Trigger-Level Electron Counting in the Light Dark Matter eXperiment using Artificial Neural Networks
(2024) FYSK04 20241Department of Physics
Particle and nuclear physics
- Abstract
- The Light Dark Matter eXperiment is a fixed target missing-momentum experiment that searches for light dark matter production via the process of Dark Bremsstrahlung by analysing the energy of beam electrons after they hit a tungsten target. Electron counting is an important part of the experiment as this forms one of the two components of the missing energy trigger. The thesis looks into viability of Artificial Neural Network models in the electron counting procedure, as electron counting is a classification problem, a type of problem in which ANNs have seen large success.
The project looks into two ANN architectures, Convolutional Neural Networks and Recurring Neural Networks. The ultimate aim is an accuracy of 1 misclassification in... (More) - The Light Dark Matter eXperiment is a fixed target missing-momentum experiment that searches for light dark matter production via the process of Dark Bremsstrahlung by analysing the energy of beam electrons after they hit a tungsten target. Electron counting is an important part of the experiment as this forms one of the two components of the missing energy trigger. The thesis looks into viability of Artificial Neural Network models in the electron counting procedure, as electron counting is a classification problem, a type of problem in which ANNs have seen large success.
The project looks into two ANN architectures, Convolutional Neural Networks and Recurring Neural Networks. The ultimate aim is an accuracy of 1 misclassification in 100000 events. We analyze the difference between the performance of various architectures. The best results is a CNN network with 0.9944 accuracy and an RNN network with 0.9598 accuracy. RNN architecture misclassified events mainly based on the number of readout hits present in an event, while the CNN relies on a combination of total energy and event geometry when classifying an event. In particular we show that performance is not degraded when using a CNN with data pooled in pre-processing to match the "Trigger Cells" which will provide the calorimetric input to the trigger decision. In contrast, both shorter training time and an improved accuracy can be obtained. (Less) - Popular Abstract
- Dark matter is estimated to make up about a quarter of our universe. This is about five times more than the scientific community has been able to describe using the largely successful Standard Model. Dark matter does not interact electromagnetically and thus is invisible for the majority of techniques used by scientists to explore matter. We know little about the nature of dark matter and its properties. The reason we know of dark matter's existence is because we can infer it from the gravitational effects it has on the visible matter around it.
The Light Dark Matter eXperiment aims to explore dark matter production through a process called dark photon bremsstrahlung. In this process a dark photon is released from an electron close to a... (More) - Dark matter is estimated to make up about a quarter of our universe. This is about five times more than the scientific community has been able to describe using the largely successful Standard Model. Dark matter does not interact electromagnetically and thus is invisible for the majority of techniques used by scientists to explore matter. We know little about the nature of dark matter and its properties. The reason we know of dark matter's existence is because we can infer it from the gravitational effects it has on the visible matter around it.
The Light Dark Matter eXperiment aims to explore dark matter production through a process called dark photon bremsstrahlung. In this process a dark photon is released from an electron close to a tungsten nucleus. This dark photon is theorised to decay into a dark particle anti-particle pair. In the experiment a beam of electrons is fired at a thin tungsten target with the energy of electrons being recorded in a calorimeter. A large amount of missing energy could indicate dark matter production.
Dark photon bremsstrahlung is expected to be a rare event. The electron beam produces 37 million events every second, however the majority of the events produced have no value to the research. Only 5 thousand events are to be read out by the the detector for analysis. In order to find the possible dark bremsstrahlung events, a system of triggers detects if the total deposited energy is lower than a cutoff energy. The cutoff energy varies based on the electron multiplicity of the event. Thus for the appropriate cutoff energy to be applied one needs to know the electron multiplicity of the event.
Artificial Neural Networks have been historically strong in solving classification problems. The electron counting is a classification problem based on the inputs from the electromagnetic calorimeter and the trigger scintillator of the LDMX detector. It was shown in previous research that Artificial Neural Networks are capable of achieving better than 95% classification accuracy. However the "black box" nature of these models means that we have little knowledge of what the classification is based on. This project analyzes the effectiveness of Artificial Neural Networks in counting electrons in the LDMX detector, compares the performance of various Artificial Neural Network architectures to understand what event qualities affect the classification process. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9165518
- author
- Oshchepkov, Pavel LU
- supervisor
-
- Lene Bryngemark LU
- Ruth Pöttgen LU
- organization
- course
- FYSK04 20241
- year
- 2024
- type
- M2 - Bachelor Degree
- subject
- keywords
- LDMX, Dark Matter, Artificial Neural Network, CNN, RNN
- language
- English
- id
- 9165518
- date added to LUP
- 2024-06-19 10:14:20
- date last changed
- 2024-06-19 10:14:20
@misc{9165518,
abstract = {{The Light Dark Matter eXperiment is a fixed target missing-momentum experiment that searches for light dark matter production via the process of Dark Bremsstrahlung by analysing the energy of beam electrons after they hit a tungsten target. Electron counting is an important part of the experiment as this forms one of the two components of the missing energy trigger. The thesis looks into viability of Artificial Neural Network models in the electron counting procedure, as electron counting is a classification problem, a type of problem in which ANNs have seen large success.
The project looks into two ANN architectures, Convolutional Neural Networks and Recurring Neural Networks. The ultimate aim is an accuracy of 1 misclassification in 100000 events. We analyze the difference between the performance of various architectures. The best results is a CNN network with 0.9944 accuracy and an RNN network with 0.9598 accuracy. RNN architecture misclassified events mainly based on the number of readout hits present in an event, while the CNN relies on a combination of total energy and event geometry when classifying an event. In particular we show that performance is not degraded when using a CNN with data pooled in pre-processing to match the "Trigger Cells" which will provide the calorimetric input to the trigger decision. In contrast, both shorter training time and an improved accuracy can be obtained.}},
author = {{Oshchepkov, Pavel}},
language = {{eng}},
note = {{Student Paper}},
title = {{Trigger-Level Electron Counting in the Light Dark Matter eXperiment using Artificial Neural Networks}},
year = {{2024}},
}