A noise kaleidoscope — that’s one way to describe the data streaming from the LIGO gravitational waves detectors, built by Rainer Weiss, Kip Thorne, and Barry Barish. These instruments catch ripples in spacetime from distant cosmic cataclysms, but they also pick up everything else: ground tremors, electrical noise, even vibrations from passing traffic. Each blip is a short disturbance, or glitch, that can closely mimic a real signal. As detector sensitivity improved, the kaleidoscope spun faster: glitches multiplied and became more varied.

To tame this chaos, scientists turned to Machine Learning — specifically, the DANTE V3 algorithm. First, it performs Spectral Analysis: decomposing each signal across four time scales and building time-frequency portraits. Then a neural networks (a program that learns from examples), built as a transformers (a specialized pattern recognizer) and trained via self-supervised learning (without explicit labels), slices up the portrait and picks out the weirdest bits. At the same time, Digital Image Processing and signal processing techniques help distinguish the meaningful from the background. But the real secret is guarding against noise drift. Detectors change their behavior over time: what was normal yesterday might look odd today. To avoid mistaking slow drifts for anomalies, DANTE compares each suspicious signal with a library of 1216 typical noise patterns, collected during quiet periods. On top of that, it checks whether the same blip shows up in seismic sensors or magnetometers (environmental monitoring channels). If it does, the signal is rejected as earthly noise.

After this cleanup, out of 10 372 candidates, only 28 percent remained. They formed a fuzzy cloud with no clear clusters — just random noise. Only two events behaved differently. The first was firmly vetoed: its signal matched the control system output almost perfectly (a coefficient of 0.987). The second survived: none of the seven environmental sensors saw anything similar, and the twin detector registered no response. But the signal’s shape doesn’t match any known gravitational-wave template, so it’s considered an unknown glitch of the instrument itself. The upper limit on the rate of such unseen disturbances is about 6 per year for each detector.

The DANTE V3 story makes it clear: if you don’t ground your analysis in the real noise background, computer algorithms will find ghosts where none exist. The new protocol, calibrated on real data rather than theoretical models, will serve as a reliable filter for future observing runs. Next up is a version that can filter data in real time, turning the noise kaleidoscope into a map where the true ripples of spacetime become ever sharper.