astronomy/ai produced the result/Physical review. D/Physical review. D. 2025 · v2
Neural network ranks gravitational wave triggers in search for neutron star mergers
Researchers built a search pipeline for colliding neutron stars in LIGO data. A convolutional neural network, trained on simulated signals injected into real detector noise, produced the score used to decide which candidates count as detections.
spectrum · one line per step, placed by what the step does · bright lines used AI
Binary neutron star merger search pipeline powered by deep learning
Physical review. D/Physical review. D., 2025
doi:10.1103/physrevd.111.024035 · record aix-00215 v2 · checked 2026-10-09
- AI was for
- Detection
- Model family
- Convolutional neural network
- Checked by
- Benchmark2800 tested
- Code
- available
The finding the paper is about came from the AI.
What this research was about
When two neutron stars spiral together and collide, they shake the fabric of space, and detectors such as LIGO in the United States can register the tremor. The difficulty is that the tremor is faint and the detectors are noisy. Instruments register the sea, the traffic and their own electronics, and they also suffer short bursts of disturbance known as glitches. The usual approach is matched filtering: the data are compared against a large bank of predicted waveforms, and the comparison produces a signal-to-noise ratio, a running measure of how well the data match each template. Deciding which peaks in that measure are real collisions, and which are noise dressed up as a signal, is the hard part.
The researchers set out to build a complete search pipeline, from template bank to final detection list, in which that final judgement is made by a trained network rather than by the hand-built statistics used in existing searches. They then ran it on data from LIGO's third observing run and compared what it found against three established pipelines, PyCBC, GstLAL and MBTA.
Where AI came in
The artificial intelligence here is a convolutional neural network, a kind of model that learns to spot patterns in a sequence of numbers. It was trained from scratch, with one branch for each of the two LIGO detectors and a section that combines them, and it was fed not the raw detector data but the signal-to-noise ratio time series produced by matched filtering. Training used simulated merger signals injected into real detector noise, including noise containing glitches, so the model saw both what a signal looks like and what noise looks like, and was told which was which.
Its output is the pipeline's ranking statistic: the number that says how merger-like each second of data is. The false alarm rates, the sensitivity figures and the recovery of the two known neutron star mergers all follow from that number, so the network stands in place of the hand-designed statistics that conventional pipelines use to rank their candidates.
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 authors trained a convolutional neural network to identify binary neutron star mergers in the signal-to-noise ratio time series produced by matched filtering real LIGO Hanford and Livingston data from the third observing run. Running the pipeline on 2,800 binary neutron star injections from the GWTC-3 O3 sensitivity injection set, they report a sensitivity comparable to the PyCBC, GstLAL and MBTA offline pipelines below the 1 per 2 months detection threshold, and a 12% increase in the total number of detected events at that threshold when their pipeline's detections are added to those of the three existing pipelines. The pipeline also recovered both confirmed binary neutron star events, GW170817, which has a loud blip glitch about one second before the merger in Livingston, and GW190425, analysed with the Hanford branch output set to zero. A comparison run of PyCBC with the authors' BNS-only template bank was used to estimate how much sensitivity would be lost if the deep learning search covered the full binary parameter space.
How AI was used
Matched filtering with a geometric aligned-spin template bank of 30,858 templates converts detector strain into signal-to-noise ratio time series, which are the network's input rather than the strain itself. Training data were built by sampling BNS parameters from Bilby priors with a distance-proportional distribution and a network SNR threshold of 6, injecting SpinTaylorT4 waveforms into real Hanford and Livingston noise from the first week of O3 with specified glitch fractions of 0.1 and 0.15, filtering each injection with ten templates chosen by approximate overlap, and slicing one-second windows: 750,000 signal SNR time series from 75,000 injection waveforms, 750,000 noise samples from 75,000 noise realisations, and 200,000 validation samples. The model is a residual-block convolutional network with one branch per interferometer, joined by an addition layer and followed by four dense layers, with the peak-time difference passed to the combiner, trained in TensorFlow with binary crossentropy, the ADAM optimiser, dropout, SNR-based sample weighting and early stopping on a custom logarithmically spaced area-under-ROC metric. In the search, SNR time series are computed over the whole bank in chirp-mass-ordered clusters, a coincident peak-finder produces one trigger per cluster per second above SNR 4, and the network predicts at 16 Hz on the highest-SNR series, with the moving average of those predictions used as the ranking statistic. Background was collected on one week of O3 noise extended by 200 time shifts applied to the single-detector subnetwork outputs before the combiner, with a Gaussian fit to the tail used to extrapolate beyond the collected background.
The shape of the work
Structural · the record, drawn
no AI
Generate BNS template bank
Cleaning, filtering, normalising or labelling data already obtained.
We used PyCBC’s pycbc_geom_aligned_bank method to generate our template bank.where the paper describes this · verbatim
no AI
Select real detector noise and identify glitches
Obtaining raw data, whether by measurement, download or retrieval.
We used Omicron to identify the glitches in the noise.where the paper describes this · verbatim
no AI
Inject simulated waveforms and matched filter into SNR time series
Encoding data into features, descriptors, embeddings or graphs.
The strain is then filtered with the 10 selected templateswhere the paper describes this · verbatim
AI
Train the two-branch CNN
Fitting model parameters, including fine-tuning an existing model. The AI stood in for conventional algorithm.
The network was trained with the binary crossentropy loss function using the ADAM optimiser, and an initial learning rate of 10−4.where the paper describes this · verbatim
no AI
Matched filter search data and select triggers
Reducing a candidate set by filtering or ranking, in a single pass.
First, the SNR time series are computed using the entire template bank.where the paper describes this · verbatim
AI
Network inference and ranking statistic, including time-shifted background
Running a trained model over new data to predict, classify or score. The AI stood in for conventional algorithm.
The neural network then makes 16 predictions per second on the SNR time serieswhere the paper describes this · verbatim
no AI
Assign false alarm rates and compare sensitivity
Testing outputs against ground truth.
we take the pipeline’s reported false alarm rates for the set of 2,800 injections and calculate their sensitive distanceswhere 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 neural network's output is the pipeline's ranking statistic, so every detection and every sensitivity figure the paper reports comes from the trained model
By training a convolutional neural network to detect binary neutron star mergers in the signal-to-noise ratio time serieswhere the paper describes this · verbatim
we were left with a set of 2,800 injections to test our search’s sensitivity withwhere the paper describes this · verbatim
The code we developed for collecting our background and running our search is available at.where the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- Trained model weightsWhether the trained model is available is not stated.
- Version of two-detector convolutional neural network (unnamed in paper)Which version of the model was used is not stated.
About this article
Record aix-00215, version 2, checked by a person on 2026-10-09. 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