Automated reconstruction of LAr events at Warwick J.J. Back, G.J. Barker, S.B. Boyd, A.J. Bennieston, B. Morgan, YR IOP meeting Manchester 28/11/2012 Y. Ramachers Challenges Single electron, 2 GeV in LAr: ● Easy 'by-eye' in isolation ● Challenging for computer IOP meeting Manchester 28/11/2012 Y. Ramachers Challenges Spot the electron! ● ● ● ● 2 GeV e- as before 81 cosmic muons in 14x14x20 m3 volume in 1.4 ms readout interval. Analysis 'by-eye' becomes interesting... A big challenge for any automated analysis. IOP meeting Manchester 28/11/2012 Y. Ramachers Automated reconstruction at Warwick Raw data: (x,y,z,charge) Detector/Simulation incl. charge quench in Genie/Geant4 Simulation: (x,y,z,charge,truth_Id) Pre-processing DBScan event noise filter, dEdx filter, charge smoothing, masking structures, feature finding, utilities Being validated Shower/Track segmentation decision Exists, tested Track segmentation Shower Segmentation - Cellular automaton - LPC algorithm LPC shower finder, total energy, direction, extent Physics Structure container Clusters of hits, each a reconstructed particle Vertex finder Particle Identification Six discrimination variables and multi-variate analysis TMVA Truth value validation, track length calculation, observables, ... LPC – vertex finder Post-recon analysis Kinematics calculator IOP meeting Manchester 28/11/2012 Kalman filter fit Y. Ramachers Current status ● Shower / Track decision code being validated ● Track analysis automated, validated for CA, LPC not yet ● Shower analysis exists, no validation yet ● Particle ID, finished ● Papers: – “Interest point detection for reconstruction in high granularity tracking detectors”, B. Morgan, JINST 5 (2010) P07006 – “Electron-Hadron shower discrimination in a LArTPC”, submitted (arXiv:1210.2215) – CA paper in preparation – Local Principal Curves (LPC) paper in preparation IOP meeting Manchester 28/11/2012 Y. Ramachers Current status Code structure: Module catalogue: ● Hit filter: ● Number of hits ● dE/dx filter ● DBScan Noise ● DBScan clustering ● Boolean cuts on truth values (for simulated data) ● Charge smoothing (for sim. Data) ● Cluster merging and stitching ● Feature point finder ● Hit/Region masking ● PCA transformation ● Cluster range calculation IOP meeting Manchester 28/11/2012 Y. Ramachers Automated reconstruction at Warwick Raw data: (x,y,z,charge) Detector/Simulation incl. charge quench in Genie/Geant4 Simulation: (x,y,z,charge,truth_Id) Pre-processing DBScan event noise filter, dEdx filter, charge smoothing, masking structures, feature finding, utilities Being validated Shower/Track segmentation decision Exists, tested Track segmentation Shower Segmentation - Cellular automaton - LPC algorithm LPC shower finder, total energy, direction, extent Physics Structure container Clusters of hits, each a reconstructed particle Vertex finder Particle Identification Six discrimination variables and multi-variate analysis TMVA Truth value validation, track length calculation, observables, ... LPC – vertex finder Post-recon analysis Kinematics calculator IOP meeting Manchester 28/11/2012 Kalman filter fit Y. Ramachers Cellular Automaton (CA) for tracks µ µ µ IOP meeting Manchester 28/11/2012 Validation on low energy CCQE muon + proton events and CCQE muon + proton + pion events. Y. Ramachers Cellular Automaton (CA) for tracks Muon efficiency Muon purity Combined muon, proton factors Have two dominant parameters: ● CA-angle ● Stitching angle Overall particle efficiency / purity is at: Muon Proton Eff. 95-97% 95-98% Pur. 91-92% 95-96% IOP meeting Manchester 28/11/2012 Y. Ramachers Automated reconstruction at Warwick Raw data: (x,y,z,charge) Detector/Simulation incl. charge quench in Genie/Geant4 Simulation: (x,y,z,charge,truth_Id) Pre-processing DBScan event noise filter, dEdx filter, charge smoothing, masking structures, feature finding, utilities Being validated Shower/Track segmentation decision Exists, tested Track segmentation Shower Segmentation - Cellular automaton - LPC algorithm LPC shower finder, total energy, direction, extent Physics Structure container Clusters of hits, each a reconstructed particle Vertex finder Particle Identification Six discrimination variables and multi-variate analysis TMVA Truth value validation, track length calculation, observables, ... LPC – vertex finder Post-recon analysis Kinematics calculator IOP meeting Manchester 28/11/2012 Kalman filter fit Y. Ramachers Most versatile: LPC LPC on low energy electron showers. IOP meeting Manchester 28/11/2012 Y. Ramachers LPC capabilities On tracks: N-dimensional feature finder ● Track separation at feature points ● Vertex finder (Blue point, top left pict.) On showers: ● Shower/Track discrimination decision ● Shower finder / analysis IOP meeting Manchester 28/11/2012 Y. Ramachers LPC on vertex finding ● ● ● ● IOP meeting Manchester 28/11/2012 Segment event structure at feature point Fit line to segment ends Calculate line intersection or closest approach for vertex finding Better than 1 cm resolution on low energy muon + proton CCQE events here. Y. Ramachers Automated reconstruction at Warwick Raw data: (x,y,z,charge) Detector/Simulation incl. charge quench in Genie/Geant4 Simulation: (x,y,z,charge,truth_Id) Pre-processing DBScan event noise filter, dEdx filter, charge smoothing, masking structures, feature finding, utilities Being validated Shower/Track segmentation decision Exists, tested Track segmentation Shower Segmentation - Cellular automaton - LPC algorithm LPC shower finder, total energy, direction, extent Physics Structure container Clusters of hits, each a reconstructed particle Vertex finder Particle Identification Six discrimination variables and multi-variate analysis TMVA Truth value validation, track length calculation, observables, ... LPC – vertex finder Post-recon analysis Kinematics calculator IOP meeting Manchester 28/11/2012 Kalman filter fit Y. Ramachers Particle ID ● Assumption: Isolated hit structures obtained from a previous step, i.e. assume to analyse single clusters! ● ● ● Step 1: Principal component transformation of hit cloud. Step 2: Calculate 6 discrimination variables: (1) Lateral structure – Core-to-Total ratio ('lat') (2) Hit concentration – Coulomb energy of hits as a measure of concentration ('con') (3) Initial dE/dx – slice cluster longitudinally, take starting part only, ('dedx') (4) Spatial extent - Calculate convex hull, get spatial extent of structure in all three principal axes, ('extx, exty, extz') Step 3: Multivariate Analysis – run TMVA4 and pick best method (typically boosted decision trees) IOP meeting Manchester 28/11/2012 Y. Ramachers Particle ID Example low energy π0 shower and PCA transform Core 9.61cm x 9.61cm for 'lat' variable Main p rincipa l axis Lateral projection plane Axes units in [mm] IOP meeting Manchester 28/11/2012 Y. Ramachers EM/(all pions+protons+muons) Signal is an isolated EM shower; background all muons, protons, pions, other mesons Signal/Background distribution from CNGS neutrino beam IOP meeting Manchester 28/11/2012 Signal/Background distribution uniform between 10 MeV and 4.5GeV. Y. Ramachers Conclusion ● Fully automated reconstruction of LAr events is possible with this software, how well is not known yet. ● Step-wise validation is ongoing. ● Most promising algorithm so far: local principal curves (LPC) ● ● ● All methods at Warwick are N-dimensional – can run on 2D images, not only full 3D. Two more publications in the pipeline: CA and LPC, particle ID is submitted and on arXiv. Final target would be to build a physics reconstruction study. IOP meeting Manchester 28/11/2012 Y. Ramachers