An AI system can turn the gravitational-wave signal from merging neutron stars into a detailed description of the event in about a second. The March 2025 DINGO-BNS study demonstrated a substantial acceleration of analysis that can otherwise occupy much longer, giving astronomers a useful route towards faster follow-up observations.[2]
The value of that second comes from what happens outside the computer. A collision can produce a brief flash and a changing glow across different wavelengths. Getting a telescope to the right patch of sky sooner can reveal parts of the event that would otherwise pass unobserved.[3]
One collision showed what combined observations can reveal
On 17 August 2017, the merger known as GW170817 produced gravitational waves and a gamma-ray burst. NASA’s Fermi telescope detected the high-energy light; other space and ground observatories subsequently studied the expanding debris. The source was in the galaxy NGC 4993, about 130 million light-years away.[3]
Those observations changed with time. Swift saw a bright ultraviolet source about 15 hours after the initial signals, but the ultraviolet emission had faded beyond detection when it looked again on 29 August. The evolving light supplied information that the gravitational waves alone could not provide.[3]
That creates a practical scientific question: how quickly can the first signal tell astronomers where to look, how far away the source is and what kind of event deserves their attention? A faster answer can improve the use of telescopes that already exist.
What the AI demonstration established
DINGO-BNS results and their evidence
| Task | Result | Evidence |
|---|---|---|
| Characterising a merger | About one second for the reported parameter inference | Reanalysis of GW170817 and GW190425, alongside simulated eventsSource 1 |
| Locating the source | Around 30% smaller median sky region | Simulated-event comparison with BAYESTAR at 90% credible levelSource 1 |
| Working before the collision | Parameter estimates from an incomplete incoming signal | Retrospective and simulation demonstrationsSource 1 |
The output covers 17 parameters, including the stars’ masses, spins, distance, direction and response to tidal forces. It supplies probability distributions: a range of possibilities with different levels of support. The one-second result therefore represents considerably more information than a single point placed on a sky map.[1]
Learning the answer space before an event arrives
The wider Dingo framework learns from simulated signals whose physical parameters are known. Training a neural network requires work upfront, but that investment can serve many later observations. When new data arrive, the model proposes combinations of parameters that could have produced them.[4]
Its published software includes a second stage that checks and corrects those proposals through importance sampling. In plain terms, it weighs the model’s suggestions using the underlying physical likelihood calculation. The neural network concentrates effort on plausible answers while the calculation provides a route to refining their probabilities.[4]
Neutron-star signals required additional engineering because their long duration makes them difficult to process quickly. The researchers combined their model with knowledge of the expected signal to compress the data adaptively. This was a collaboration between physical understanding and machine learning, with the compression suited to the event being analysed.[1]
That division of work explains why speed need not come simply from ignoring inconvenient information. The model can do preparatory learning before the event, and the subsequent calculation can focus on the part of the problem that the observed signal makes relevant.
From one-second analysis to more useful observations
The same ability to handle partial signals makes repeated updates possible as two stars approach merger. Astronomers could receive an improving description while there is still time to organise observations. The study demonstrated the analysis capability; catching a previously unseen collision’s earliest light with it remains a prospective application.[2]
Development of the broader approach continued in 2026. A Physical Review Letters study introduced Dingo-T1, a transformer-based model tested on 48 binary-black-hole events from the third LIGO–Virgo–KAGRA observing run. It handled different detector combinations and frequency ranges within one flexible model, making it easier to examine how analysis choices affect a result.[5]
That later work addresses a related need: useful scientific software must adapt to the data available on a particular occasion. Together, the studies show progress towards rapid interpretation that can accommodate real observational choices. The opportunity is to spend less of a fleeting astronomical event waiting for its measurements to become intelligible.
Sources & further reading
- Real-time inference for binary neutron star mergers using machine learning Nature · 5 Mar 2025 · Accessed 11 Oct 2026
- Neural network deciphers gravitational waves from merging neutron stars in a second Max Planck Institute for Gravitational Physics · 5 Mar 2025 · Accessed 11 Oct 2026
- NASA Missions Catch First Light from a Gravitational-Wave Event NASA · 16 Oct 2017 · Accessed 11 Oct 2026
- Dingo: Deep inference for gravitational-wave observations Dingo researchers / GitHub · Accessed 11 Oct 2026
- Flexible Gravitational-Wave Parameter Estimation with Transformers Physical Review Letters · 26 Jun 2026 · Accessed 11 Oct 2026