Gravitational-wave astronomy turns to AI

16 March 2022
Computer simulation of gravitational waves emitted by a supernova. Credit: J Powell / B Mueller

New frontiers in gravitational-wave (GW) astronomy were discussed in the charming and culturally vibrant region of Oaxaca, Mexico from 14 to 19 November. Around 37 participants attended the hybrid Banff International Research Station for Mathematical Innovation and Discovery (BIRS) “Detection and Analysis of Gravitational Waves in the Era of Multi-Messenger Astronomy: From Mathematical Modelling to Machine Learning’’ workshop. Topics ranged from numerical relativity to observational astrophysics and computer science, including the latest applications of machine-learning algorithms for the analysis of GW data.

GW observations are a new way to explore the universe’s deepest mysteries. They allow researchers to test gravity in extreme conditions, to get important clues on the mathematical structure and possible extension of general relativity, and to understand the origin of matter and the evolution of the universe. As more GW observations with increased detector sensitivities spur astrophysical and theoretical investigations, the analysis and interpretation of GW data faces new challenges which require close collaboration with all GW researchers. The Oaxaca workshop focused on a topic that is currently receiving a lot of attention: the development of efficient machine-learning (ML) methods and numerical-analysis algorithms for the detection and analysis of GWs. The programme gave participants an overview of new-physics phenomena that could be probed by current or next-generation GW detectors, as well as data-analysis tools that are being developed to search for astrophysical signals in noisy data.

Since their first detections in 2015, the LIGO and Virgo detectors have reached an unprecedented GW sensitivity. They have observed signals from binary black-hole mergers and a handful of signals from binary neutron star and mixed black hole-neutron star systems. In discussing the role that numerical relativity plays in unveiling the GW sky, Pablo Laguna and Deirdre Shoemaker (U. Texas) showed how it can help in understanding the physical signatures of GW events, for example by distinguishing black hole-neutron star binaries from binary black-hole mergers. On the observational side, several talks focused on possible signatures of new physics in future detections. Adam Coogan (U. de Montréal and Mila) and Gianfranco Bertone (U. of Amsterdam, and chair of EuCAPT) discussed dark-matter halos around black holes. Distinctive GW signals  could help to determine whether dark matter is made of a cold, collisionless particle via signatures of intermediate mass-ratio inspirals embedded in dark-matter halos. In addition, primordial black holes could be dark-matter candidates.

Bernard Mueller (U. Monash) and Pablo Cerdá-Durán (U. de Valencia) described GW emission from core-collapse supernovae. The range of current detectors is limited to the Milky Way, where the rate of supernovae is about one per century. However, if and when a galactic supernova happens, its GW signature will be within reach of existing detectors. Lorena Magaña Zertuche (U. of Mississippi) talked about the physics of black-hole ringdown – the process whereby gravitational waves are emitted in the aftermath of a binary black-hole merger – which is crucial for understanding astrophysical black holes and testing general relativity. Finally, Leïla Haegel (U. de Paris) described how the detection of GW dispersion would indicate the breaking of Lorentz symmetry: if a GW propagates according to a modified dispersion relation, its frequency modes will propagate at different speeds, changing  the phase evolution of the signals with respect to general relativity.

Machine learning
Applications of different flavours of ML algorithms to GW astronomy, ranging from the detection of GWs to their characterisation in detector simulations, were the focus of the rest of the workshop.

ML has seen a huge development in recent years and has been increasingly used in many fields of science. In GW astronomy, a variety of supervised, unsupervised, and reinforcement ML algorithms, such as deep learning, neural networks, genetic programming and support vector machines, have been developed. They have been used to successfully deal with noise in the detector, signal processing, data analysis for signal detections and for reducing the non-astrophysical background of GW searches. These algorithms must be able to deal with large data sets and demand  a high accuracy to model  theoretical waveforms and to perform  searches at the limit of instrument sensitivities. The next step for a successful use of ML in GW science will be the integration of ML techniques with more traditional numerical-analysis methods that have been developed for the modelling, real-time detection, and analysis of signals.

The BIRS workshop provided a broad overview of the latest advances in this field, as well as open questions that need to be solved to apply robust ML techniques to a wide range of problems. These include reliable background estimation, modelling gravitational waveforms in regions of the parameter space not covered by full numerical relativity simulations, and determining populations of GW sources and their properties. Although ML for GW astronomy is in its infancy, there is no doubt that it will play an increasingly important role in the detection and characterization of GWs leading to new discoveries.

bright-rec iop pub iop-science physcis connect