Simulation-based inference of black hole ringdowns in the time domain

Gravitational waves emitted by a ringing black hole allow us to perform precision tests of general relativity in the strong field regime. With improvements to our current gravitational wave detectors and upcoming next-generation detectors, developing likelihood-free parameter inference infrastructure is critical as we will face complications like nonstandard noise properties, partial data, and incomplete signal modeling that may not allow for an analytically tractable likelihood function. In this work, we present a proof-of-concept strategy to perform likelihood-free Bayesian inference on ringdown gravitational waves using simulation based inference. Specifically, our method is based on truncated sequential neural posterior estimation, which trains a neural density estimator of the posterior for a specific observed data segment. We setup the ringdown parameter estimation directly in the time domain. We show that the parameter estimation results obtained using our trained networks are in agreement with well-established Markov-chain methods for simulated injections as well as analysis on real detector data corresponding to GW150914. Additionally, to assess our approach's internal consistency, we show that the density estimators pass a Bayesian coverage test.