A neural network already used to find black hole mergers has now proven it can hunt binary neutron star collisions just as accurately as the conventional method — at a fraction of the computing cost.
Researchers adapted Aframe, an AI search algorithm deployed during LIGO-Virgo-KAGRA's fourth observing run, to detect signals from merging neutron stars. The original pipeline for that task requires matching incoming detector data against roughly one million reference waveforms, a job that can demand up to a thousand CPU cores running in real time. Aframe sidesteps that by training a neural network to recognize the presence of a signal directly in the data. To handle the longer signal duration that neutron star mergers produce compared to black hole pairs, the team applied a preprocessing step called heterodyning before feeding data to the network. The result: sensitivity on par with matched-filter pipelines, on a single non-flagship GPU.
The stakes for binary neutron star detection are unusually high. When two neutron stars merge, they emit not just gravitational waves but light and neutrinos simultaneously — the kind of multi-messenger event that lets physicists probe nuclear matter, test general relativity, and measure the expansion rate of the universe all at once. The 2017 detection of GW170817 remains the only confirmed example, and the entire field has been waiting for a second one. Cutting latency and compute costs makes real-time alerts more feasible, which is when multi-messenger follow-up is most valuable.
Aframe was already the first AI pipeline to detect multiple binary black hole mergers live; this extension to the lower-mass neutron star regime is a meaningful step, though a detected event — not a benchmark — will be the real proof of concept.