astronomy/ai produced the result/Classical and Quantum Gravity 2023 · v2
Neural network sorts six hundred thousand noise blips in LIGO gravitational-wave data
Researchers catalogued the transient noise recorded by the two Advanced LIGO detectors during their first three observing runs. A convolutional neural network, trained on images labelled by experts and volunteers, sorted each noise blip into a class and gave a confidence score.
spectrum · one line per step, placed by what the step does · bright lines used AI
Data quality up to the third observing run of Advanced LIGO: Gravity Spy glitch classifications
Classical and Quantum Gravity, 2023
doi:10.1088/1361-6382/acb633 · record aix-00019 v2 · checked 2026-10-07
- AI was for
- Classification
- Model family
- Convolutional neural network
- Checked by
- Held-out1203 tested
- Code
- not reported
The finding the paper is about came from the AI.
What this research was about
Gravitational waves are tiny ripples in space itself, and the LIGO observatories look for them by measuring minute changes in the length of two long arms of laser light. The signals are so faint that almost anything else can drown them out: traffic, weather, the machinery of the instrument. These short bursts of unwanted noise are known as glitches. They come in recognisable shapes when drawn as a spectrogram, a picture showing how loud each frequency was over a fraction of a second. Knowing which kinds of glitch occur, how often and in which detector matters, because a glitch at the wrong moment can mimic or mask a real signal.
The team set out to record what the noise actually looked like across the first three Advanced LIGO observing runs. Each short noise burst was turned into spectrograms at four different time spans, and then sorted by shape into named classes, with the results published as a public data set.
Where AI came in
The sorting was done by a convolutional neural network, a kind of program that learns to recognise patterns in images. It was shown pictures of glitches that detector specialists and volunteers on the Zooniverse platform had already labelled, and learned to tell the classes apart. One version worked with twenty-two classes for the first two runs; another, with twenty-three classes, was used for the third. For each spectrogram it returned a confidence for every class, and the glitch was filed under whichever scored highest.
The network stood in for people looking at images one by one. It produced the catalogue itself: 233,981 glitches at the Hanford detector and 379,805 at Livingston. Results quoted in the paper keep only classifications above a 90% confidence level. On that basis, Scattered Light accounted for about 47% of third-run glitches at Hanford, while Fast Scattering dominated at Livingston, at 9.05 an hour against 0.22 an hour at Hanford.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
The Gravity Spy project ran convolutional neural networks over time–frequency spectrograms of transient noise (glitches) in Advanced LIGO strain data from the first three observing runs. The analysis classified 233981 glitches from LIGO Hanford and 379805 glitches from LIGO Livingston into morphological classes, each with a per-class confidence score. The distribution of classes differed between the two sites: Scattered Light made up about 47% of O3 glitches classified with confidence above 90% at Hanford, while Fast Scattering was the most common class in Livingston data, occurring at 9.05 per hour there against 0.22 per hour at Hanford. The classifications, together with Omicron trigger metadata and confidence scores, are released on Zenodo.
How AI was used
Excess-power triggers were identified in the LIGO strain channel by the Omicron pipeline, and each trigger was rendered as Omega-scan spectrograms at four durations (0.5 s, 1 s, 2 s and 4 s) centred on the transient. A convolutional neural network with four convolutional layers, each followed by max-pooling and ReLU, then a fully connected layer and a softmax layer, took these spectrograms as input and returned a confidence for each morphological glitch class, with the trigger assigned to the highest-confidence class. Training sets were built by detector-characterisation experts and Zooniverse volunteers: a 22-class set of 7932 samples for the model applied to O1 and O2 data, and a 23-class set of 9631 samples (8427 training, 1203 validation) for the model applied to O3 data, the latter adding Fast Scattering and Blip Low Frequency and dropping None of the Above. The resulting classifications were then filtered at a fiducial 90% confidence threshold for the reported class counts, SNR distributions and hourly rates, and cross-referenced by time against gravitational-wave candidates within a ±5 s window.
The shape of the work
Structural · the record, drawn
no AI
Record detector strain data
Obtaining raw data, whether by measurement, download or retrieval.
The primary data output of these observatories is the strain measured by the interferometerswhere the paper describes this · verbatim
no AI
Identify noise transients with Omicron
Cleaning, filtering, normalising or labelling data already obtained.
All the noise transients analyzed in this paper were detected by the Omicron algorithm analysing the gravitational-wave strain channelwhere the paper describes this · verbatim
no AI
Render time–frequency spectrograms
Encoding data into features, descriptors, embeddings or graphs.
These time–frequency spectrograms are used as the input to Gravity Spy.where the paper describes this · verbatim
no AI
Assemble labelled training set
Cleaning, filtering, normalising or labelling data already obtained.
The original LIGO data set used to train the Gravity Spy CNN was created by detector-characterisation experts and Gravity Spy volunteers.where the paper describes this · verbatim
AI
Train CNN glitch classifier
Fitting model parameters, including fine-tuning an existing model. The AI stood in for manual curation.
this current training data set contains 9631 glitch samples distributed over 23 classeswhere the paper describes this · verbatim
AI
Classify glitches and assign confidence
Running a trained model over new data to predict, classify or score. The AI stood in for manual curation.
For every image input to the CNN, the probability (or confidence) p of belonging to each class is calculatedwhere the paper describes this · verbatim
no AI
Threshold classifications and compute glitch rates
Reducing a candidate set by filtering or ranking, in a single pass.
We mainly use a fiducial 90% confidence threshold for our quoted results.where the paper describes this · verbatim
no AI
Cross-reference classifications with gravitational-wave candidates
Extracting understanding from model behaviour.
This list was compiled by cross-referencing the times associated with public alerts and high-significance candidates from offline analyseswhere the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
The paper's reported result is the catalogue of classified glitches, which was produced by the convolutional neural network; the glitch classes and rates presented depend on the model's output
Gravity Spy uses a CNN, a deep-learning algorithm used primarily for image classification, to analyse the Omega scans.where the paper describes this · verbatim
of these 8427 were used for training and 1203 were used for validationwhere the paper describes this · verbatim
The data release is available from Zenodowhere the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- CodeWhether the code is available is not stated.
- Trained model weightsWhether the trained model is available is not stated.
- ComputeThe hardware or time used is not stated.
- Version of Gravity Spy CNN (O1–O2 model, 22 classes)Which version of the model was used is not stated.
- Version of Gravity Spy CNN (O3 model, 23 classes)Which version of the model was used is not stated.
About this article
Record aix-00019, version 2, checked by a person on 2026-10-07. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY-4.0; quotations are at most 25 words. How we work · Report an error