Title: PID Capabilities using Artificial Neural Networks in the TRD
1PID Capabilities using Artificial Neural Networks
in the TRD
Alexander Wilk IKP Münster August 22nd 2008 ALICE
Workshop Sibiu
2gt Outline
- PID methods for the TRD and their principles
- performance results
- test beams
- simulations
- electrons from g-conversions as reference data
- contamination studies
- summary
3gt Definitions
- design goal for the ALICE TRD PID
- pion efficiency lt 1 at 90 electron
efficiency - electron efficiency is the rate of electrons
which are identified correctly - pion efficiency is the rate of pions which are
- misidentified as electrons
- pion suppression is the reciprocal of pion
efficiency - non-electron efficiency is the same as pion
efficiency, but the pions here are a mixture
of hadrons
4gt non-Electron Efficiency - Particle
Ratios in pp Collisions
- ratios of primary particles are taken into
account for the calculation of the efficiencies - negative sample
- is a mixture of all other particle types,
depending on their ratios in pp collisions
e
m
10
10
0
0
K
p
10
10
0
0
1
p
ratio
0
0
10
0
10
pT (GeV/c)
pT (GeV/c)
5I. PID Methods
6gt Principles of PID in the TRD
- TRD signal is composed of 2 components
- energy loss of charged particles
- (all particles)
- absorption of TR photons
- (g gt 1000, electrons only)
- transition radiation photon absorbed early
- in chamber ? late in drift time
- signal in one chamber correlated
- in time (TRF), but independent between
- different chambers
- 6 layers of TRD chambers
- 24 time measurements per chamber
- ? 6 x 24 144 measurements per track
TR Peak
7gt 2-dimensional Likelihood (A. Bercuci, GSI)
- information of 2 time slices per chamber are
connected in one histogram - idea
- signal of electrons should be significantly
- larger in the second time slice due to
- transition radiation
- advantages
- consideration of time information
- consideration of correlations between
- different time slices
- disadvantages
- not extendable to more than 3 time bins
- (statistical problem)
1
2
http//indico.cern.ch/getFile.py/access?resId0ma
terialId slidescontribIdI9sessionId2subContI
d6confId7104
8gt Artificial Neural Networks
- information of more time slices is usable (here 8
slices) - time information is not lost
- network topology (for 1 chamber)
- 8 (or more) input neurons
- 2 hidden layer
- 2 or 5 output neurons
- calculation of particle probabilities for 6
chambers analogue to likelihood method
Input
http//indico.cern.ch/getFile.py/access?contribId
10resId1materialIdslidesconfId20791
9gt Artificial Neural Networks
- information of more time slices is usable (here 8
slices) - time information is not lost
- network topology (for 1 chamber)
- 8 (or more) input neurons
- 2 hidden layer
- 2 or 5 output neurons
- calculation of particle probabilities for 6
chambers analogue to likelihood method
e m p K p
Input
http//indico.cern.ch/getFile.py/access?contribId
10resId1materialIdslidesconfId20791
10gt Artificial Neural Networks
- information of more time slices is usable (here 8
slices) - time information is not lost
- network topology (for 1 chamber)
- 8 (or more) input neurons
- 2 hidden layer
- 2 or 5 output neurons
- calculation of particle probabilities for 6
chambers analogue to likelihood method
e m p K p
Input
,
http//indico.cern.ch/getFile.py/access?contribId
10resId1materialIdslidesconfId20791
11gt Test Beam Times
- 3 larger test beam times at CERN PS
- 2002 (prototype tests)
- 2004 (stack of 6 large chambers)
- 2007 (super module III)
- secondary beam composed of electrons and pions
- momenta between 1 GeV/c and 10 GeV/c
- independent PID using a Cherenkov detector and a
PbGl calorimeter
12gt Test Beam Time 2002 - Comparison 1-dim LQ/NNs
- first application of artificial neural networks
in the TRD - 2002 extrapolation from 4 to 6 chambers
- design goal reached using 1-dim likelihood and
NNs - NNs give (factor 3) better results than LQ
- NNs exploit time
- information
design goal for 3 GeV/c
13gt Test Beam Times Overview - Results NNs
- later beam time results demonstrate NNs advantage
compared to 1-dim likelihood - unfortunately results from 2002 not completely
reached
design goal for 3 GeV/c
14gt Test Beam Times Overview - Number of Input
Neurons
- rising number of input neurons
- better pion efficiency
Slices
number of input neurons
15gt PID in Simulations
- goal implementation of PID using NNs in AliRoot
and building of reference data - for training of the networks no complete
simulation of collisions necessary - simulation of
- 5 particle types (e, m, p, K, p)
- 11 momenta (0.6, 0.8, 1.0, 1.5, 2.0, 3.0, 4.0,
5.0, 6.0, 8.0 und 10.0 GeV/c) - 20 particles and anti-particles per particle type
in each event (200 particles per event)
16gt Results from Simulations
- 2-dim likelihood and
- NNs significantly better than 1-dim LQ
- NNs about a factor of 2 better than 2-dim
likelihood
17gt Results for Hadron PID
- TRD can help with hadron identification at small
momenta
18gt Summary of the PID methods
- particle identification in the TRD is based on
the deposited charge of the crossing particle and
the absorption of created transition radiation - two powerful methods were developed and
implemented in AliRoot - 2-dim likelihood method
- artificial neural networks
- suppression factor for hadrons is about 1000
- for 2 GeV/c in simulations
19II. Reference Data
20gt Types of Reference Data
- simulations
- at the moment used in AliRoot
- special PID simualtions using 5 particle types
with a given momentum - test beam data 2007
- no magnetic field (no Lorentz angle)
- only one beam position in the SM and only one
incident angle - problems with the PID not usable as reference
data for NNs (?) - reference data from real data needed!
21gt Particles from Displaced Vertices (M. Heide,
IKP)
- use particles from displaced vertices in pp
collisions as reference - (e.g. g-gtee-)
- independent PID information
- reconstruction of invariant mass
- location of the decay
- additional PID of the other barrel detectors
(ITS, TPC and TOF) - study done for TPC dE/dx calibration (A.
Kalweit, GSI)
http//indico.cern.ch/getFile.py/access?contribId
4resId 1materialIdslidesconfId8438
22gt Example g Conversions
- probability for g conversions in the inner barrel
10 - reconstruction in AliRoot via the V0 finder
23gt Spectrum of Invariant Mass
- spectrum after cuts on
- pointing angle lt 0.03
- TRD hits 6
- clear peak at minv 0
- small background
- electrons also parts of the background
all reconstructed particles reconstructed
g background
24gt pT Spectrum of the Electrons
- pT spectrum after cuts
- pointing angle lt 0.03
- TRD hits 6
- TPC PID gt 0.5
- small background 20 non electrons
- statistics for 400k events
e from reconstructed g e in background non e
background
25gt Summary Reference Data
- reference data from real data taking needed for
- 2-dim LQ reference histograms
- training of artificial neural networks
- very pure samples of electrons can be extracted
for momenta lt 3 GeV/c - to do
- study with large statistics simulations of g
conversions with high pT (TPC PID) - extract ps and ps from K0s and Ls
- think about kaons
26III. Contamination Studies
27gt Contamination Studies - Motivation
- pT spectrum of particles reconstructed with
V0-finder (no cuts) - background can be reduced by a very large factor
(depending on cuts) - how much background is acceptable for training
the NNs?
background e from reconstructed g electrons in
background
(Markus Heide, IKP)
28gt Contamination of the Samples
- contamination is the rate of not correctly
identified (or assigned) particles - particles in the contamination are a mixture of
particles given by the rates of different
background particles (extracted by Markus Heide) - contamination is done using gRandom
- if contamination is done (depending on cont.
level) - which particle type is used (depending on
particle ratios) - only training samples contaminated, test samples
consist of pure particle samples
29gt Example Contamination 30
each sample consists of 70 correct assigned
particles 30 background
background (2GeV/c) e 5.8 m 0.5
p 65.2 K 7.7 p 20.8
30gt Example Contamination 30
each sample consists of 70 correct assigned
particles 30 background
background (2GeV/c) e 5.8 m 0.5
p 65.2 K 7.7 p 20.8
31gt Results (2GeV/c)
- contamination until 15 shows no big effect on
PID - up to 70 PID is not that bad
- step from 15 to 20 not understood yet
32gt Results (10GeV/c)
- contamination until 15 shows no big effect on
PID - up to 70 PID is not that bad
- step from 15 to 20 not understood yet
- for 10 GeV/c comparable results
33gt Test Beam Data 2007
- some problems in Test Beam Data
- clearly identified pions show very high electron
probability - -gt overlapping particle
- tracks
- using some cuts, the cleaning of samples is
possible, but - bad statistics or
- bias on electron and pion samples
- -gt no reasonable PID calculation possible
e-
p-
34gt Test Beam Data NNs
- neural networks were trained with test beam data
- cuts were used to reduce rate of overlapping
tracks - anyway bad results for pion efficiency due to
contamination with overlapping tracks - what happens if networks from test beam were
applied to really clean data? - -gt apply networks on data from simulations
35gt The Testing Sample
- data from simulations
- data that was used in AliRoot to train the
reference neural networks - about 20k events per particle type and momentum
- calibration
- by hand using dE/dx distributions and output of
neural networks - same calibration factor of 0.5 on dE/dx worked
for all momenta (only for 1 GeV/c a higher factor
0.8 was needed -gt attachment due to gas leak)
36gt Results of Test Beam Networks using
Simulated Data
- results fit well to the results from former test
beam times - for 1 and 2 GeV/c pion efficiency lt 0.2
37gt Summary
- we have two powerful PID methods for the TRD
- 2-dim likelihood
- PID with artificial neural networks
- both PID methods fulfill design goal and were
tested in simulations (both) and with test
beam data (NNs only) - study of getting reference data from displaced
vertices shows that very pure samples of
electrons can be extracted
(for momenta lt 3 GeV/c) - contamination study shows that a contamination up
to 15 in the reference data has nearly no effect
on PID performance - networks trained with 2007 test beam data show
nice performance on simulated data
38Thank you!
39Backup
40gt dE/dx in the TRD
41gt Pion Efficiency in Simulations
42gt Outlook
- study with large statistics simulations of g
conversions with high pT - test K0 and L decays for extraction of pions and
protons, think about possibilities of kaon
identification - framework for extraction of reference data is in
progress