A Machine Learning Application in Collider Physics: Distinguishing Between Different Kinds of Multiple-Displaced-Vertex Events.
Lafayette College
File access restricted to members of the Lafayette College community.
Abstract
In Ref. [1] my collaborators and I proposed a new kind of experimental signature, called a tumbler, which could be detectable and discernible signal at the Large Hadron Collider (LHC). In this paper we discussed the underlying physics of a model for dark matter that could produce this tumbler decay, how to turn data taken at the LHC into useful information about the particles involved, and how to address issues that could arise in data manipulation and the detection of these phenomena. One of these issues was that decays can occur in a particle detector which look like tumbler events, but are actually the result of different physical phenomena. Specifically, a tumbler is a sequential decay chain of undetectable particles, i.e., processes in which one undetectable particle decays into at least two detectable particles and another undetectable particle, this undetectable particle in turn decays in to another two or more detectable particles and another undetectable particle, and so on. However, if two undetectable particles that are not from the same decay chain happen to decay near each other, or in some pattern that otherwise makes the decays seem connected, it is possible that an event that is not a tumbler could be mistaken for one.
In our previous paper, we addressed this issue in an “aggregate” fashion; we used well motivated data cuts, which is to say we developed background-rejection techniques, on a simulated set of events believed to be tumblers in order to help us infer the values for physical quantities which describe the new particles involved in these events — quantities which can be extracted more readily from the tumbler events than from non-tumbler events because of the characteristic kinematic properties of the tumbler events. This approach works as long as an appreciable number of events are measured. However, if only a handful of events of this type were to be observed, using our previous data techniques may not be optimal or even possible as there might not be enough data to make any statistically meaningful observations. Here, we investigate the prospects of using a machine learning classifier to distinguish between simulated tumbler and non-tumbler (non-sequential) multiple decay events. If a sufficiently general classifier can be constructed, it could be possible to use a classifier trained on simulated data to make a prediction as to whether an experimental event was a tumbler or not. Thus, we could potentially identify a signature of tumblers at the LHC or other future colliders on the basis of far less data – data which takes massive amounts of time and energy to collect.
Description
Honors thesis that investigates the prospects of using a machine learning classifier to distinguish between simulated tumbler and non-tumbler (non-sequential) multiple decay events. If a sufficiently general classifier can be constructed, it could be possible to use a classifier trained on simulated data to make a prediction as to whether an experimental event was a tumbler or not. Thus, potentially a signature of tumblers at the Large Hadron Collider or other future colliders could be identified on the basis of far less data – data which takes massive amounts of time and energy to collect.
Title
A Machine Learning Application in Collider Physics: Distinguishing Between Different Kinds of Multiple-Displaced-Vertex Events.
Digital collection of student honors theses, beginning in academic year 2021-2022.
Past theses written by Lafayette students through academic year 2020-2021 are kept in Special Collections and College Archives. Information about the honors theses in Special Collections is available in the Library Catalog.