Parameter estimation of microlensed gravitational waves with Conditional Variational Autoencoders

Abstract

Gravitational lensing of gravitational waves (GWs) provides a unique opportunity to study cosmology and astrophysics at multiple scales. Detecting microlensing signatures, in particular, requires efficient parameter estimation methods due to the high computational cost of traditional Bayesian inference. In this paper we explore the use of deep learning, namely Conditional Variational Autoencoders (CVAE), to estimate parameters of microlensed binary black hole (simulated) waveforms. We find that our CVAE model yields accurate parameter estimation and significant computational savings compared to Bayesian methods such as Bilby (up to five orders of magnitude faster inferences). Moreover, the incorporation of CVAE-generated priors into Bilby, based on the 95% confidence intervals of the CVAE posterior for the lensing parameters, reduces Bilby's average runtime by around 48% without any penalty on accuracy. Our results suggest that a CVAE model is a promising tool for future low-latency searches of lensed signals. Further applications to actual signals and integration with advanced pipelines could help extend the capabilities of GW observatories in detecting microlensing events.