Every black hole merger leaves its own time-frequency pattern on detectors like LIGO—a subtle ripple climbing in frequency and trembling under the influence of complex physics. But sometimes two completely different phenomena paint suspiciously similar pictures. When a gravitational waves from a black hole merger passes a massive object, the effect of microlensing kicks in—the signal's frequency and amplitude begin to oscillate, as if an interference pattern is overlaid on the detector. According to general relativity by Albert Einstein, massive bodies warp spacetime, acting as lenses—this is gravitational lensing. On the flip side, if one of the black holes is spinning and its axis isn't aligned with the orbital angular momentum, precession occurs—the orbital plane starts to rotate, modulating the signal in a different way. Telling these scenarios apart quickly and reliably is critical, because a wrong interpretation would skew our understanding of dark matter, spin distributions, and even challenge the very geometry of the universe.
The authors created a library of 60 000 synthetic signals of three types: microlensed without spin (ML), unlensed with precession (UP), and plain non-spinning (UN). Each signal was run through realistic noise models—from idealized Gaussian to real recordings from the H1 detector (LIGO Hanford) during the O4a run. The key step is transforming the signal into a Q-transform (QT), which turns a time series into an image with time on the horizontal axis, frequency on the vertical, and brightness representing energy. These 'fingerprints' were fed into a Convolutional Neural Networks with six convolutional blocks, trained using Machine Learning techniques. Simulations used the GWMAT, Bilby, and gwpopulation packages, with binary and lens parameters drawn from astrophysical distributions, including constraints from the GWTC-3 catalog. Special care was taken to select signals with a signal-to-noise ratio above 20 and a microlensing delay less than 0,15 seconds—so the effects would be distinguishable.
Classification between ML and UP on Q-transform images reached 95% accuracy in numerical simulation and 82% in real O4a noise. For ML versus UN, accuracy was around 80% in real noise, while UN versus UP could not be confidently separated even under ideal conditions—the precession pattern often gets lost in instrumental noise. Error analysis showed that confusion arises in two regimes: microlensed signals with a large impact parameter and low lens mass masquerade as unlensed ones because the distortions become barely noticeable; conversely, unlensed systems with high total mass and extreme mass ratio produce a short 'chirp' resembling the oscillations induced by a lens. Testing on 124 real events from the O4a catalog confirmed viability: the model trained on Gaussian noise correctly labeled most of them as UN, whereas the version trained on real noise showed poorer generalization.
For the first time, a signal processing low-latency system for microlensing identification has been proposed and implemented directly in the gravitational-wave data stream. This opens the door to a large-scale search for compact dark objects, which could be primordial black holes—candidates for dark matter. Moreover, timely recognition of lensed events will prevent systematic errors in determining binary parameters, including distance and spin, which are critical for testing cosmological models and alternative gravity theories.
The next generation of classifiers will incorporate more complex physics: eccentric orbits, higher multipole radiation modes, and hybrid scenarios where both microlensing and precession are present. Improving neural network architectures and moving to ensemble methods will boost noise robustness. Expanding the training set with realistic O4 data and future O5 data, along with multi-detector information, will make the model suitable for routine use within the LIGO–Virgo–Kagra collaboration.
Gravitational-wave astronomy, cosmology (dark matter searches via microlensing), compact object astrophysics, fast data processing algorithms.
Carry out a large-scale background study on thousands of events to calibrate the classification threshold, incorporate data from the L1 and Virgo detectors into the model, add simulations of precessing signals with eccentricity.
The work is directly tied to the mystery of dark matter's nature (microlenses as primordial black hole candidates), the question of black hole spin distributions in binaries, and the general challenge of disentangling degenerate parameters in gravitational-wave observations—essential for testing relativity in strong fields.
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The authors generated 60,000 synthetic signals of three types, converted them into Q-transforms, and trained a convolutional neural network to distinguish microlensed from precessing signals, achieving 95% accuracy in Gaussian noise and 82% in real O4a noise.
Where its strength is
This work introduces the first low-latency classification system capable of separating microlensing effects from spin precession in real time, paving the way for rapid dark matter searches and studies of compact object populations.
Where it could go next
Worth trying
Why this may matter
Fast and reliable separation of microlensing and precession is critical for correctly interpreting the nature of dark matter and black hole spins; the proposed extensions could turn this method into a standard tool for online LIGO–Virgo–Kagra analysis, influencing alert strategies and follow-up observations.
Nearby works in our archive
The closest work in the archive sits at 0.69 by meaning — the usual level for our corpus.
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.