Original title: Self-supervised reconstruction of transients in data from space-borne gravitational-wave detectors
Yuxiang Xu, Minghui Du, Bo Liang, Zihao Xiao +2 · September 2026
Future space observatories will hunt for gravitational waves, which are ripples in space. Sometimes these detectors catch short, surprise signals with shapes nobody can predict. Usually, artificial intelligence needs clean examples to learn how to spot specific signal shapes. That fails when researchers do not know what shape to expect. To fix this, researchers built a self-supervised system inspired by Noise2Noise. It learns directly from noisy data without needing clean examples or pre-made templates. The team trained one model on noisy black hole data. Surprisingly, it still recovered cosmic string bursts and detector glitches. It also worked well even when signals overlapped. Finally, when a glitch was found, scientists subtracted its shape from continuous test data, clearing out unwanted noise power without an anomaly template.
Imagine trying to learn a friend's voice by listening to a crowded party. You have never heard the friend alone in a quiet room, so you do not know exactly what their voice sounds like. But you hear them in many different noisy moments, and after a while you can pick out that voice even in a fresh crowd. In this paper, the system learns to recognize a gravitational-wave event the same way: it is trained only on noisy observations, with no clean copy to compare against, so it can pull out a brief event even when the exact shape was never seen before.
Space-based gravitational-wave detectors may hear signals we have no template for. This method lets those signals be pulled out and studied anyway, and once an instrumental glitch is identified, its waveform can be subtracted from the data without a template. That means cleaner data and a way to characterize brand-new sources.
The abstract demonstrates recovery in source-confused test data and a simulated continuous data stream, not real detector data. It shows high-overlap recovery for MBHB signals and dominant morphologies of glitches and cosmic string bursts, but does not claim all transient types. The subtraction step requires independent information identifying a candidate as instrumental, so it does not apply to unclassified transients.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Self-supervised reconstruction of transients in data from space-borne gravitational-wave detectors — Yuxiang Xu, Minghui Du, Bo Liang et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.