GlitchGAN — Synthetic LIGO Glitch Generator

GlitchGAN is a class-conditional generative model that synthesizes realistic gravitational-wave detector glitches directly in the time domain. It is built on the cDVGAN (Conditional Derivative GAN) architecture, which uses a first-derivative discriminator alongside the standard signal discriminator to enforce temporal smoothness in generated waveforms.

The model is trained on high-quality glitch reconstructions produced by DeepExtractor, a U-Net framework for glitch reconstruction from LIGO strain data, covering seven common glitch classes observed during LIGO's third observing run (O3).

Supported glitch classes

Index	Class	Description
0	Blip	Short, broadband transient
1	Fast_Scattering	Arches at low frequency from scattered light
2	Koi_Fish	Low-frequency dispersive burst
3	Low_Frequency_Burst	Broadband low-frequency transient
4	Scattered_Light	Repeated arch pattern from optical scattering
5	Tomte	Short, low-frequency transient
6	Whistle	Narrow-band high-frequency chirp

Quick start

pip install glitchgan[tensorflow]

from cdvgan import GlitchGAN

# Download pretrained weights automatically (cached after first call)
model = GlitchGAN.from_pretrained()

# Generate 10 Blip glitches at SNR 50
blips = model.generate("Blip", n=10, snr=50)
# blips.shape → (10, 8192)  — 2 s at 4096 Hz

# Generate a mix of classes
glitches = model.generate(["Blip", "Tomte", "Whistle"], n=3, snr=30)

# Class interpolation — hybrid Blip/Koi Fish morphology
hybrids = model.interpolate([0.5, 0, 0.5, 0, 0, 0, 0], n=5)

Output waveforms are 2-second segments at 4096 Hz (8192 samples). Pass snr=None (default) to get the raw normalized generator output, or specify a target SNR to rescale to a physically meaningful amplitude.

Model details

Property	Value
Architecture	cDVGAN (conditional WGAN-GP + derivative discriminator)
Signal length	8192 samples (2 s at 4096 Hz)
Noise dim	100
Num classes	7
Training epochs	210
Optimizer	RMSprop (lr=1e-4)
Gradient penalty weight	10
Framework	TensorFlow / Keras 3

Training data

The model was trained on DeepExtractor reconstructions of real LIGO glitches from O3 (2019–2020), curated using Gravity Spy classifications. Seven glitch classes were selected with balanced class sizes. Waveforms are whitened and normalized before training; the generator therefore learns a normalized waveform distribution and SNR scaling is applied at inference time.

Evaluation

Synthetic glitches were evaluated against real data using:

1. Gravity Spy classification — the fraction of generated samples correctly classified by the state-of-the-art Gravity Spy classifier.
2. UMAP embeddings — unsupervised 3-D projections to compare the morphological distributions of real and synthetic samples.

Selected Gravity Spy results at SNR 100 (100 samples per class):

Class	Accuracy
Blip	88%
Fast Scattering	72%
Koi Fish	20% (SNR-sensitive; 79% at SNR 150)
Low Frequency Burst	100%
Scattered Light	42% (often confused with Fast Scattering)
Tomte	92%
Whistle	94%

UMAP embeddings show strong overlap between real and synthetic samples for most classes. Whistle glitches show partial separation, attributed to the model's difficulty with fine-scale high-frequency structure.

Limitations

• SNR sensitivity: GlitchGAN does not model an SNR distribution. Generated waveforms must be manually rescaled. Classification accuracy for morphologically ambiguous classes (Koi Fish, Scattered Light) is strongly SNR-dependent.
• Whistle class: Synthetic Whistle glitches form a partially distinct cluster from real data in UMAP space, indicating incomplete capture of high-frequency structure.
• Seven classes only: The model covers the seven classes present in the O3 training set. It cannot generate unseen glitch morphologies.
• LIGO-specific: Trained on LIGO H1/L1 data. Generalization to Virgo or KAGRA data has not been evaluated.

Citation

If you use GlitchGAN in your work, please cite:

@article{Dooney2026GlitchGAN,
  title   = {Realistic Time-Domain Synthesis of Gravitational-Wave Detector
             Glitches using Class-Conditional Derivative Generative Adversarial Networks},
  author  = {Dooney, Tom and de Boer, Mees and Narola, Harsh and Lopez, Melissa
             and Bromuri, Stefano and Tan, Daniel Stanley and Van Den Broeck, Chris},
  journal = {Physical Review D},
  year    = {2026},
}

License

MIT