Time-Domain Synthesis of Gravitational-Wave Detector Glitches using Class-Conditional Derivative Generative Adversarial Networks
Abstract
Gravitational-wave detectors such as LIGO, Virgo, and KAGRA are highly sensitive instruments susceptible to many noise sources. Short-duration transient noise events, known as glitches, pose a particular challenge for data analysis pipelines, as they can mimic or obscure astrophysical signals. We present GlitchGAN, a class-conditional generative model built on the Conditional Derivative GAN (cDVGAN) architecture, capable of synthesizing seven glitch types from LIGO's third observing run (O3) directly in the time domain. GlitchGAN generalizes effectively, learning to reproduce a diverse glitch space consistent with high-quality DeepExtractor reconstructions, and can generate hybrid glitch morphologies by interpolating across its class-conditioning vector. It generates 1000 glitches in under 22 seconds on a CPU, suitable for detector simulations, mock data challenges, and pipeline validation. Synthetic glitches are validated using the Gravity Spy classifier and UMAP embeddings, both showing strong agreement with real data. To probe residual distributional differences, we train a separate holdout GlitchGAN model and use a downstream CNN to distinguish held-out real glitches from synthetic ones: detectability is high in a clean representation but drops substantially once both populations are injected into realistic detector noise, the condition under which they would typically be used. Despite this, GlitchGAN-generated glitches remain practically useful: augmenting real training sets with synthetic samples matches simple duplication of real data when data is abundant, and increasingly outperforms it as real data becomes scarce. Finally, we highlight a limitation of magnitude-only spectrograms: magnitude Q-transform classifiers can confidently misclassify physically unrealistic glitches from less robust models, underscoring the need for validation methods that preserve phase information.
Related
- Bridging the Population Synthesis of Supermassive Binary Black Holes and the Gravitational Wave Background
- Gravitational Wave Bias in IllustrisTNG300 from Machine-Learned Population-Synthesis Calibrated Merger Rates
- THFR-1.2: Spatiotemporal Metric as a Topological Derivative of Informational Latency
- Cosmology in sixth-order higher-derivative theories of gravity
- Finite-domain inverse reconstruction of a Morris--Thorne-type wormhole candidate with a derivative-consistent f(R) curvature sector
BibTeX
@misc{dooney2026timedomain,
title = {{Time-Domain Synthesis of Gravitational-Wave Detector Glitches using Class-Conditional Derivative Generative Adversarial Networks}},
author = {Tom Dooney and Mees de Boer and Harsh Narola and Melissa Lopez and Stefano Bromuri and Daniel Stanley Tan and Chris Van Den Broeck},
howpublished = {arXiv},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2606.27227},
eprint = {2606.27227},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2606.27227},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Time-Domain Synthesis of Gravitational-Wave Detector Glitches using Class-Conditional Derivative Generative Adversarial Networks
AU - Tom Dooney
AU - Mees de Boer
AU - Harsh Narola
AU - Melissa Lopez
AU - Stefano Bromuri
AU - Daniel Stanley Tan
AU - Chris Van Den Broeck
PY - 2026
DA - 2026/09/14
PB - arXiv
DO - 10.48550/arxiv.2606.27227
UR - https://arxiv.org/abs/2606.27227
AB - Gravitational-wave detectors such as LIGO, Virgo, and KAGRA are highly sensitive instruments susceptible to many noise sources. Short-duration transient noise events, known as glitches, pose a particular challenge for data analysis pipelines, as they can mimic or obscure astrophysical signals. We present GlitchGAN, a class-conditional generative model built on the Conditional Derivative GAN (cDVGAN) architecture, capable of synthesizing seven glitch types from LIGO's third observing run (O3) directly in the time domain. GlitchGAN generalizes effectively, learning to reproduce a diverse glitch space consistent with high-quality DeepExtractor reconstructions, and can generate hybrid glitch morphologies by interpolating across its class-conditioning vector. It generates 1000 glitches in under 22 seconds on a CPU, suitable for detector simulations, mock data challenges, and pipeline validation. Synthetic glitches are validated using the Gravity Spy classifier and UMAP embeddings, both showing strong agreement with real data. To probe residual distributional differences, we train a separate holdout GlitchGAN model and use a downstream CNN to distinguish held-out real glitches from synthetic ones: detectability is high in a clean representation but drops substantially once both populations are injected into realistic detector noise, the condition under which they would typically be used. Despite this, GlitchGAN-generated glitches remain practically useful: augmenting real training sets with synthetic samples matches simple duplication of real data when data is abundant, and increasingly outperforms it as real data becomes scarce. Finally, we highlight a limitation of magnitude-only spectrograms: magnitude Q-transform classifiers can confidently misclassify physically unrealistic glitches from less robust models, underscoring the need for validation methods that preserve phase information.
ER -