Using Neural Networks to Perform Rapid High-Dimensional Kilonova Parameter Inference

Abstract

On the 17th of August, 2017 came the simultaneous detections of GW170817, a gravitational wave that originated from the coalescence of two neutron stars, along with the gamma-ray burst GRB170817A, and the kilonova counterpart AT2017gfo. Since then, there has been much excitement surrounding the study of neutron star mergers, both observationally, using a variety of tools, and theoretically, with the development of complex models describing the gravitational-wave and electromagnetic signals. In this work, we improve upon our pipeline to infer kilonova properties from observed light-curves by employing a Neural-Network framework that reduces execution time and handles much larger simulation sets than previously possible. In particular, we use the radiative transfer code POSSIS to construct 5-dimensional kilonova grids where we employ different functional forms for the angular dependence of the dynamical ejecta component. We find that incorporating an angular dependence improves the fit to the AT2017gfo light-curves by up to ~50% when quantified in terms of the weighted Mean Square Error.