Master's Thesis : Lightning Gravitational Wave Parameter Inference through Neural Amortization

Abstract

Gravitational waves analysis relies on a simulator governed by the nonlinear field equations of general relativity for binary systems. Such analysis is computationally very expensive and necessitates a large-scale exploration of the likelihood surface over the full parameter space. Neural networks have been gaining popularity as tools for gravitational waves analysis for the last few years. They lead to fast gravitational wave detection and parameter inference and hence complement classical slower techniques. This work explores simulation based inference which relies on likelihood-to-evidence ratio estimation for the parameters of binary black holes mergers. We build a neural network modeling this ratio and use it in place of the simulator allowing to perform parameter inference in few minutes. The performances are assessed on both gravitational wave generated by the simulator and emitted by real black holes.