Summary of downstream PID MICE collaboration meeting Fermilab 2006-06-10 Rikard Sandström
Dec 19, 2015
Outline
• Modular analysis approach• Fits• PID variables• PID performance• PID and emittance• Which events are miss identified?• New useful tool?• Summary
Modular analysis
• All fits used in analysis has been updated.– Wider range in all phase space coordinates
• Robustness!
• Fits are using a modular approach.– For example, given (x,y,z,t,px,py,pz,E) at exit of
tracker, return expected track parameters after TOF2.
– The same thing is then repeated for track after layer 0 of calorimeter, but using the track that was expected coming out of TOF2.
– If one object changes, or fits are deemed poor, only that module needs refitting!
• Robustness!
Energy loss fits
• Bethe-Bloch:
• Difficult for the fitter, Taylor expand!– dE/dx = k0/β2 + k1/ β + k2 + k3 β + k4 β2 +ordo(β3)
• Path length correction by cos()
(β = p/E)
Range in EMCal layer 1-10
• Range is well parameterized by βγ.• Due to projection onto z axis, correct with a factor
cos().
Visible energy EMCal layer 1-10
• This kinetic energy after layer 0.• Note how the calorimeter starts leaking energy at high E.
Reconstructed E vs MC Truth range
• Distribution of ADC counts in layers gives information on Bragg peak -> muon range & momentum
• For muons punching through Bragg peak is often not found.
Fits -> Useful PID variables
• Given the fits and information in a detector, the response in of another variable can be anticipated. – If the expected value is wrong the particle could be
background.
• Discrepancy variables:– D = 1-expected/measured– D = 0 indicates signal event (muon)
• Used in Neural Net analysis– Barycenter disc, total ADC disc, tof, tof disc, range disc,
tdc peaks, holes/range, high threshold adc/ low threshold adc, adc layer0/ total adc, adc layer0/ adc layer1.
– Would need a better tool for evaluating their real impact on PID.
• Some variables are correlated.
PID performance achieved
• Scenario: – Aug’05 with 7.6mm diffuser– RF turned off, empty absorbers
• With same statistics as in Osaka, – Efficiency 99.90% -> rejecting 97.7%
• Input purity 99.58% ->final purity 99.99%
– Efficiency 99.87% -> rejecting 99.5%
• Curiously, exactly same data but fewer events gave– Efficiency 99.90% -> rejecting 99.4%
• These results are even better than what I presented 2 days ago.– Due to longer NN training.
Steepness!
Impact on emittance
• Is 99.9% efficiency, 99.8% purity enough?
• I used no field approximation (thanks Chris)– ε = sqrt(x
2y2-xy
2)/m
– Exit of TOF2, probably worst difference between signal & background.
• x direction– ε = 10.3 pi mm rad– dε/ ε = 0.89 ppm
• y direction– ε = 9.9 pi mm rad– dε/ ε = 0.58 ppm
Impact of wrong badness tagging
• Assuming PID is perfect, what is the effect of poor tagging as good/bad event?– x, px gave 0.22 ppm difference at exit of TOF2.
• Effect is lower than PID, but at the same scale.
• Cause exclusively by tracker and tof resolutions.– Miss tagging straight tracks and tracks close to
active volume edges.
• Note: emittance is using reconstructed tracks (smeared) for both MC truth badness and reconstructed badness.
What signal events are miss IDed?
• For w<0.2, – 8% of muons are stopped in TOF2.
• Will be worse a lower momentum.
– 8% are leaving TOF2 with very large angle and misses calorimeter.
• Will be worse a lower momentum.• Move calorimeter even closer to TOF2.
– 8% decay between TOF2 and calorimeter.• Move calorimeter even closer to TOF2.
– 60% are muons decaying in ADC gate and too close in time to its own track that only one TDC peak is registered.
• Tweaking TDC threshold could help.• Harmless!
New useful tool?
• Stumbled upon a ROOT package, T Multi Variable Analysis (TMVA) which is made for evaluating different methods of separating signal from background.
• Yesterday, I installed it and ran it.– To my big surprise it worked
without problem.
• We could use this tool to see if a better way than using my neural net exists.– Can use the ROOT trees I
have already prepared for comparison.