[Paper Review] Real-Time Detection of Unmodelled Gravitational-Wave Transients Using Convolutional Neural Networks

V. Skliris, M. R. Norman+1 more|arXiv (Cornell University)|Sep 30, 2020

Pulsars and Gravitational Waves Research9 citations

TL;DR

This paper presents MLy, a novel convolutional neural network pipeline that detects unmodeled gravitational-wave transients in real time by analyzing both detector strain data and inter-detector Pearson cross-correlations, achieving sensitivity comparable to standard unmodeled transient searches with ~1-second latency and orders-of-magnitude lower computational cost than traditional methods.

ABSTRACT

Convolutional Neural Networks (CNNs) have demonstrated potential for the real-time analysis of data from gravitational-wave detector networks for the specific case of signals from coalescing compact-object binaries such as black-hole binaries. Unfortunately, training these CNNs requires a precise model of the target signal; they are therefore not applicable to a wide class of potential gravitational-wave sources, such as core-collapse supernovae and long gamma-ray bursts, where unknown physics or computational limitations prevent the development of comprehensive signal models. We demonstrate for the first time a CNN with the ability to detect generic signals -- those without a precise model -- with sensitivity across a wide parameter space. Our CNN has a novel structure that uses not only the network strain data but also the Pearson cross-correlation between detectors to distinguish correlated gravitational-wave signals from uncorrelated noise transients. We demonstrate the efficacy of our CNN using data from the second LIGO-Virgo observing run, and show that it has sensitivity comparable to that of the "gold-standard" transient searches currently used by LIGO-Virgo, at extremely low (order of 1 second) latency and using only a fraction of the computing power required by existing searches, allowing our models the possibility of true real-time detection of gravitational-wave transients associated with gamma-ray bursts, core-collapse supernovae, and other relativistic astrophysical phenomena.

Motivation & Objective

• To develop a real-time gravitational-wave detection pipeline capable of identifying generic, unmodeled transients without relying on precise signal models.
• To overcome the limitation of existing CNN-based methods that require specific signal templates for training.
• To enable low-latency, computationally efficient detection of transient signals from unknown sources such as core-collapse supernovae and gamma-ray burst-associated bursts.
• To improve sensitivity to gravitational-wave bursts (GWBs) by leveraging detector coherence rather than signal morphology.
• To support multi-messenger astronomy by issuing rapid alerts compatible with electromagnetic follow-up windows.

Proposed method

• The pipeline uses a dual-branch CNN architecture: one branch processes whitened strain data from individual detectors, and the other processes Pearson cross-correlation timeseries between detectors across all physically allowed light-travel time delays.
• The model is trained on randomized, featureless signals to learn to detect coherent, correlated transients across detectors rather than specific signal waveforms.
• Training data includes simulated gravitational-wave bursts and noise glitches with identical morphologies to force the network to rely on inter-detector coherence for detection.
• The system estimates false alarm rates using ~10^3 time-shifted data segments, enabling robust threshold setting with minimal computational overhead.
• A single A100-SXM4 GPU can handle the full analysis in real time, drastically reducing the need for hundreds of CPUs used in conventional low-latency searches.
• The architecture separates detection of transient energy from detection of inter-detector coherence, allowing robust identification of correlated signals while rejecting coincident noise glitches.

Experimental results

Research questions

• RQ1 Can a deep learning model detect unmodeled gravitational-wave transients without being trained on specific signal waveforms?
• RQ2 Can inter-detector cross-correlation information improve detection sensitivity for generic transient signals?
• RQ3 Can such a model achieve real-time performance with minimal computational cost compared to standard unmodeled transient searches?
• RQ4 How does the performance of this model compare to established low-latency search methods like cWB in terms of sensitivity and false alarm rate?
• RQ5 Can the model reliably distinguish true gravitational-wave signals from coincident noise glitches using coherence alone?

Key findings

• MLy achieves sensitivity to gravitational-wave transients approaching that of the cWB search, a gold-standard unmodeled transient search, as shown in Table 1.
• The pipeline operates with a latency of approximately 1 second, enabling real-time alert generation for multi-messenger follow-up.
• MLy requires only a fraction of the computing power of traditional methods—approximately three to four A100-SXM4 GPUs can sustain real-time operation, compared to hundreds of CPUs in standard pipelines.
• The model successfully detects a variety of simulated GWB morphologies not present in the training data, demonstrating generalization to unknown waveforms.
• The background distribution of false alarms is dominated by known glitch types, suggesting that post-processing with glitch classifiers could further reduce false positives.
• The method's reliance on coherence rather than shape enables detection of signals from sources with unknown or computationally intractable physics, such as core-collapse supernovae and cosmic string cusps.