Séminaires : Documents

Enhancing Bayesian analysis and post-processing with machine-learning
Virginia D'Emilio (Cardiff University)

Gravitational-wave (GW) astrophysics is largely based on Bayesian inference results of individual observations. These are what we call parameter estimation products. During the last GW observing run, the LIGO-Virgo-KAGRA collaboration detected around 90 signals from compact binaries coalescences. We expect the number of such detections to increase to O(1000) over the next two observing runs, which is challenging for our current analysis and post-processing pipelines. Several efforts in our community have been devoted to boost parameter estimation, especially with the use of machine-learning. In many cases, neural networks seem to offer a powerful alternative to tackle expensive computation, an example being simulation-based inference. But in other cases, as shown by Williams and Rasmussen (1996), real world data modelling problems are well solved by sensible smoothing methods, such as Gaussian Processes. In this talk, I will outline how both these techniques address important interdisciplinary problems and as such are both being investigated in GW and cosmology. Finally, as a concrete example of how they could improve the accuracy and speed of Bayesian data analysis, I will consider the case of dark sirens cosmology with GW.