Thesis

The Laser Interferometer Space Antenna (LISA) is an upcoming space-based observatory designed to detect low-frequency gravitational waves, primarily from the collisions of supermassive black holes. Because these massive binaries can be observed days or weeks before they finally merge, LISA presents a groundbreaking opportunity for multi-messenger astronomy: issuing “early warning” alerts so that traditional electromagnetic-based telescopes can point at the right patch of sky to capture the explosive event simultaneously.

However, generating these real-time alerts poses a severe statistical challenge. Early in the observation, the faint signal yields highly ambiguous, “multimodal” data, meaning the physical symmetries of the detector make it look like the event could originate from up to eight completely different locations in the sky. Standard statistical sampling methods (like Markov Chain Monte Carlo) are far too slow to update these precise predictions within the required few-minute latency budget. Conversely, conventional neural networks struggle to adapt to the unpredictable, ever-changing nature of a live, interrupted space data stream.

This PhD project sits at the intersection of astrophysical data analysis, Bayesian statistics, and scientific machine learning to solve this challenge. The successful candidate will develop a fast, sequential Variational Inference (VI) framework that translates a slow sampling method into a rapid, GPU-accelerated optimization problem. By embedding the known physics of the LISA detector into the algorithm, the method will “fold” the complex probability space to fit the incoming data, and then “unfold” it to provide astronomers with a real-time, multi-modal map of where to look. Over three years, the student will scale this physics-informed machine learning approach from idealized models to a robust, open-source pipeline capable of handling realistic, noisy mission data, directly enabling LISA’s early-warning capabilities.