In 2019, the LIGO facility recorded a gravitational waves, a ripple in the fabric of spacetime, whose time-frequency map resembled a strange comb. Each such signal from a black hole merger leaves a unique pattern, like a fingerprint. But the problem is that two completely different phenomena can produce nearly indistinguishable pictures.
If a gravitational wave passes by a massive object, it experiences microlensing — the focusing of gravitational waves by a compact object. The signal's frequency and amplitude start oscillating, as if the detector picks up interference from multiple copies of the same event. On the other hand, when one of the black holes spins and its axis is not aligned with the orbital angular momentum, the orbital plane starts to rotate: this is precession. It also modulates the signal, but in a completely different way. Separating these scenarios quickly and reliably is like distinguishing two similar fingerprints left at different times and with different pressure.
The researchers created a database of 60,000 synthetic fingerprints — numerical simulations of signals of three types: microlensed non-spinning, non-lensed precessing, and ordinary non-spinning. Each signal was passed through realistic noise models — from idealized Gaussian to actual recordings from the LIGO Hanford detector during observing run O4a. Then they were converted into Q-transforms — time-frequency images where brightness corresponds to energy. These two-dimensional fingerprints were fed into a Convolutional Neural Networks with six blocks, trained using Machine Learning methods.
The result surprised: in ideal Gaussian noise, the accuracy of distinguishing microlensed and precessing signals reached 95%. In real O4a noise, it dropped to 82%—but that is still a breakthrough for signal processing low-latency data processing. Errors occurred in two cases: when the lensing distortion was barely noticeable (large impact parameter and small lens mass), and when a short 'chirp' from a massive binary with extreme mass ratio resembled lensing oscillations. However, a test on 124 real events from the O4a catalog confirmed: the model trained on Gaussian noise reliably identified ordinary signals, not confusing them with exotic ones.
This is the first step toward a mass search for compact dark objects — candidates for primordial black holes that could constitute dark matter. Rapid recognition of lensed events will prevent systematic errors in distance and spin measurements, critical for testing cosmological models and general relativity itself. In the future, neural network architectures will become more complex, incorporating eccentric orbits, higher-order emission modes, and hybrid scenarios with simultaneous lensing and precession. For now, each new fingerprint captured from the fabric of spacetime brings us closer to unraveling the dark side of the Universe.
A microlens with a mass around a thousand solar masses can alter a merger signal so much that even an automated classifier would notice: the amplitude oscillates at tens of hertz, as if the detector is picking up 'beats' from overlapping ray paths.