Overview

DeepExtractor is a deep learning framework for reconstructing transient noise artefacts (glitches) and gravitational-wave signals in LIGO detector strain data.

The key idea

LIGO strain data contains both astrophysical signals and instrumental glitches — short-duration noise transients that can mimic or obscure real gravitational-wave events. DeepExtractor addresses this by framing glitch reconstruction as a supervised denoising problem:

Given a 2-second stretch of whitened strain containing a glitch, the model learns to output a clean background estimate. Subtracting that background from the input recovers the glitch (or signal) waveform.

Input:  h(t)  = background noise  +  glitch/signal
Output: n̂(t)  = predicted background
Result: ĝ(t)  = h(t) − n̂(t)   ← reconstructed glitch or signal

Architecture

The model is a U-Net operating on STFT spectrograms (magnitude + phase). The input strain is transformed into a 2-channel time-frequency representation, processed by the U-Net encoder- decoder, then converted back to the time domain via iSTFT. The default model, DeepExtractor_257, uses a 4-level U-Net with feature maps [64, 128, 256, 512] and produces 257×257 spectrograms.

Pretrained models

Two pretrained variants of DeepExtractor_257 are provided, both fine-tuned on LIGO O3 data:

Variant	Trained on	Best for
bilby_noise	Simulated LIGO/Virgo noise (bilby)	Simulated data, injection studies
real_noise	Real LIGO O3 strain	Real LIGO O3 detector data

Weights are downloaded automatically from Hugging Face Hub on first use.

Citation

If you use DeepExtractor in your work, please cite:

@article{s91m-c2jw,
  title     = {Time-domain reconstruction of signals and glitches in gravitational wave data with deep learning},
  author    = {Dooney, Tom and Narola, Harsh and Bromuri, Stefano and Curier, R. Lyana and Van Den Broeck, Chris and Caudill, Sarah and Tan, Daniel Stanley},
  journal   = {Phys. Rev. D},
  volume    = {112},
  issue     = {4},
  pages     = {044022},
  numpages  = {24},
  year      = {2025},
  month     = {Aug},
  publisher = {American Physical Society},
  doi       = {10.1103/s91m-c2jw},
  url       = {https://link.aps.org/doi/10.1103/s91m-c2jw}
}