Cut on the curvature of the projected circle passing through the two points and the calculated vertex 4. 4. “Helix matching cut” …where a is the corresponding circle arc of the projection in the xy plane

19. Implementation: 3 – procedure “Step by step” procedure (removing the points used at the end of each step) Many curvature cut steps, with increasing cut value Sectioning of the ITS barrel into N azymuthal sectors RISK: edge effects the tracks crossing a sector boundary will not be recognizable by the ANN tracker

20. Implementation: 4 – reconstruction Track reconstruction: Kalman Filter. (ref.: A. Badalà et al., NIM A(2002) in press and references therein). Track reconstruction: Kalman Filter. (ref.: A. Badalà et al., NIM A(2002) in press and references therein).  “vertex constrained” seed.   A helix is estimated by using the two outermost points and the experimental vertex (the same which is used for neuron creation cut).  two operational phases: 1.vertex  layer 6. 2.layer 6  vertex.

21. Test trial ingredients Test on a simulation produced with the HIJING event generator interface (developed within the AliRoot framework), and tracks transported through the detector by GEANT 3.21: Test on a simulation produced with the HIJING event generator interface (developed within the AliRoot framework), and tracks transported through the detector by GEANT 3.21:  All detectors and all physical effects turned “on”.  Fully detailed geometry, simulation and reconstruction in the ITS.  ALICE “default” number of primary tracks (84210 in the pseudorapidity region |  | < 8.0). Track definition for efficiency evaluation Criterion GOOD TRACK (fake otherwise) FINDABLE TRACK “SOFT” at least 5 right points Has at least 5 points in ITS “HARD” all 6 point must be correct Has a point for each layer Efficiency = # good tracks (fake tracks) / # findable tracks

22. “Signal-to-noise ratio” Layer123456Avera ge Good / All46%60%65 % 69 % 77 % 74 % 65% Unused good / All unused 21%37%45 % 51 % 68 % 63 % 47%

23. Stand-alone tracking: results (I) Number of found tracks, efficiency and CPU time as a function of the # of sectors. Only one event analyzed. Test choice: 18 sectors CPU time: ~10% of the time requested the whole ITS at once PC used: PIII 1 GHz

24. SOFT good fake Stand-alone tracking: results (II)

25. Stand-alone tracking: results (III) Dip angle ( ) resolution (in mrad) sigma = 3.69  0.01 Azimuthal angle (  ) resolution (in mrad) sigma = 4.71  0.01 p t resolution (in % of true value) sigma = 13.4  0.3 % (only 6 points!)

26. Stand-alone tracking: results (III) Transverse impact parameter resolution (in microns) sigma = 79.7  0.1 Longitudinal impact parameter resolution (in microns) sigma = 265.6  0.4

27. Stand-alone tracking results (III) Parameters resolution NeuralKalman (without vertex. constr.) p t (%) 13.4  0.31.57  0.02  (mrad)4.71  0.011.40  0.08  (mrad)3.69  0.011.60  0.08 D t (  m)79.7  0.1 ~ 50 D z  (  m)265.6  0.4 ~150 Efficiency for tracks with p t  1 GeV / c Efficiency (%)Fake (%) Neural “soft” 78.2  3.09.9  0.9 Kalman “soft” 72.8  2.94.9  0.6

28. “Combined” tracking: results (III) Results for Pt  1 GeV / c:Kalman Efficiency: 72.8  2.9 Fake prob.: 4.9  0.6 Efficiency per particle [p t  1 GeV/c]  K Kalman 74.7  2.764.5  0.8

29. “Combined” tracking: results (III) The “findable” tracks are counted among all ITS findable tracks (even the ones which are NOT findable in the TPC) Results for Pt  1 GeV / c:KalmanCombined Efficiency: 72.8  2.983.0  3.0 Fake prob.: 4.9  0.67.0  0.7 Efficiency per particle [p t  1 GeV/c]  K Kalman 74.7  2.764.5  0.8 Combined 84.7  2.976.2  0.8 10% increase!

30. Conclusions & work in progress  The Neural Network tracking algorithm has been successfully adapted to the unprecedented ALICE multiplicity  Implementation has been done in the official AliRoot off-line framework based on ROOT.  Recognition efficiency is comparable with the Kalman Filter one, in the range of p t > 1 GeV/c. Under study: Under study:  Improving the neural algorithm performances for LOW transverse momentum tracks [ p t < 0.2 GeV/c ] (not a trivial task!).  Alternative possible techniques for the same purpose (adapting some existing algorithms like elastic tracking, elastic arms algorithm, or developing a genetic algorithm). Future developments (for “combined” tracking). Future developments (for “combined” tracking).  Improving track parameter resolution by including also the TPC/TRD points “unused” by Kalman tracking.