Machine learning assisted parameter-space searches for lensed gravitational waves

When a gravitational wave encounters a massive object along the line of sight, repeated copies of the original signal may be produced due to gravitational lensing. In this paper, we develop a series of new machine-learning-based statistical methods to identify promising strong lensing candidates in gravitational wave catalogs. We employ state-of-the-art normalizing flow generative models to perform statistical calculations on the posterior distributions of gravitational wave events that would otherwise be computationally unfeasible. Our lensing identification strategy, developed on two simulated gravitational wave catalogs that test noise realization and event signal variations, selects event pairs with low parameter differences in the optimal detector basis that also have a high information content and favorable likelihood for coincident parameters. We then apply our method to the GWTC-3 catalog and find a single pair still consistent with the lensing hypothesis. This pair has been previously identified through more costly evidence ratio techniques, but rejected on astrophysical grounds, which further validates our technique.