PSA with Neural Network, simulation

prudencewooshΤεχνίτη Νοημοσύνη και Ρομποτική

19 Οκτ 2013 (πριν από 3 χρόνια και 1 μήνα)

56 εμφανίσεις

PSA with Neural Network,simulation
Oleksandr Volynets
Max-Planck-Institute for Physics
ddate:Pungenday,Chaos 68,3178 YOLD/Mar 09,2012
Alex (MPI for Physics)
PSA with Neural Network,simulation
Mar 09,2012 1/13
A pulse
Idea:use features of individual pulses to distinguish signal/background;
single-site/multi-site
Time [ns]
0 200 400 600 800 1000 1200 1400 1600
Normalized charge [A.U.]
0
0.2
0.4
0.6
0.8
1
Pulse example
Alex (MPI for Physics)
PSA with Neural Network,simulation
Mar 09,2012 2/13
Simulation:training configuration
Two event samples of interest:

212
Bi,1620 keV:MSE dominated

208
Tl DEP,1592 keV:SSE dominated
Two source configurations (
only homog.will be discussed
):

“Homogeneous” set:

SSE
:DEP generated homogeneously in the detector (forcing pair
production at given points),WITHOUT Compton and bremsstrahlung.

MSE
:1620 gammas shot from a disk on top of the detector,only
Compton+bremsstrahlung selected;

“Top” set:both 1620 keV and 2.6 MeV gammas shot from a disk on
top of the detector,with/without Compton+bremsstrahlung as in
“Homog.” set
Alex (MPI for Physics)
PSA with Neural Network,simulation
Mar 09,2012 3/13
Reminder:NN parameters
NN input:

2 types of pulses:background DEP 1593 keV (type 0.0),signal 1620
keV (type 1.0);

40 sample of each pulse (533 ns),50 % ± 20 samples;
NN parameters:

1 input layer with 40 neurons;

1 hidden layer with 40 neurons;

1 output neuron:pulse type.
Alex (MPI for Physics)
PSA with Neural Network,simulation
Mar 09,2012 4/13
Reminder:Training output,homog.set
­0.4 ­0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4
0
100
200
300
400
500
Background
Signal
Neural net output (neuron 0)
Alex (MPI for Physics)
PSA with Neural Network,simulation
Mar 09,2012 5/13
Reminder:Rejection-efficiency compromise

REJ ×EFF
0 0.2 0.4 0.6 0.8 1 1.2
0
0.2
0.4
0.6
0.8
1
The maximum corresponds to the optimal value of the cut
Alex (MPI for Physics)
PSA with Neural Network,simulation
Mar 09,2012 6/13
Bug/feature found
If taking 50% ± 20 samples,and first sample is <0th,shift the region
until it is 0th.
Time [ns]
0 200 400 600 800 1000 1200 1400 1600
Normalized charge [A.U.]
0
0.2
0.4
0.6
0.8
1
Pulse example
Additionally,the jitter shifts the pulse →by rand[0–6].This
compensated
the bug by chance.
Alex (MPI for Physics)
PSA with Neural Network,simulation
Mar 09,2012 7/13
Feature 2
Distr.of pulse samples,e.g.50%-7 (7th to left from the central sample
used):
Charge amplitude
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
Entries
0
200
400
600
800
1000
1200
1400
Signal, DEP, training
Background, 1620, training
0nbb, evaluation
Position dependent!Each peak corresponds to particular hit positions -
checked.
Alex (MPI for Physics)
PSA with Neural Network,simulation
Mar 09,2012 8/13
Application to 0νββ
Old (with the bug/feature):
x [mm]
­40 ­30 ­20 ­10 0 10 20 30 40
y [mm]
­40
­30
­20
­10
0
10
20
30
40
0
0.2
0.4
0.6
0.8
1
Alex (MPI for Physics)
PSA with Neural Network,simulation
Mar 09,2012 9/13
Application to 0νββ
New (NN>0.55):
­40 ­30 ­20 ­10 0 10 20 30 40
­40
­30
­20
­10
0
10
20
30
40
Alex (MPI for Physics)
PSA with Neural Network,simulation
Mar 09,2012 10/13
Application to 0νββ:variations
New (NN > 0.45):
­40 ­30 ­20 ­10 0 10 20 30 40
­40
­30
­20
­10
0
10
20
30
40
Alex (MPI for Physics)
PSA with Neural Network,simulation
Mar 09,2012 11/13
Application to 0νββ:variations
New (NN > 0.8):
­40 ­30 ­20 ­10 0 10 20 30 40
­40
­30
­20
­10
0
10
20
30
40
Alex (MPI for Physics)
PSA with Neural Network,simulation
Mar 09,2012 12/13
Plan
s

Try with 1125 MHz and stretch/squeeze pulses to normalize by rise
time - probably removes the inpact of the rise time parameter
Alex (MPI for Physics)
PSA with Neural Network,simulation
Mar 09,2012 13/13