When two black holes merge, a tremor ripples through the universe — a gravitational wave. Detectors on Earth catch this tremor, but to understand which black holes collided, the signal must be compared with thousands of templates. It's like searching for a specific melody in a huge music library by playing every record. Each year, more signals arrive, and old methods can't keep up.

The neural network from this work acts like a simultaneous interpreter: first it quickly conveys the gist, then it refines the details. The first part of the network compresses data about the masses and spins of the black holes into a short code and draws the wave from it. The second part corrects errors, especially at the moment of merger — the trickiest part. As a result, the waveform comes out nearly right: the error is less than one hundredth.

Previously, it was unclear whether such fast models could be trusted. Now it's clear: they can, if you know where they go wrong. This will speed up processing of future observations and help find interesting events faster.