1
Progress report on
Calorimeter design comparison simulations
MICE detector phone conference2006-01-27
Rikard Sandström
2
Before I begin: Scraping in trackers
• At 6 pi mm, partial scraping in trackers.– Particles still make it
through the experiment.
• Manually filtering events with more than 7 MeV energy loss in a tracker.– Done by using MC truth
values.
• Tracker people will have to deal with this.
3
Outline
• The alternative detector geometry• Techniques and methods used
– PID simulations in 14 steps
• Present status• Results so far
4
The alternative calorimeter
• The alternative calorimeter consists of one KLOE light layer in front, then ten plastic layers.
– Used to call this smörgås, sandwich has two KLOE layers. The latter is no good idea, so sandwich now means the variant with only one KL layer.
• Plastic layers contain 9 cells each, at increasing thickness. (1 cm to 12 cm).– Increasing thickness gives best range(p) resolution for
money.
• Total number of channels is constant between designs.– KL: 4x30x2 = 240 channels.– SW: (30+10x9)x2 = 240 channels.
• Abbreviations:– KL = KLOE Light (4 KLOE Light layers)– SW = Sandwich (1 KLOE Light layer, then plastic)
5
Reminder of run plan
• Stage 1– Pi & Mu
• 100<pz<300 MeV/c
• Stage 6– Mu & mu-decay
• 140 MeV/c• 170 MeV/c• 200 MeV/c• 240 MeV/c• Tilley’s TURTLE beam, with diffuser
6
Method #1 (examples follows)
1. Write a document explaining what to do and why• Not in the document = not on the table.
2. Simulate beams of 10k events, wide distributions.3. Use those to find useful variables for PID.4. Find combinations of detectors, such that given A,
expect B.5. Make fits for all expected values, and create
“discrepancy variables” 1-expected/measured.• Zero means very muon like.
6. Run 120k events of muons per experimental scenario.• ~ 2Gb of data per file
7. For every such scenario, also run 120k muons with 40 ns lifetime to generate background.• Muons not decayed at TOF2 are filtered out of analysis.
7
Method #2 (examples follows)
8. Digitize every simulated beam.9. Convert to ROOT trees, and tag good/bad event.10. For every scenario, merge the muon sample with
the background sample. 11. Filter out events while trying to not lose any muons.12. Train a Neural Net on the half of the merged &
filtered sample (training sample).13. Using the weights acquired by Neural Net, assign a
weight all other events (the test sample).14. Evaluate the PID capabilities by looking at weights
for the test sample.
8
100<pz<300 MeV/c
9
100<pz<300 MeV/c
Sandwich
10
Example of a fit
11
Example of “discrepancy variable” used for Neural Net
Discrepancy = 1-expected/measured
12Discrepancy = 1-expected/measured
13
Stage 1, 100<pz<300 MeV/c
Stage 1 Mu, KL Pi, KL Mu, SW Pi, SW
Simulation
100% 100% 100% 100%
Digitisation
100% 100% 100% 100%
Fits 100% 100% 100% 100%
RootEvent
100% 100% 100% 100%
Neural Net 100% 100%
14
Stage 6, 140±14 MeV/c
Stage 6140 MeV/c
Mu, KL BG, KL Mu, SW BG, SW
Simulation
100% 100% 100% 100%
Digitisation
100% 100% 100% 100%
Fits 100% 100% 100% 100%
RootEvent
100% 100% Problem!(bug 107)
Problem!(bug 107)
Neural Net 100% 0%
15
Stage 6, 170±17 MeV/c
Stage 6170 MeV/c
Mu, KL BG, KL Mu, SW BG, SW
Simulation
100% 100% 100% 100%
Digitisation
100% 100% 100% Problem!(bug 107)
Fits 100% 100% 100% 100%
RootEvent
100% 0% Problem!(bug 107)
0%
Neural Net
0% 0%
16
Stage 6, 200±20 MeV/c
Stage 6200 MeV/c
Mu, KL BG, KL Mu, SW BG, SW
Simulation
100% 100% 100% 100%
Digitisation
100% 100% 100% Problem!(bug 107)
Fits 100% 100% 100% 100%
RootEvent
Problem!(bug 107)
0% 0% Problem!(bug 107)
Neural Net
0% 0%
17
Stage 6, 240±24 MeV/c
Stage 6240 MeV/c
Mu, KL BG, KL Mu, SW BG, SW
Simulation
0% 30% 100% 100%
Digitisation
0% 0% 100% Problem!(bug 107)
Fits 100% 100% 100% 100%
RootEvent
0% 0% 0% 0%
Neural Net
0% 0%
18
Stage 6, Tilley’s TURTLE beam
• A problem with the diffuser does not allow it to be placed.
• Without a diffuser, too low emittance.• If I have time I will try to solve the problem
before Japan.
19
Bug 107
• A vector in EmCalHit holding pointers to EmCalDigits seems to be corrupt.
• Very rare makes it hard to debug.– Why rare?
• Since the RootEvent converter uses the same class both Digitization and RootEvent suffers.
• Could be compiler/machine specific problem.– Then move all files to another computer, but we
are talking of ~ 50 Gb of data.
20
Results - Stage 1• Neural Net
– For training, used only muons which stayed muons until downstream TOF or beyond.• Same for pions.
– For testing, pions decaying to muons between TOFs where 1. treated as background. 2. omitted from analysis.
– KLOE Light:• Strongest variables are
based on: – tof, barycenter, and
fraction of energy in first layer.
– Sandwich:• Strongest variables are
based on: – tof, barycenter, and
total energy in calorimeter
Signal acc. BG rej.(with ->µ)
BG rej (no ->µ)
99.5% 49.2% 54.4%
99.0% 56.3% 62.2%
90.0% 80.5% 86.5%
Signal acc. BG rej.(with ->µ)
BG rej (no ->µ)
99.5% 61.0% 68.1%
99.0% 68.2% 75.3%
90.0% 79.1% 84.1%
KLOE Light
Sandwich
21
Results - Stage 6
• Only 140 MeV/c, KLOE Light is finished.– Results are very promising, but I wait with
presenting them until I can compare the different detectors.
22
Comments
• All momentum and tof measurements are MC truth.– Still waiting for tracker reconstruction to come
back online.– For tof, might simply add a Gaussian.
23
Summary
• Stage 1 is finished– Only a matter of how to present it.
• Most of stage 6 is simulated, but only partly digitized.– A bug most be fixed to continue.
• The first stage 6 beam that could be analyzed looks promising.