Machine learning method could revolutionise multi-messenger astronomy
Observing binary neutron star mergers is high on the wish list of astronomers. In a study published in the scientific journal Nature, an interdisciplinary team of researchers including ETH Zurich postdoc Maximilian Dax and Professor Bernhard Schölkopf presents a novel machine learning method to analyse gravitational waves emitted from neutron star collisions almost instantaneously – even before the merger is fully observed.
Binary neutron star mergers occur millions of light-years away from Earth. Interpreting the gravitational waves they produce presents a major challenge for traditional data-analysis methods. These signals correspond to minutes of data from current detectors and potentially hours to days of data from future observatories. Analysing such massive data sets is computationally expensive and time-consuming.
An international team of scientists has developed a machine learning algorithm, called DINGO-BNS (Deep INference for Gravitational-wave Observations from Binary Neutron Stars) that saves valuable time in interpreting gravitational waves emitted by binary neutron star mergers. They trained a neural network to fully characterize systems of merging neutron stars in about a second, compared to about an hour for the fastest traditional methods.
Why is real-time computation important?
Neutron star mergers emit visible light (in the subsequent kilonova explosion) and other electromagnetic radiation in addition to gravitational waves. “Rapid and accurate analysis of the gravitational-wave data is crucial to localize the source and point telescopes in the right direction as quickly as possible to observe all the accompanying signals,” says Maximilian Dax, first author of the paper, former doctoral student in the Empirical Inference Department at the Max Planck Institute for Intelligent Systems (MPI-IS) and now postdoc at ETH Zurich and at the ELLIS Institute Tübingen.
The real-time method could set a new standard for data analysis of neutron star mergers, giving the broader astronomy community more time to point their telescopes toward the merging neutron stars as soon as the large detectors of the LIGO-Virgo-KAGRA (LVK) collaboration identify them. “Current rapid analysis algorithms used by the LVK make approximations that sacrifice accuracy. Our new study addresses these shortcomings,” says Jonathan Gair, a group leader in the Astrophysical and Cosmological Relativity Department at the Max Planck Institute for Gravitational Physics in the Potsdam Science Park.
“Rapid and accurate analysis of the gravitational-wave data is crucial to localize the source and point telescopes in the right direction as quickly as possible to observe all the accompanying signals. ”Maximilian Dax, postdoc at ETH Zurich and first author of the paper
Indeed, the machine learning framework fully characterizes the neutron star merger (e.g., its masses, spins, and location) in just one second without making such approximations. This allows, among other things, to quickly determine the sky position 30% more precisely. Because it works so quickly and accurately, the neural network can provide critical information for joint observations of gravitational-wave detectors and other telescopes. It can help to search for the light and other electromagnetic signals produced by the merger and to make the best possible use of the expensive telescope observing time.
Catching a neutron star merger in the act
“Gravitational wave analysis is particularly challenging for binary neutron stars, so for DINGO-BNS, we had to develop various technical innovations. This includes for example a method for event-adaptive data compression,” says Stephen Green, UKRI Future Leaders Fellow at the University of Nottingham. Bernhard Schölkopf, Professor of Computer Science at ETH Zurich, Director of the Empirical Inference Department at MPI-IS and at the ELLIS Institute Tübingen, adds: “Our study showcases the effectiveness of combining modern machine learning methods with physical domain knowledge.”
DINGO-BNS could one day help to observe electromagnetic signals before and at the time of the collision of the two neutron stars. “Such early multi-messenger observations could provide new insights into the merger process and the subsequent kilonova, which are still mysterious,” says Alessandra Buonanno, Director of the Astrophysical and Cosmological Relativity Department at the Max Planck Institute for Gravitational Physics.
The results of the paper ‘Real-time inference for binary neutron star mergers using machine learning’ were published in Nature on 5 March 2025. external page Read article.
Reference
Maximilian Dax (1,2,3), Stephen R. Green (4), Jonathan Gair (5), Nihar Gupte (5,6), Michael Pürrer (7,8), Vivien Raymond (9), Jonas Wildberger (3), Jakob H. Macke (1,10), Alessandra Buonanno (5,6) and Bernhard Schölkopf (1,2,3)
1 Max Planck Institute for Intelligent Systems, Tübingen, Germany. 2 ETH Zurich, Zurich, Switzerland. 3 ELLIS Institute Tübingen, Tübingen, Germany. 4 School of Mathematical Sciences, University, of Nottingham, Nottingham, UK. 5 Max Planck Institute for Gravitational Physics (Albert Einstein Institute), Potsdam, Germany. 6 Department of Physics, University of Maryland, College Park, MD, USA. 7 Department of Physics, University of Rhode Island, Kingston, RI, USA. 8 Center for Computational Research, University of Rhode Island, Kingston, RI, USA. 9 Gravity Exploration Institute, Cardiff University, Cardiff, UK. 10 Machine Learning in Science, University of Tübingen & Tübingen AI Center, Tübingen, Germany.
DOI 10.1038/s41586-025-08593-z