Detectors of gravitational waves usually compare the signal with ready-made templates, as if searching for a familiar melody in the noise. This allows them to track only known scenarios, such as black hole mergers. But explosions of supernovae or sudden tremors of neutron stars go unnoticed – there are no templates for them.
The CoAD method bypasses this limitation. Two neural networks work independently on two distant detectors like a pair of sensitive ears: each hears its own noise, but a signal is considered real only if both ears catch the same sound simultaneously. Training happens without hints – the networks learn on their own to pick out coincidences from the tangled data.
This approach not only finds unusual waves but also reconstructs their shape – like putting together a composite sketch of an unseen phenomenon. In tests, CoAD recognized over 90% of neutron star mergers and up to 85% of low-frequency bursts, making mistakes less than once a year. For future ultra-sensitive detectors, it will become indispensable.
The most sensitive gravitational wave detectors can capture vibrations a thousand times smaller than the diameter of a proton.