Millisecond pulsars are among nature’s most precise clocks, and together are used as a galaxy-sized detector for low-frequency gravitational waves (GWs) called a pulsar timing array (PTA). While the main goal of a PTA is GW detection, there are a variety of astrophysical effects, such as pulse propagation effects from the interstellar medium (ISM) that are radio-frequency dependent, that also affect pulsar timing. These are interesting effects to study in their own right, but they also complicate GW analyses by acting as sources of noise. Since every pulsar has different intrinsic properties, is located in a different location of the galaxy, and has been observed for different amounts of time and by different telescope receivers, the observed noise in each pulsar may vary dramatically. This suggests we use noise models which are individually-tailored to different pulsars when performing GW searches.

This talk will focus on how different pulsar noise models can be implemented and compared under the highly flexible and interpretable formalism of Gaussian processes, a workhorse tool in PTA analyses. We will first look at a comparison of different noise models in 6 pulsars from the NANOGrav 15-year dataset, as well as how those same models perform when applied to those same pulsars from a different dataset, EPTA DR2. We focus specifically on identifying cases of model misspecification and understanding how the misspecification biases GW measurements. We will then look at some preliminary results of creating tailored noise models for each pulsar in the NANOGrav 15-year dataset using Bayesian model selection methods, a project in collaboration with the Hazgrav lab. We will end by discussing how combining data from multiple PTAs can dramatically improve our characterization of pulsar noise.

Bio: Bjorn Larsen is a 4th year PhD student in NANOGrav and the Mingarelli Lab at Yale. His primary research interests include detecting gravitational waves and modeling noise in PTA datasets, learning how to visualize these interesting and complicated signals, and trying not to misuse Bayesian statistics.