astronomy/ai produced the result/Machine Learning Science and Technology 2024 · v2
Eight machine learning methods sorted supernova gravitational wave signals by nuclear physics model
Researchers simulated gravitational waves from collapsing stellar cores under four descriptions of dense nuclear matter, then trained eight kinds of machine learning model to tell, from a waveform alone, which description produced it.
spectrum · one line per step, placed by what the step does · bright lines used AI
Evaluating machine learning models for supernova gravitational wave signal classification
Machine Learning Science and Technology, 2024
doi:10.1088/2632-2153/ada33a · record aix-00049 v2 · checked 2026-10-08
- AI was for
- Classification
- Model family
- Convolutional neural network, Recurrent neural network, Random forest, Support vector machine, Gradient-boosted trees, Linear model, Probabilistic graphical model
- Checked by
- Held-out
- Code
- not reported
The finding the paper is about came from the AI.
What this research was about
When a massive star runs out of fuel, its core collapses in a fraction of a second until it becomes too stiff to squeeze further and bounces. That bounce shakes space itself, sending out gravitational waves: tiny ripples in the geometry of space and time that detectors on Earth can in principle pick up. How hard the core resists being squeezed depends on its equation of state, the relationship between the density of nuclear matter and the pressure it pushes back with. That relationship is not settled, because matter at such densities cannot be made in a laboratory.
Because the bounce signal depends on how stiff the matter is, the shape of the wave carries information about the equation of state. The difficulty is that the shape also depends on how fast the core spins, and the differences between competing equations of state are subtle. The researchers simulated bounce waveforms for four equations of state and around a hundred rotation settings of a single progenitor star, then asked whether a computer could read a waveform and name the equation of state behind it.
Where AI came in
The machine learning models are the measuring instrument here. Each was trained on labelled simulated waveforms and asked to sort an unseen waveform into one of four classes, one per equation of state. The set spanned two neural networks, a ten-layer convolutional network and a four-layer recurrent one, alongside six more conventional methods including a random forest, a support vector machine and gradient boosting. All but naive Bayes averaged above 90 per cent accuracy on the general-relativistic waveforms; the support vector machine reached 99.5±1.0 per cent, while naive Bayes managed 48.9±5 per cent.
The models also served as a test of a shortcut used in simulations. Full general-relativistic gravity is costly to compute, so some codes approximate it with a relativistic correction to Newtonian gravity, known as GREP. A support vector machine trained on GREP waveforms and then shown general-relativistic ones scored 29.9±2.5 per cent, rising to 68.0±4.3 per cent once time was rescaled by the peak wave frequency. The conclusion about the approximation rests on how the classifiers performed, not on a direct comparison of the waveforms.
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
Bounce gravitational wave signals from rotating core-collapse supernova simulations were generated for four nuclear equations of state, giving 452 general-relativistic waveforms and 412 waveforms computed with the general relativistic effective potential (GREP) approximation. Eight machine learning model families were trained to classify which equation of state produced a waveform. All models except naive Bayes exceeded 90 per cent mean accuracy on the general-relativistic data; the support vector machine reached 99.5±1.0 per cent and naive Bayes 48.9±5 per cent. When a support vector machine trained on GREP waveforms was applied to general-relativistic waveforms, accuracy was 29.9±2.5 per cent, rising to 68.0±4.3 per cent after time was normalized by the peak signal frequency.
How AI was used
Gravitational waveforms were produced with the CoCoNuT code for four equations of state and roughly 100 rotational configurations of a single s12 progenitor, in two sets: general relativity with the conformal-flatness condition, and Newtonian hydrodynamics with the general relativistic effective potential. Waveforms were sampled at 10 kHz over the interval -2 to 6 ms around bounce and divided by their amplitude D·Δh. Two deep models (a ten-layer CNN and a four-layer SimpleRNN) and six classical algorithms (random forest, support vector machine, naive Bayes, logistic regression, k-nearest neighbours and XGBoost) were trained as four-class equation-of-state classifiers on the general-relativistic data. Classical model hyperparameters were selected by grid search with 5-fold cross-validation; the deep models used sparse categorical cross-entropy, the Adam optimizer and early stopping on validation loss. Data were split 64:16:20 into training, validation and test sets, and accuracy, recall and precision were averaged over 100 random train-test splits. The same models were then trained on GREP waveforms and applied to general-relativistic waveforms, both as simulated and after normalizing time by the peak gravitational wave frequency. Principal component analysis reduced waveforms from 81 to 2 dimensions so that a linear-kernel support vector machine's decision regions could be inspected, and a random forest was re-run across different signal windows, Tukey tapering parameters and sampling rates.
The shape of the work
Structural · the record, drawn
no AI
Simulate rotating core-collapse gravitational waveforms
Numerical or physics simulation, including where a learned surrogate replaces it.
We obtain GWs from numerical simulations using the code CoCoNuT.where the paper describes this · verbatim
no AI
Preprocess waveforms into fixed-length time series
Cleaning, filtering, normalising or labelling data already obtained.
Before the ML analysis, all waveforms are normalized by dividing them by their amplitudes D⋅Δh.where the paper describes this · verbatim
AI
Tune hyperparameters by grid search with cross-validation
Iterative search over a space.
using the larger portion for hyperparameter tuning through the Grid Search Cross-Validation (GridSearchCV) techniquewhere the paper describes this · verbatim
AI
Train classifiers on general-relativistic waveforms
Fitting model parameters, including fine-tuning an existing model.
We train our models and optimize their hyperparameters using the GR data.where the paper describes this · verbatim
AI
Classify held-out waveforms and score over repeated splits
Running a trained model over new data to predict, classify or score.
Our evaluation process involves repeating the calculations 100 times, with each iteration involving a random train-test split.where the paper describes this · verbatim
AI
Test GREP-trained models on general-relativistic waveforms
Testing outputs against ground truth.
we assess the ability of ML models trained on GREP data to classify the EOS from realistic GW signalswhere the paper describes this · verbatim
AI
Reduce dimensionality and inspect decision boundaries
Extracting understanding from model behaviour.
we apply Principal Component Analysis (PCA) to reduce the dimensionality of the waveforms in both datasets from 81 to 2where the paper describes this · verbatim
AI
Sweep preprocessing choices and re-score accuracy
Iterative search over a space. Its result feeds back into an earlier step.
We first study the influence of the signal length and duration on the accuracy of the classifier using a sliding window approach.where 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 results are the classification accuracies of the ML models themselves, including the conclusion about the GREP approximation, which is drawn from model performance
We use two deep learning algorithms, Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN)where the paper describes this · verbatim
After determining the optimal set of hyperparameters, we split the dataset into training, validation, and test sets with a 64:16:20 ratio.where the paper describes this · verbatim
The gravitational waveforms are publicly accessible at https://doi.org/10.5281/zenodo.13774509.where 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.
- How many were testedThe paper gives no count of what was tested.
- Version of Convolutional neural network (custom, ten layers)Which version of the model was used is not stated.
- Version of Recurrent neural network (SimpleRNN, four layers)Which version of the model was used is not stated.
- Version of Random forestWhich version of the model was used is not stated.
- Version of Support vector machineWhich version of the model was used is not stated.
- Version of Naive BayesWhich version of the model was used is not stated.
- Version of Logistic regressionWhich version of the model was used is not stated.
- Version of k-nearest neighboursWhich version of the model was used is not stated.
- Version of eXtreme gradient boosting (XGBoost)Which version of the model was used is not stated.
- Version of Principal component analysisWhich version of the model was used is not stated.
- What step 3 replacedThe paper gives no basis for what the AI stood in for.
- What step 4 replacedThe paper gives no basis for what the AI stood in for.
- What step 5 replacedThe paper gives no basis for what the AI stood in for.
- What step 6 replacedThe paper gives no basis for what the AI stood in for.
- What step 7 replacedThe paper gives no basis for what the AI stood in for.
- What step 8 replacedThe paper gives no basis for what the AI stood in for.
About this article
Record aix-00049, 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