Overview

GlitchGAN is a generative model for synthesising realistic LIGO gravitational-wave detector glitches directly in the time domain.

The key idea

Instrumental glitches — short-duration noise transients — are a major challenge for gravitational-wave astronomy. They can mimic or obscure real signals and are difficult to characterise at scale. GlitchGAN addresses this by learning to generate realistic glitch waveforms conditioned on a glitch class, enabling:

• Augmentation of training sets for glitch classifiers

• Injection studies for signal-vs-glitch discrimination

• Morphological interpolation between glitch classes

Architecture

The model is a cDVGAN (class-conditional Derivative GAN) with two Wasserstein discriminators and gradient penalty:

Noise z ~ N(0,I)  +  class vector c  →  Generator  →  ĝ(t)
                                                           ↓
                       Discriminator D₁  (waveform realism)
                       Discriminator D₂  (derivative realism)

The derivative discriminator D₂ operates on the first difference of the waveform, penalising unrealistic temporal structure that the standard discriminator misses. Class conditioning is injected at every layer of the generator via a class vector c ∈ Δ⁶ (probability simplex over 7 classes), allowing continuous interpolation between glitch morphologies.

Supported glitch classes

Class	Description
Blip	Short broadband transient, < 0.1 s, 100–2000 Hz
Fast_Scattering	Periodic arches from optical path-length modulation
Koi_Fish	Long low-frequency transient with frequency evolution
Low_Frequency_Burst	Broadband burst concentrated below 100 Hz
Scattered_Light	Repeating low-frequency arches from mirror motion
Tomte	Short, loud transient with characteristic frequency evolution
Whistle	Narrowband frequency-sweeping transient

Signal specifications

• Sample rate: 4096 Hz

• Duration: 2 seconds (8192 samples)

• Domain: whitened time domain

• Pretrained epoch: 210

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},
}