Anomaly Detection for Gravitational Waves

Why This Matters

LIGO and Virgo are the most sensitive instruments humans have built. They are also fragile: ground motion, scattered light, photodiode saturation, and hundreds of subtler couplings produce non-Gaussian transients ("glitches") that mimic astrophysical signals. During Observing Run O3, glitch rates exceeded $1$ per minute per detector, with morphologies that overlap the parameter space of compact binary inspirals.

Matched filtering against template banks is the canonical detection pipeline for known waveform families (binary black holes, neutron stars). It is near-optimal under stationary Gaussian noise but degrades sharply when glitches violate the noise model. Glitch identification, classification, and rejection are now core parts of the calibration and detection chain. ML moved from auxiliary tooling to a load-bearing component between O2 and O4.

For unmodeled signals (supernova core collapse, cosmic-string cusps, or genuinely unknown astrophysics), there is no template. Detection becomes an anomaly-detection problem against an empirical noise distribution that drifts on hours-to-days timescales.

Core Ideas

Gravity Spy and citizen-science labels. The Gravity Spy project (Bahaadini et al. 2018, Information Sciences 444) couples a CNN trained on spectrogram images of LIGO glitches with crowdsourced labels from Zooniverse volunteers. The system labels 22 glitch classes (Blip, Koi Fish, Whistle, Scattered Light, etc.) with reported $> 97\%$ accuracy on held-out examples. Active learning routes uncertain examples to volunteers; high-confidence labels feed back into the training set. The labeled corpus has become the de facto benchmark for LIGO glitch ML.

CNN-based signal vs. noise discrimination. George and Huerta (2018, PRD 97; arXiv 1701.00008) showed that deep CNNs operating directly on time-series strain data can detect simulated binary black hole signals at sensitivities comparable to matched filtering, with three to four orders of magnitude lower latency. Subsequent work (Gabbard et al. 2018, PRL 120) confirmed that CNN detection statistics approach the Neyman-Pearson optimum on simulated Gaussian noise. The practical wins are speed and the ability to absorb non-Gaussian features that templates ignore.

Unsupervised methods for unmodeled bursts. Coherent WaveBurst is the classical excess-power pipeline. ML alternatives include autoencoders trained on detector noise that flag high-reconstruction-error segments as candidates, and variational methods that estimate detector-specific noise manifolds. The detection threshold is set by tail behavior of the reconstruction-error distribution; calibration against time-shifted background is mandatory.

Parameter estimation acceleration. Bayesian parameter estimation for a single binary merger requires $\sim 10^6$ likelihood evaluations and historically took hours to days. Normalizing-flow surrogates (Dax et al. 2021, PRL 127; arXiv 2106.12594) produce posterior samples in seconds with quality comparable to nested sampling, enabling real-time multimessenger alerts.

Common Confusions

Watch Out

High classifier accuracy is not low false-alarm rate

A glitch classifier with 99% accuracy on a balanced test set can still produce $10^4$ false alarms per day at the rates seen in raw LIGO data. Operating points must be set against the actual class prior and trigger rate, not balanced-set accuracy. The relevant metric is the false-alarm rate at fixed detection efficiency, evaluated on time-slid background.

References

Related Topics

• Convolutional Neural Networks
• Signal Detection Theory
• Anomaly Detection
• Out-of-Distribution Detection
• Detection Theory