Speaker
Description
Gravitational-wave detectors like LIGO, Virgo, and KAGRA are extremely sensitive instruments that pick up faint ripples in spacetime from distant cosmic events. However, that same sensitivity means they also pick up "glitches": brief bursts of instrumental or environmental noise that can mimic genuine astrophysical signals or mask their true characteristics.
We present DeepExtractor, a deep learning method that separates a signal or glitch from the surrounding detector noise regardless of the source or morphology. Rather than learning to recognize specific signal patterns, the model learns what the background noise itself typically looks like, predicts it, and subtracts it away. Whatever's left over is the reconstructed signal or glitch. This makes it broadly applicable, since it isn't tied to any particular source or morphology.
We validated the approach on simulated glitches, on real noise artifacts from LIGO data, and by reconstructing real gravitational-wave events, despite the model never having been trained on real astrophysical signals. It matches or exceeds the accuracy of BayesWave, a leading but computationally expensive existing method, while running significantly faster.
We then extend the method to where a signal and a glitch overlap in time and frequency. By jointly separating both from the noise at once, we show this reduces the bias a glitch would otherwise introduce into the physical properties later inferred from the recovered signal. Finally, we discuss the method's potential for fast, real-time detection pipelines, as a template-free alternative to existing search methods.