Calorimetry Simulations

1
Calorimetry Simulations

• Norman A. Graf
• for the SLAC Group
• January 10, 2003

2
Analysis Infrastructure

• Redefine detector segmentation Detector
  definitions changed and CalorimeterHits rewritten
  for each event. Allows using existing data sets
  for comparing different detector segmentations.
• Generate events with fine segmentation, gang at
  analysis level to study effects of cell size.
• Redefine particles (MCParticle) to which
  calorimeter energy is assigned. Especially useful
  for particles that interact in inner walls of
  calorimeters.
• Finer control over which MC particle is
  considered shower initiator.

3
Clustering Algorithms

• Current clusterer (SimpleClusterFinder) All
  adjacent hits in a calorimeter form a cluster
• Extend idea of adjacency across EM-HAD border.
• Adjacency extended to an integer number of bins
  in theta, phi, and layer.
• Need to extend across Barrel-Endcap borders.
• Fixed-Cone clusterer developed for analysis of EM
  showers.
• Fast, efficient.

4
Cluster Algorithm Evaluation

• Clustering analyzer in progress will ultimately
  produce efficiency plots, purity plots, and
  energy resolution for each class of particle.
  (EM, Charged hadron, Neutral hadron)

5
Energy Assignments

• Work continuing to understand sampling fraction
  differences.
• EM vs HAD
• Barrel vs Endcap (esp. in 5T SD)
• Gismo/Geant4 differences in energy deposition
  under investigation Total fraction of photon
  energy to ionization, fraction of ionization in
  active material, and effect of magnetic field all
  different in Geant4.

6
ClusterID Algorithm

• Goal is to identify particle type (photon,
  charged hadron, neutral hadron, , fragment) that
  created each cal cluster.
• Based on a set of discriminators measured for
  each cluster ( shape and pointing parameters, )
• Now using a Neural Net for discrimination.

7
ClusterID Neural Net

• Input variables for the neural net are composed
  of cluster shape quantities, e.g.
• Normalized cluster energy tensor eigenvalues
• Cluster extent, number of hit cells in cluster,
  cluster energy,
• Position and angular difference wrt IP
• Total of 15 inputs
• 4 outputs Photon, Charged Hadron, Neutral
  Hadron, Fragment (assign cluster highest ID).

8
Neural Net Training

• Neural Nets have been trained on single particle
  samples.
• Tested on both single particle samples and
  physics samples.
• Work ongoing to refine input variables.
• Framework and trained nets exist and have been
  released in latest hep.lcd distribution.
• Aim to release fully retrainable cluster ID and
  general NN application code when JAS3LCD is
  ready.

9
ClusterID Current Status

• Study Z decays at Z pole in SD.
• Gives Z mass width twice the width of perfect
  reconstruction.
• Correctly IDs 90 of gamma energy.
• Incorrectly IDs 6 of gamma energy.
• Correctly IDs 66 of neutral had energy.
• Incorrectly IDs 27 of neutral had energy. (27
  goes to 42 misID with cal gap)

10
Zmass at Zpole
11
EM Id

• Developed simple fixed-cone algorithm for finding
  EM clusters.
• Fast, efficient.
• Implemented fully covariant ?2 calculation for
  longitudinal shower shape analysis.
• Use shower width for ?-?0 discrimination.
• In addition, use track-match and E/p for electron
  id.

12
Cone Algorithm

• Currently using fixed cone radius of 0.03 in ?,?
  space on EM Calorimeter hit cells.
• Based on energy contained within cone.
• Based on number of clusters.
• Could also use a cone radius based on energy of
  seed cell.
• Currently split clusters whose cones overlap by
  associating cells to nearest cone axis.
• Could also search for NN clusters within cone.

13
Longitudinal HMatrix

• Use longitudinal energy depositions and their
  correlations to create a cluster ?2.
• Mild, smooth energy dependence (logE)

14
Charged Hadron Id (No clustering)

• Continuing to characterize pion shower shapes in
  calorimeters as function of momentum and
  direction.
• PionShower class being developed to encapsulate
  the association of hit calorimeter cells with
  extrapolated tracks.
• Follows MIP trace to shower start.
• Characterize hit-track association with ?2.
• Will allow association to proceed until a limit
  is reached on either match ?2 or E/p.

15
ReconstructedParticle

• A class which encapsulates the behavior of an
  object which can be used for physics analysis.
• mirrors MCParticle
• Kinematics determined by track momentum or
  calorimeter cluster energy at time of creation.
• ID determined later by particle ID algorithms,
  e.g. track dE/dx, cluster shape, or combination
  of detector element variables.

16
Detector Designs

• New SD geometry without physical gap between EM
  and HAD.
• Propose strongback followed by sampling layer.
• Working with T. Behnke, have first implementation
  of T detector.
• Approximation to Tesla detector using simplified
  geometries (barrels and disks) and projective
  readout.
• Should simplify EFlow analysis comparisons.

17
Fast Simulations

• Current fast simulation does not populate
  calorimeter cells, only smears Clusters.
• Working on more realistic fast simulation
• Using parameterizations for longitudinal and
  lateral shower shapes.
• Fast prototyping of materials, segmentation, etc.
• Using a shower library.
• Fast simulation of large samples for fixed
  detector.

18
New Functionality

• Ganging of calorimeter cells during analysis.
• Simple (NN) clustering across EM-HAD.
• User-defined neighborhood size for clustering.
• Cone algorithm HMatrix for EM showers.
• Neural Net applied to ClusterID.
• ReconstructedParticle definitions arising.
• Integrated Eflow package being developed.
• T Detector implemented for comparison.

19
Acknowledgements

• Have just presented an overview of work being
  done.
• Details can be found in talks presented (or to be
  presented) in ALCPG calorimeter meetings.
• Thanks to
• T. Behnke, G. Bower, R. Cassell, T. Johnson,
• W. Langeveld