Neural network catches up with classical methods in the search for neutron star mergers

Neutron stars are the super-dense remnants of exploded supernovae, predicted by Fritz Zwicky. Their rapidly spinning varieties — pulsars — were discovered by Jocelyn Bell Burnell. When two such stars spiral together and merge, they create gravitational waves — ripples in the fabric of spacetime. These events not only shake the cosmos but also spew out heavy elements like gold and platinum, and produce flashes of light that telescopes can catch. The challenge is to spot the faint signal amid detector noise — like picking out a familiar tune in a noisy crowd.

Previously, the search was done "head-on": scientists pre-calculated thousands of possible signal shapes and continuously compared them with sensor readings — akin to riffling through a deck of cards looking for a specific one. It’s reliable, but slow. The Aframe neural network does things differently. The long "chirping" signal is first compressed using heterodyning — a trick similar to tuning an old radio: you strip away the carrier frequency and leave just the pure, slow sound. That leaves a snippet only a second and a half long, and the neural network instantly says whether it contains a merger.

Tests on data from the LIGO gravitational-wave observatories, built with the involvement of Rainer Weiss, showed that Aframe finds neutron star mergers just as well, and for heavier systems even better than classical algorithms. An alert about the event reaches astronomers in less than five seconds. That gives a real chance to slew telescopes and see the light from a cosmic cataclysm, while also peering into the hearts of stars where matter is squeezed to unimaginable densities and follows the Standard Model of physics at its limits.

In the future, such neural networks could search for any mergers — with black holes, and mixed pairs. This will make multi-messenger astronomy truly real-time.