Particle Identification in MINERvA using Artificial Neural Networks

1
Particle Identification in MINERvA using
Artificial Neural Networks
P. Stamoulis G.Tzanakos University of Athens,
Athens, Greece
MINERvA Collaboration Meeting, 14 May 2004
2
Stating the obvious

• Cross-section measurements in Minerva
• Interaction classification
• Particle Identification
• In principle, it should work with simple cuts in
  some very clearly
• discriminating variables (e.g. length,dE/dx for
  p/?)
• But there are some special discrimination tasks
• Low energy particles, penetrating ?/?, ?/?-
• There, a multivariate approach like ANNs can do
  better

3
ANN basics

• Define characteristic,
• separating variables for each
• class Form patterns (MC
• events).
• Train the ANN with these
• patterns.
• Use trained ANN to separate
• real events.
• The result is the Bayes a
• posteriori probability that the
• event/particle belongs to the class

step functions
Many available packages (simple or not,
supported or not) MLPfit, SNNS, ROOT embedded
TNeurons
We have widely used ANNs for interaction
classification -Niki Saoulidou in DONUT -Niki
and Athens U. group in MINOS Far Detector -PID in
MINOS CalDet
4
Performance example The CalDet e/(h,?)
separation at 0.6 GeV/c
The Input
Sample of (25 in total) separating variables
usedTotally overlapping, but with distinct shapes
The Result
5
Examples of p/?/? separation Stating the problem
Simple MC event display (á la Howard, using
ROOT) 2 resonance scattering examples shown (from
Steves ntuples)
Seeing many of these, we decide on possible
separating variables
6
Discriminating variables
CAUTIONResults before digitization and any
tracking/clustering!!
7
The hard part Separating pis of different charge
Important to separate different resonance
production interactions

• dE/dx, total track length, opening angle,
  obviously wont
• work
• Chose to look at stopping patterns of produced
  particles
• showers - electrons
• b. Shower energy Electron energy
• c. Shower/Electron opening angle (?)
• d. Track length before interaction or decay
  (Inconclusive before tracking proceeds)

Modified MC Hugh
?-
?
1 GeV/c
8
Different stopping behaviour of pi- and pi
0.1 GeV/c
0.4 GeV/c
0.8 GeV/c
1.0 GeV/c
Electron and shower multiplicity
?
?-
Electron and shower total energy
9
What needs to be done -Actually create an ANN,
train on MC and see discrimination
quality. -TRACKING !! Working with the whole MC
truth info does not tell the whole truth of the
the power and limitations of the technique. -MC
tuning (?) It is possible that this effect will
go away or become less significant when using a
different hadronic interaction MC. -Expansion Deci
de whether to use the technique on PID or
interaction separation.