Neural likelihood estimators for flexible gravitational wave data analysis

In this paper, we develop a Neural Likelihood Estimator and apply it to analyse real gravitational-wave (GW) data for the first time. We assess the usability of neural likelihood for GW parameter estimation and report the parameter space where neural likelihood performs as a robust estimator to output posterior probability distributions using modest computational resources. In addition, we demonstrate that the trained Neural likelihood can also be used in further analysis, enabling us to obtain the evidence corresponding to a hypothesis, making our method a complete tool for parameter estimation. Particularly, our method requires around 100 times fewer likelihood evaluations than standard Bayesian algorithms to infer properties of a GW signal from a binary black hole system as observed by current generation ground-based detectors. The fairly simple neural network architecture chosen makes for cheap training, which allows our method to be used on-the-fly without the need for special hardware and ensures our method is flexible to use any waveform model, noise model, or prior. We show results from simulations as well as results from GW150914 as proof of the effectiveness of our algorithm.