Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Using Artificial Neural Networks to optimize scattering probabilities

Rustas, Erik LU (2021) FYTK02 20211
Theoretical Particle Physics
Abstract
Monte Carlo event generators are used by theoretical particle physicists to get a better understanding of the phenomena in particle physics. Given the improvements in precision and accuracy of event generators, using these tools can be very CPU intensive. A method of unweighting events using artificial neural networks is presented to improve the efficiency of event generation. An introduction to machine learning as well as an introduction to the unweighting procedure is given as a basis. Results are given by comparing the “classical” and artificial neural network unweighting. The efficiency is expressed as factors of computing time for the matrix element and the model’s predicted value.
Popular Abstract
For centuries humans believed that matter was made of indivisible small particles called atoms (from the word atomos Greek for "indivisible"). We now know that atoms are made of even smaller constituents of matter. These are called fundamental particles such as the electron or the Higgs boson. How these fundamental particles interact with each other is best explained by the theory called the Standard Model. Further research in the field of particle physics is still an ongoing subject even today.

One way of studying how particles interact is by colliding particles together such as the Large Hadron Collider at CERN. The Large Hadron Collider accelerates particles to nearly the speed of light in two different tubes that run opposite each... (More)
For centuries humans believed that matter was made of indivisible small particles called atoms (from the word atomos Greek for "indivisible"). We now know that atoms are made of even smaller constituents of matter. These are called fundamental particles such as the electron or the Higgs boson. How these fundamental particles interact with each other is best explained by the theory called the Standard Model. Further research in the field of particle physics is still an ongoing subject even today.

One way of studying how particles interact is by colliding particles together such as the Large Hadron Collider at CERN. The Large Hadron Collider accelerates particles to nearly the speed of light in two different tubes that run opposite each other in circular loops. When these tubes intersect, the particles collide into each other creating extremely high energy collisions. From these collisions one can analyze the data and find new particles. However one does not need a collider to extract useful information about particle physics. Theorists use computer simulation programs called event generators. These event generators simulate high energy particle collisions by Monte Carlo methods. The event generators are essential tools to theorists as they allow for confirmation of theoretical predictions. Given their significance, the search for higher precision composes a highly demanding computational cost. Optimizing these event generators using machine learning techniques will be this project's main goal.

Machine learning is part of the field of artificial intelligence, which learns by processing sample data. These types of artificial intelligence are used in for example data filtering such as spam mail. This project will use an artificial neural network, which is a branch of machine learning inspired by the structural construct of the brain. The brain is amazing at remembering and learning new things, so building a program inspired by how the brain works is only natural.

One of the Monte Carlo methods that event generators use is called the "unweighting" method. The generated events from the event generator have a value, or weight, mainly determined from the matrix element. Making these weighted events have a more nature-like distribution, where the events have an equal unit weight, is the process of unweighting. This process scales rapidly with the amount of particles in a given interaction. Where the complexity of the matrix element increases as well as the amount of times the matrix element has to be evaluated. Machine learning unweighting will be presented and tested in this project. This method will use mainly the artificial neural network's prediction of the matrix element instead of the actual matrix element. Machine learning and classical unweighting methods will be tested where the break even point of efficiency will be found. This is the factor that compares the time to do one evaluation of the matrix element, and the artificial neural networks prediction. The break-even point of efficiency will be how much more computational time the matrix element has to be for the machine learning unweighting to be a more favorable method. (Less)
Please use this url to cite or link to this publication:
author
Rustas, Erik LU
supervisor
organization
course
FYTK02 20211
year
type
M2 - Bachelor Degree
subject
keywords
Machine Learning, Artificial Neural Network, Monte Carlo Method, Unweighting
language
English
id
9057975
date added to LUP
2021-06-24 09:10:57
date last changed
2021-07-09 15:57:22
@misc{9057975,
  abstract     = {{Monte Carlo event generators are used by theoretical particle physicists to get a better understanding of the phenomena in particle physics. Given the improvements in precision and accuracy of event generators, using these tools can be very CPU intensive. A method of unweighting events using artificial neural networks is presented to improve the efficiency of event generation. An introduction to machine learning as well as an introduction to the unweighting procedure is given as a basis. Results are given by comparing the “classical” and artificial neural network unweighting. The efficiency is expressed as factors of computing time for the matrix element and the model’s predicted value.}},
  author       = {{Rustas, Erik}},
  language     = {{eng}},
  note         = {{Student Paper}},
  title        = {{Using Artificial Neural Networks to optimize scattering probabilities}},
  year         = {{2021}},
}