Abstract:
Gravitational-wave (GW) astronomy has led to groundbreaking discoveries in the past decade, and with the development of next-generation detectors, its potential for future breakthroughs continues to grow. This field hinges on the ability to accurately characterize GW sources based on measured data. However, computational demands of existing inference methods impede their application to large-scale or real-time data analysis. We here present DINGO, a probabilistic machine learning framework for Bayesian GW inference that addresses these limitations with an unprecedented combination of speed and accuracy. Building on neural posterior estimation (NPE), DINGO trains deep neural networks on GW simulations to learn the mapping between measured data and GW source parameters.
We first introduce DINGO for binary black hole mergers, the most common GW source. We develop techniques to integrate symmetries (called GNPE) and to rapidly adapt to varying detector noise properties. We then augment NPE with importance sampling (NPE-IS) to correct for potential network inaccuracies. This enables asymptotically exact inference, independent verification and unbiased estimates of the Bayesian evidence, addressing important limitations of deep learning-based inference. Finally, we extend DINGO to binary neutron star mergers. We develop techniques to effectively compress long signals based on event-adaptive priors (prior conditioning) and to enable inference even before the merger. With inference times of less than a second, this provides crucial real-time information for directing searches for electromagnetic counterparts.
Our experimental evaluations encompass more than 50 real events and thousands of simulations, three different waveform models, two types of sources and two experimental setups (LIGO-Virgo-KAGRA and next-generation detectors). DINGO consistently achieves comparable accuracy to established inference methods while being orders of magnitude faster. This prepares GW data analysis for increasing detection rates, facilitates large-scale studies and can improve searches for electromagnetic counterparts. Beyond GW astronomy, DINGO contributes several broadly applicable techniques to the field of simulation-based inference, including GNPE, NPE-IS and prior-conditioning.
Machine Learning