Sky localization of Un-modeled Gravitational Wave Sources with Machine Learning

During the first 3 observing runs the LIGO-Virgo-Kagra collaboration (LVK) has made over 90 detections of gravitational waves, all from compact binary coalescences (CBCs). Beyond these is another category of possible sources, short-duration (around ~1s), un-modeled transients. These sources do not have well-modeled templates of the gravitational wave emission. Plausible astrophysical progenitors include core-collapse supernovae (CCSNe), pulsar-glitches, with the additional possibility of detecting an unexpected source. Accurate and rapid parameter estimation of these un-modeled gravitational wave transients is vital for informing electromagnetic follow up. Many Bayesian methods can take hours to days to converge to accurate posterior estimates. As an alternative, machine learning approaches have been shown to offer comparable accuracies with significantly faster inference times. In this work, we present a machine learning framework that utilizes normalizing flows to perform real-time parameter estimation for un-modeled gravitational wave sources. To train our model, we use ad-hoc Sine-Gaussian waveforms that make minimal assumptions about the source. We demonstrate the ability of our model to accurately recover sky localization parameters from sources modeled with Sine-Gaussians embedded in real data from the LVK third observing run.