Using Pulsar Classifying Machine Learning Algorithms To Increase Pulsar Timing Array Sensitivity

A pulsar is a neutron star that emits electromagnetic radiation along its magnetic pole and spins extremely fast. Millisecond pulsars, in particular, rotate hundreds or thousands of times per second and to atomic clock-level precision. Knowing the precision of a millisecond pulsar allows scientists to observe gravitational waves using the discrepancies in pulse timing as measured on Earth. Discovering more pulsars will enable us to increase the sensitivity of these extensive collections of pulsar data, also known as Pulsar Timing Arrays. Radio telescopes detect new pulsar candidates, but the majority of candidate data consists of non-pulsar radio sources, such as Radio Frequency Interference (RFI) and noise. Humans have manually verified pulsar candidates, but using machine learning algorithms (MLA) to sort the candidates can save incredible amounts of time and increase efficiency. This project focuses on an existing pulsar classifying MLA, PulsAr Classifier - Machine-learning Algorithm with Neural Networks (PACMANN). One difficulty encountered during the use of PACMANN is that we have far more data for non-pulsar candidates than pulsar candidates, which means that the MLA is learning from either an unbalanced data set or an incredibly small one, which could decrease the MLA’s overall performance. Proposed solutions include injecting simulated pulsar signals into the dataset or balancing the existing data using class weights. For this project, we focused on implementing both solutions and aimed to use PACMANN to identify new pulsar candidates.