Machine-learning interpolation of population-synthesis simulations to interpret gravitational-wave observations: a case study

TL;DR

This paper develops a machine-learning emulator trained on population-synthesis simulations to interpret gravitational-wave data, enabling more accurate astrophysical inferences about black hole binaries.

Contribution

It introduces a novel machine-learning approach integrated with Bayesian analysis to interpolate simulation results for gravitational-wave event interpretation.

Findings

Black holes in binaries likely receive natal kicks with velocity dispersion ~105 km/s.

Omission of some event parameters can cause systematic errors in population inference.

The method demonstrates potential for improved astrophysical insights from gravitational-wave observations.

Abstract

We report on advances to interpret current and future gravitational-wave events in light of astrophysical simulations. A machine-learning emulator is trained on numerical population-synthesis predictions and inserted into a Bayesian hierarchical framework. In this case study, a modest but state-of-the-art suite of simulations of isolated binary stars is interpolated across two event parameters and one population parameter. The validation process of our pipelines highlights how omitting some of the event parameters might cause errors in estimating selection effects, which propagates as systematics to the final population inference. Using LIGO/Virgo data from O1 and O2 we infer that black holes in binaries are most likely to receive natal kicks with one-dimensional velocity dispersion $\sigma$ = 105+44 km/s. Our results showcase potential applications of machine-learning tools in conjunction with population-synthesis simulations and gravitational-wave data.