Deep learning to detect gravitational waves from binary close encounters: Fast parameter estimation using normalizing flows

A yet undetected class of gravitational wave signals is represented by the close encounters between compact objects in highly-eccentric (e1) orbits, that can occur in binary compact systems formed in dense environments such as globular clusters. The expected gravitational signals from these close encounters are short-duration pulses that would repeat over a much longer timescale in case of multiple passages at periastron. These sources represent a unique opportunity of exploring astrophysical formation channels as well as a different way of testing general relativity. Furthermore, in the case of binary systems containing neutron stars, the observation of these sources could help to constrain the neutron star equation of state, thanks to the signature left in the gravitational wave signal by the f-modes excitation that can occur during the encounter. The detection and parameter estimation of these signals is however challenging given the short duration of expected signals and the sensitivities of current ground-based gravitational wave interferometers. We present a novel approach to perform fast detection and parameter estimation of gravitational wave signals from binary close encounters that exploits probabilistic machine learning. We have used conditional normalizing flows to model complex probability distributions and therefore infer posterior distributions for the source parameters. This architecture is able to perform inference in a very short time and its output can be directly compared with classical methods. Fast detection and parameter estimation is very important as it could trigger electromagnetic follow-up campaigns and offer the possibility to study these events in a multimessenger context. To develop and test the algorithm, we have focused on the simulations of single bursts emission obtained using the Effective Fly-by formalism and embedded in the noise of Advanced LIGO and Virgo during their third observing run (O3). Our proposed model outperforms standard Bayesian methods in accuracy and is 5 orders of magnitude faster, being able to produce 510^4 posterior samples in just 0. 5 s. The results are extremely promising and constitute the first successful attempt for a fast and complete parameter estimation of binary close encounters using deep learning, offering a new approach to study the evolution of orbital parameters of compact binary systems.