Digital Conductor of Gravitational Waves: Neural Network Conducts Black Hole Mergers ✛
In September 2015, two four-kilometer laser beams compressed by a thousandth of a proton — the LIGO interferometry detectors caught a gravitational waves for the first time. Its source turned out to be the merger of two black holes 1,3 billion light-years away. Since then, hundreds of such signals have been captured, and each is a cipher: from the wave shape, astrophysicists reconstruct the masses, spins, and trajectories of the colliding giants.
But decoding requires comparison with millions of templates. Templates are obtained by solving Einstein field equations on supercomputers — such a calculation can take a week even for a single scenario. Therefore, surrogates were created — fast approximate models. However, even they begin to slow down when next-generation detectors start registering an event every few minutes.
The authors of the new study turned the problem into an orchestral score. Instead of one cumbersome calculation, they broke the wave into 25 'instruments': separately predicting the orbital frequency, the spin of the entire system, the spins of each black hole, and 21 overtones — the gravitational modes. Each element is handled by its own tiny Machine Learning network. It's a digital conductor: it assigns parts and then assembles them into a single sound — the final gravitational wave.
The wave assembled in this way is practically indistinguishable from the benchmark numerical simulation. On 10,000 test systems, the discrepancy was only 8×10^(-5) — a hundred times less than the intrinsic error of the original surrogate. On a graphics processor, one template is generated in 1 millisecond. When batch-processing 64 signals, the acceleration reaches 140 times.
The main gain is not only speed, but also the differentiability of the entire chain. This allows applying gradient methods for parameter estimation, which were previously unavailable for gravitational-wave analysis. A typical event like GW150914 is now processed in four hours instead of days, and the result matches the benchmark within the margin of error.
It's already clear that the digital conductor handles the score no worse than computational monsters, but does so with ease. When detectors again register the trembling of spacetime — as in September 2015 — decoding will take not days, but minutes. And then astronomy will begin to hear the voice of the universe in real time.
One simulation of a black hole merger using numerical relativity can still take up to a month on a supercomputer. The new surrogate generates just as many waves in... one second.
In the movie Interstellar, real calculations from general relativity were used to visualize the black hole. Today's neural network models, like the one described, help not to draw pictures, but to listen to the real 'sounds' of the universe hidden in detector data.