Template-free search for gravitational wave events using coincident anomaly detection

TL;DR

This work tackles the limitation of template-based gravitational-wave searches that miss unmodeled sources by introducing Coincident Anomaly Detection (CoAD), a fully unsupervised, template-free approach that trains two detector-specific neural networks to maximize prediction coincidence across spatially separated detectors. The method defines a loss $\ ilde{F}_\beta$ based on the correlation and covariance of the networks' outputs, enabling learning from coincident signals without labeled data or background-only datasets. Applied to Codabench LIGO data with BBH and SGLF injections, CoAD achieves high PR-AUC and recall at a low false-alarm rate, with interpretability via Integrated Gradients enabling localization in time–frequency space and potential data-driven template construction. The results suggest CoAD scales to higher event rates anticipated for next-generation detectors and could generalize to other multi-detector physics searches, offering a robust path toward discovering unmodeled GW sources.

Abstract

Gravitational-wave (GW) observatories have used template-based search to detect hundreds of compact binary coalescences (CBCs). However, template-based search cannot detect astrophysical sources that lack accurate waveform models, including core-collapse supernovae, neutron star glitches, and cosmic strings. Here, we present a novel approach for template-free search using coincident anomaly detection (CoAD). CoAD requires neither labeled training examples nor background-only training sets, instead exploiting the coincidence of events across spatially separated detectors as the training loss itself: two neural networks independently analyze data from each detector and are trained to maximize coincident predictions. Additionally, we show that integrated gradient analysis can localize GW signals from the neural-network weights, providing a path toward data-driven template construction of unmodeled sources, and further improving precision by frequency matching. Using the CodaBench dataset of real LIGO backgrounds with injected simulated CBCs and sine-Gaussian low-frequency bursts, CoAD achieves recall up to 0.91 and 0.85 respectively at a false-alarm rate of one event per year, and achieves recall above 0.5 at signal-to-noise ratios below 10. The fully-unsupervised nature of CoAD makes it especially well-suited for next-generation detectors with greater sensitivity and associated increases in GW event rates.

Figures (7)

• Figure 1: Schematic of the CoAD concept applied to template-free search at LIGO. H and L data are fed into separate NNs, which each make predictions of the probability of an anomaly (H pred. and L Pred. respectively). During training, the two predictions are combined in the $\hat{F}_\beta$ loss function, which updates the NNs via back-propagation. At inference time, the networks operate independently, and predictions --- if both are available --- can be combined to improve precision.
• Figure 2: Distribution of predictions for the 0.2%, 100 event model for a single model output (top) and the product of both models' outputs (bottom), highlighting how combining the two predictions improves precision. Both outputs are passed through a sigmoid function. Note that results presented use a logical AND operation between the two outputs, rather than the product of the sigmoids shown at bottom for visualization purposes.
• Figure 3: IG visualization for three input examples: a strong SGLF at left, a weak SGLF event in the center, and a background event at right. Note that nearly the entirety of the AUC curve has a threshold $\ll 0.1$, so the middle prediction represents a less confident --- but correct --- prediction that the example has a SGLF event.
• Figure 4: Recall as a function of SNR for the 100 event SGLF and BBH models. Blue crosses show recall at FAR-1/year, assuming random backgrounds. Performance drops assuming a worst-case trigger selecting difficult backgrounds, but recovers with either a higher FAR rate (magenta circle) or inclusion of an IG filter (black triangles).
• Figure 5: Validation loss values as a function of batch for the 250 event, 0.2% anomaly rate, SGLF model. The training curve for the final model is shown in blue.