astronomy/ai produced the result/arXiv 2023 · v2
Neural network strips mains-power hum from LIGO data in about a second
Researchers trained a neural network called DeepClean on data from LIGO's third observing run to subtract 60 Hz electrical noise from gravitational-wave recordings, then checked the cleaned data against a search pipeline and a parameter-estimation run.
spectrum · one line per step, placed by what the step does · bright lines used AI
Demonstration of Machine Learning-assisted real-time noise regression in gravitational wave detectors
arXiv, 2023
doi:10.48550/arxiv.2306.11366 · record aix-00076 v2 · checked 2026-10-08
- AI was for
- Denoising
- Model family
- Convolutional neural network, Autoencoder
- Checked by
- Benchmark78 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, thrown out when massive objects such as black holes or neutron stars spiral together and merge. Detectors like LIGO sense them by watching for minute changes in the distance between mirrors kilometres apart. The signals are faint enough that the instruments also record a great deal of noise: ground motion, equipment vibration, and electrical interference from the mains power supply, which in the United States alternates at 60 times a second. That mains line and the frequencies clustered around it sit awkwardly close to the band where merging neutron stars and black holes make themselves heard.
Detectors carry extra sensors, called witness channels, that monitor the surroundings rather than the passing wave. If the noise in the main recording can be predicted from those witnesses, it can in principle be subtracted away. The researchers set out to do that subtraction with a trained network, run it fast enough to keep up with the detectors in near real time, and then test whether the cleaned data behaved well in the analyses astronomers actually use.
Where AI came in
The AI is the subtraction itself. DeepClean is a convolutional autoencoder, a network that compresses its input through a narrow middle layer and expands it again, learning in the process which patterns matter. It was given the witness-sensor recordings and trained to produce the noise they imply, with no human-labelled answer to copy: its training target was simply to reduce the power left in the main recording across the 55 to 65 Hz band. One model was trained on the first 2000 seconds of each stretch of good data, giving 47 models for the Hanford detector and 72 for Livingston.
Its predictions were then bandpassed and subtracted from the original recording. This stands in for the conventional filtering algorithms previously used to remove the 60 Hz line and its sidebands. The team ran the network both offline, over hour-long chunks, and in a streaming mode that cleans one second at a time using a short window of surrounding data, reporting output at latencies of about one to two seconds. The checks afterwards used no AI: a conventional matched-filter search with GstLAL across 25,000 injected simulated signals, and Bayesian parameter estimation on 78 binary black hole injections.
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
DeepClean, a convolutional autoencoder that predicts witnessed noise in gravitational-wave strain from auxiliary witness sensors, was trained on LIGO Hanford and Livingston data from the third observing run and used to subtract the 60 Hz power-mains line and its sidebands. The test data were a mock data challenge set containing 25,000 injected compact binary signals spread over a 20-day period. In a matched-filter search with GstLAL, more injections were recovered in the DeepClean-processed data for false-alarm rates between 2 per day and 1 per 100 years, with a slight loss of sensitivity at false-alarm rates below 1 per 100 years. Parameter estimation on 78 binary black hole injections gave posteriors from the cleaned data consistent with those from the original data, and the authors report cleaned output at latencies of about 1-2 s in a low-latency configuration.
How AI was used
Science-quality strain segments and the witness channels previously used for 60 Hz sideband subtraction were selected from the third observing run, simulated compact binary signals were injected to form the mock dataset, and the strain was downsampled to 4096 Hz and bandpass filtered to 55-65 Hz while witness channels were upsampled to match and every channel normalised to zero mean and unit variance. A symmetric fully convolutional autoencoder with four downsampling and four transpose-convolution upsampling layers, batch normalisation and tanh activations was then trained with the ADAM optimiser to minimise a loss defined as the ratio of residual to original strain power spectral density summed over the analysis band; one model was trained on the first 2000 s of each science segment, giving 47 models for Hanford and 72 for Livingston. Inference ran over 8 s kernels with 4 s overlaps, whose predictions were Hann-windowed, averaged, bandpassed and subtracted from the original strain, in 3600 s chunks for the offline analysis. A low-latency variant applied the same model to a short kernel in which the 1 s target segment is centred using preceding data and one second of future data, keeping only the target second. The cleaned frames were then passed to a GstLAL matched-filter search and to Bilby with the Dynesty sampler over a 15-dimensional parameter space, each run on both cleaned and original data for comparison.
The shape of the work
Structural · the record, drawn
no AI
Select O3 strain and witness channel data
Obtaining raw data, whether by measurement, download or retrieval.
We selected the low-latency O3 data (labeled as GDS-CALIB_STRAIN) from the 20-day periodwhere the paper describes this · verbatim
no AI
Build mock data by injecting compact binary signals
Cleaning, filtering, normalising or labelling data already obtained.
The mock data is generated by injecting compact binary signals into the O3 strain data from LIGO Hanford and LIGO Livingston.where the paper describes this · verbatim
no AI
Pre-process strain and witness timeseries
Cleaning, filtering, normalising or labelling data already obtained.
DeepClean uses an 8th-order Butterworth filter to bandpass filter the data to the 55-65 Hz range.where the paper describes this · verbatim
AI
Train DeepClean per science segment
Fitting model parameters, including fine-tuning an existing model. The AI stood in for conventional algorithm.
training performed once using the first 2000 s of the sub-segment, regardless of the lengthwhere the paper describes this · verbatim
AI
Offline noise prediction and subtraction
Running a trained model over new data to predict, classify or score. The AI stood in for conventional algorithm.
For the offline analysis, we perform inference on 3600 s-long (1 hour) chunks of datawhere the paper describes this · verbatim
AI
Low-latency cleaning of 1 s frames
Running a trained model over new data to predict, classify or score. The AI stood in for conventional algorithm.
we employ a 4-second kernel that includes 2 seconds of data before and 1 second after the 1-second target segmentwhere the paper describes this · verbatim
no AI
Compare compact binary search sensitivity
Testing outputs against ground truth.
We perform two GstLAL analyses on ∼ 20 days of O3 data to assess the performance of DeepClean.where the paper describes this · verbatim
no AI
Parameter estimation before and after cleaning
Testing outputs against ground truth.
We ran the Dynesty sampler to sample from a 15-dimensional parameter spacewhere 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 result is the cleaned strain produced by the neural network and its downstream effect on search sensitivity and parameter estimation; without the network there is no result.
DeepClean is a convolutional neural network that encodes this activation function using trainable weightswhere the paper describes this · verbatim
a Mock Data challenge (MDC) introduced by the LVK to benchmark and prepare the low latency analysis pipelineswhere the paper describes this · verbatim
our production deployment of DeepClean is highly compute-efficient due to leveraging the GPU resourceswhere 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.
- DataWhether the data are available is not stated.
- Version of DeepCleanWhich version of the model was used is not stated.
About this article
Record aix-00076, version 2, checked by a person on 2026-10-08. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error