Events

Real-Time Gravitational Wave Science with Neural Posterior Estimation

Speaker: Maximilian Dax

Location: 60 Fifth Avenue, Room 7th floor

Date: Thursday, February 2, 2023

We introduce Dingo, a machine-learning approach for rapid gravitational wave (GW) inference with unprecedented accuracy. For the first time, this enables accurate real-time analysis of black hole mergers. Dingo could be a game changer for GW science, as it enables various analyses that have previously been computationally prohibitive. We thus expect it to make a significant contribution to future data analysis by the LIGO-Virgo-KAGRA collaboration.

Dingo builds on conditional normalizing flows as surrogates for Bayesian posteriors, trained in a simulation-based inference setting. Our equivariant framework integrates physical symmetries to enhance the accuracy. In a recent extension, we combine Dingo results with existing GW simulators via importance sampling. This provides a corrected result free from potential network inaccuracies and a reliable performance metric, thereby addressing common criticisms against the use of deep learning methods in scientific applications. In my talk, I will highlight selected aspects from this line of work. References: https://arxiv.org/abs/2106.12594 (PRL 2021), https://arxiv.org/abs/2111.13139 (ICLR 2022), https://arxiv.org/abs/2210.05686 (preprint 2022)