An Analysis into Multivariate Lepton Jet Recognition
Session Number
Project ID: PHYS 16
Advisor(s)
Dr. Peter J. Dong, Illinois Mathematics and Science Academy
Discipline
Physical Science
Start Date
17-4-2024 8:15 AM
End Date
17-4-2024 8:30 AM
Abstract
Dark photons are useful indicators to explain important phenomena beyond the Standard Model of particle physics, namely in the context of many experiments. Dark photons can interact with Standard Model particles through a process called kinetic mixing, which allows them to decay into Standard Model leptons, from which lepton jets are produced. However, classification of such particles becomes challenging when introducing background effects, systematic errors, and cut values that are difficult to calibrate. Therefore, we propose a methodology in which a multivariate analysis is used to produce a neural network that differentiates true signal events from fakes. Machine learning analyses can aid in fixing this issue; moreover, several neural net methods for cut efficiency and optimal cut values were employed: specifically, Boosted Decision Trees (BDT), Multi-Layer Perceptrons (MLP), and Deep Neural Networks (DNN)’s. In summary, the most successful method of separation was through the use of BDT, where we observed optimal signal acceptance and through efficiency ROC curves; we also concluded that the current efficiency for our BDT experiment was maximized at around 92% signal acceptance, and 93% background rejection. MLPs also had remarkably similar background rejection and signal efficiency, but were generally outperformed by BDTs.
An Analysis into Multivariate Lepton Jet Recognition
Dark photons are useful indicators to explain important phenomena beyond the Standard Model of particle physics, namely in the context of many experiments. Dark photons can interact with Standard Model particles through a process called kinetic mixing, which allows them to decay into Standard Model leptons, from which lepton jets are produced. However, classification of such particles becomes challenging when introducing background effects, systematic errors, and cut values that are difficult to calibrate. Therefore, we propose a methodology in which a multivariate analysis is used to produce a neural network that differentiates true signal events from fakes. Machine learning analyses can aid in fixing this issue; moreover, several neural net methods for cut efficiency and optimal cut values were employed: specifically, Boosted Decision Trees (BDT), Multi-Layer Perceptrons (MLP), and Deep Neural Networks (DNN)’s. In summary, the most successful method of separation was through the use of BDT, where we observed optimal signal acceptance and through efficiency ROC curves; we also concluded that the current efficiency for our BDT experiment was maximized at around 92% signal acceptance, and 93% background rejection. MLPs also had remarkably similar background rejection and signal efficiency, but were generally outperformed by BDTs.