We introduce GW-Whisper, an innovative application of OpenAI’s Whisper model, originally designed for speech recognition, to gravitational wave (GW) data analysis. As the volume of data from advanced detectors like LIGO and Virgo grows, traditional methods face scalability challenges. GW-Whisper leverages Whisper’s pre-trained architecture to address critical tasks in GW astronomy, including signal detection and glitch classification, by fine-tuning the model using the parameter-efficient DoRA method. This fine-tuning updates only 0.5% of the model’s parameters, significantly reducing computational costs.
We implement two adaptations of the original Whisper model. We perform binary classification between GW signals vs. background noise and multi-class classification between GWs and noise artifacts (glitches) using off-the-shelf Whisper encoder without any domain-speciifc adjustments. For a more robust GW search over continuous 1-month LIGO data, we implement several GW-specific modifications of the original GW-Whisper configuration. This includes replacing the speech log-mel front end with per-detector Q-transforms (∼1 s, 20–1024 Hz) and a lightweight Q-Adapter (small 2D-CNN) that reshapes/normalizes features. We apply this modified architecture on the MLGWSC-1 datasets that benchmarks machine learning models against traditional search algorithms.
The architecture of the domain-specific GW-Whisper is shown:
The Laser Interferometer Gravitational-Wave Observatory (LIGO) is a revolutionary astrophysical observatory designed specifically to detect gravitational waves. These waves are ripples in spacetime caused by violent cosmic events such as the merger of black holes or neutron stars. LIGO's detections provide a novel way to study the universe, complementary to traditional electromagnetic observatories.
LIGO consists of two large-scale interferometers located in Hanford, Washington, and Livingston, Louisiana. Each interferometer has two arms, each 4 km long. A laser beam is split and sent down each arm, reflecting back from mirrors at the end. When a gravitational wave passes through, it stretches one arm and compresses the other, causing a tiny change in the length of the arms. This changes the path the light travels, leading to an interference pattern when the split beams are recombined. By measuring this interference, LIGO can detect the passage of a gravitational wave. This data, which is a frequency signal over time, is similar to audio data, which has been studied and used extensively in the AI/ML world.
| Repository | Status | Description | Paper link |
|---|---|---|---|
| Signal_vs_Noise | Public | Code repo for signal vs. noise classification | link |
| Glitch classification | Public | Code repo for glitch classification | link |
| MLGWSC-1 analysis | Public | Updated model for MLGWSC-1 analysis | link |
Each repository contains its own detailed documentation and setup instructions. This repo contains introductory information to the project and working with graviational wave data. We recommend starting with
- Our wiki page for understanding our core architecture and methodology.
- Our exploratory notebooks to begin working with graviational wave data.
The codebase is structured as follows:
GW-Whisper/
├── Signal_vs_Noise/
| ├── Efficiency_test/
| | ├── src/
│ │ └── ...
│ ├── Real_events/
│ │ ├── evaluation_real_events.py
│ │ └── ...
| ├── data/
│ ├── results/
│ ├── src/
| ├── preprocess.py
│ ├── run_train.py
│ └── ...
├── Glitch_classification/
│ ├── src/
│ ├── results/
│ └── generate_glitch_dataset.py
| └── ...
├── MLGWSC-1/
│ └── train.py
| └── inference.py
| └── evaluate.py
└── requirements.txt
└── ...
This project is developed at Vanderbilt University as a joint effort between @vanderbilt-data-science and the Physics and Astronomy department. For inquiries, please contact the project leads.
Karan Jani, Ph.D. (PI) - Associate Professor of Physics and Astronomy. Email: karan.jani@vanderbilt.edu
Chayan Chatterjee, Ph.D. (Co-PI) - DSI/Astronomy Postdoctoral Researcher. Email: chayan.chatterjee@vanderbilt.edu
Abbie Petulante, Ph.D. DSI Postdoctoral Researcher. Email: abigail.petulante@vanderbilt.edu
Jesse B Spencer-Smith, Ph.D. - Vanderbilt University DSI (DSI) Chief Data Scientist. Email: jesse.spencer-smith@vanderbilt.edu
Roy Lau - code contributor. Email: roy.lau@vanderbilt.edu
Haowei Fu - code contributor. Email: haowei.fu@vanderbilt.edu
Alara Kaymak - code contributor. Email: alara.kaymak@vanderbilt.edu
Echo Yu - code contributor. Email: tianrun.yu@vanderbilt.edu
Suyash Deshmukh - code contributor. Email: suyash.deshmukh@vanderbilt.edu
Albert Hu - code contributor. Email: yang.hu.1@vanderbilt.edu
Stephen Zhao - code contributor, project manager. Email: chong.zhao.1@vanderbilt.edu
Divya Mereddy - code contributor. Email: divya.mereddy@vanderbilt.edu
Brandon Soubasis - code contributor. Email: brandon.j.soubasis@vanderbilt.edu
Ricky Sun - code contributor. Email: qifeng.sun@vanderbilt.edu
Zack Braasch - project manager. Email: zack.braasch@vanderbilt.edu
Youngbin Kwon - code contributor. Email: youngbin.kwon@vanderbilt.edu
Ziyao Zhang - code contributor. Email: ziyao.zhang@vanderbilt.edu