Searching for gravitational waves from binary close encounters: a deep learning analysis approach

Since 2015 Advanced LIGO and Virgo have detected 93 gravitational wave transients produced by compact object binaries in quasi-circular orbits. Dynamical capture scenarios in dense stellar environments are predicted to form eccentric orbits persisting until the merger. At each periastron passage, the close encounters between the member of the compact object pair should emit gravitational wave signals. Waveforms expected in this scenario are not modeled with the same precision as coalescences, making it a good area to explore innovative analysis methods. The thesis is focused on the application of deep learning to detect gravitational waves from close encounters in compact binary systems. We developed a classification algorithm based on convolutional neural networks capable of detecting transient signals associated with close encounters and discriminating them from noise glitches. The algorithm training is based on custom simulated data of close encounters embedded in colored noise from Advanced Virgo. To better study the contamination by noise glitches, we developed a tool to test Gaussianity based on the Rayleigh test, which has been tested on real Virgo data, and an interactive web analysis tool for the quick look of glitches. The performance of the deep learning architecture developed in this thesis has an accuracy (~ 99.9%) comparable to other deep learning-based studies on gravitational wave transients. The results open new observational scenarios for the detection of close encounters, with the final goal of making them interesting targets for future electromagnetic follow-up campaigns.