Sequentially learning gravity Scalable simulation-based inference for gravitational waves

Gravitational-wave astronomy has rapidly evolved into a precision science, demanding inference methods that scale with increasing data volumes, source complexity, and instrumental configurations. Traditional likelihood-based techniques, though robust, struggle under high-dimensional, computationally intensive scenarios—especially when dealing with overlapping signals from similar or varying source types. This thesis develops and applies scalable, simulation-based inference (SBI) frameworks tailored to gravitational-wave data analysis, where likelihood functions are often intractable.
At its core, the work leverages neural ratio estimation (NRE), a sequential SBI technique that recasts inference as a classification problem. Building on this, the thesis introduces two open-source pipelines: Peregrine for transient gravitational-wave events, and Saqqara for persistent signals such as the stochastic gravitational wave background. These pipelines implement Truncated Marginal Neural Ratio Estimation (TMNRE), an efficient, iterative variant of NRE designed to obtain precision by reducing data variance over a sequence of rounds.
Through case studies involving binary black hole (BBH) coalescences and overlapping transient events as well as coincident transient and persistent signals, this thesis demonstrates that the proposed SBI methods can match the precision of traditional approaches with high simulation efficiency. Extensive comparisons with likelihood-based samplers and posterior coverage tests validate their reliability. Overall, this work lays the foundation for tackling the increasingly complex and high-dimensional inference challenges that will define the future of gravitational-wave astrophysics.