1. Identifying Particle
Ancestor using Compton
Scatter Analysis for EXO-
200
CS Division III by Beryl Bell
Supervised by Andrea Pocar, Sarah Hews, Lee Spector, and
Sereres Johnson

2. Simple Definition
The input data is a series of charge
clusters, with 3 spatial coordinates and
one energy value
Want to group these clusters based on
the particle that created them
The Compton equation relates angle of
scatter to energy difference
2

3. Simple Definition
The input data is a series of charge
clusters, with 3 spatial coordinates and
one energy value
Want to group these clusters based on
the particle that created them
The Compton equation relates angle of
scatter to energy difference
3

4. Outline
1. Physics Background
a. Brief introduction to neutrinos
b. What is neutrinoless double beta decay and what are its
implications
c. What decay process is being examined here
2. Compton Algorithm
a. How will the algorithm be incorporated into the machine learning
framework
b. Structure of the Compton algorithm
c. Performance of the algorithm and future work
4

5. Physics Background
5

6. The Standard Model1
Currently the most accurate model of
physics that we know
A definition of particles and the ways
that these particles interact
1: Image from Symmetry Magazine
6

7. Neutrinos
Very light, uncharged particles
Difficult to detect due to low rates of
interaction
First theorized about when electrons
emitted in decay processes had a
spectrum of energy instead of single
value
Potential Majorana particle
7

8. Neutron undergoing beta decay2
2: Image from Wikipedia 8

9. Double Beta Decay3
Two neutron undergo beta decay side
by side and emit two electrons and two
neutrinos
Relatively common process
Xenon 136 commonly undergoes this
process, the EXO200 project uses 200
kg of Xenon to conduct its search
3: Image from Erlangen Centre for Astroparticle Physics website
9

10. Neutrinoless Double Beta Decay4
Only possible if the neutrino is a
Majorana particle
The first emitted neutrino essentially
triggers the second decay, no neutrinos
are released in this decay
Much rarer process
4: Image from Erlangen Centre for Astroparticle Physics website
10

11. Compton Algorithm
11

12. Application of Compton Algorithm
Used for the analysis of the excited state decay
A machine learning process uses different parameters to
differentiate between signal and background
It has been shown that the Compton algorithm, working with
100% accuracy, could double the machine learning’s ability to
differentiate between signal and background
12

13. Excited State5
The decay to the 01
+
excited state has
never been observed before
This decay emits two photons of known
energies
Characterization of this decay would
allow for more accurate
characterization of the neutrinoless
double beta decay
5: From Inspire HEP webpage 13

14. Excited State5
In the detector this decay is detected as
a group of charge clusters, resulting
from the two photons and the electrons
In some events part of this energy
escapes or remains undetected, either
the photon energy or the electron
energy
5: From Inspire HEP webpage 14

15. Compton Equation6
156: Image from Wikipedia

16. Minimizing Angle Difference
The angle difference is the difference
between the scaler angle among
clusters (θ) and the angle calculated
using the Compton equation (θE
)
The Compton algorithm attempts to
find the grouping of clusters that
minimizes this angle difference, called
superclusters
The Compton algorithm tests all the
permutations
16

17. Histograms of Cluster Energy
Performs at 33% accuracy, same as random assignment
17

18. Addition of Energy
Weighting
In order to account for some instances
where to algorithm struggled, a
matching of the total supercluster
energy to the expected energy
This difference was then also
minimized, with a weight to make the
scale comparable to the angle
difference
Over a range of weights, a maximum
was found
18

19. Histograms of Cluster Energy
Performs at 43% accuracy, better by ~9%
19

20. Challenges to Algorithm
Events with only three clusters are difficult to categorize
using scatters as no scatters can be draw. This is partially
addressed with the energy weighting
20

21. Challenges to Algorithm For Future Work
Some clusters have a mixture of energy from both the
electrons and the photons, making the Compton algorithm
ineffective
Not all events have electron clusters, but the algorithm
assumes one as the origin. Alternative definitions for the
origin definition could be found in the future
Modifications to the energy resolution of the clusters would
make the algorithm more accurate overall
21

22. Conclusion
The Compton algorithm by itself does not provide enough
accuracy to be used by the machine learning program at this
time
Future work could bring the algorithm to the point where it is
a useful parameter for distinguishing between signal and
background events
22

23. Extra Slides
23

24. Correlation of Angle Difference and
Accuracy
Comparison of the minimized angle
difference and the accuracy shows that
a larger accuracy does not always
correlate with a lower minimized angle
difference
24

25. Boosted Decision
Trees
Machine learning process generates a
forest of boosted decision trees
These trees, based on input
parameters, split events into
background and signal events
All the trees vote to classify a single
event as somewhere between
background (-1) and signal (1)
25

26. Detector
The EXO-200 detector is a Time
Projection Chamber (TCP) which
identifies the x-y and energy
coordinates with induction wires and
calculates the z coordinate by the
difference between the detection of the
ionization and scintillation events
26

27. Detailed Compton
Algorithm
Given an event, such as the
following, the algorithm assigns
the first cluster as the electron
cluster ( the beta cluster) and the
rest as 818, then calculate the
average angle difference of these
assignments in this permutation
The algorithm then scrolls through
the rest of the assignments as
follows, calculating the angle
difference for each
27

28. Detailed Compton
Algorithm
Given an event, such as the
following, the algorithm assigns
the first cluster as the electron
cluster ( the beta cluster) and the
rest as 818, then calculate the
average angle difference of these
assignments in this permutation
The algorithm then scrolls through
the rest of the assignments as
follows, calculating the angle
difference for each
28

29. Detailed Compton
Algorithm
Given an event, such as the
following, the algorithm assigns
the first cluster as the electron
cluster ( the beta cluster) and the
rest as 818, then calculate the
average angle difference of these
assignments in this permutation
The algorithm then scrolls through
the rest of the assignments as
follows, calculating the angle
difference for each
29

30. Detailed Compton
Algorithm
Given an event, such as the
following, the algorithm assigns
the first cluster as the electron
cluster ( the beta cluster) and the
rest as 818, then calculate the
average angle difference of these
assignments in this permutation
The algorithm then scrolls through
the rest of the assignments as
follows, calculating the angle
difference for each
30

31. Detailed Compton
Algorithm
Given an event, such as the
following, the algorithm assigns
the first cluster as the electron
cluster ( the beta cluster) and the
rest as 818, then calculate the
average angle difference of these
assignments in this permutation
The algorithm then scrolls through
the rest of the assignments as
follows, calculating the angle
difference for each
31

32. Detailed Compton
Algorithm
After scrolling through all these
steps, then generates the next
permutation to scroll through
Out of this process the
minimized average angle
difference is found
32