Entropy-based classification of core-collapse supernova gravitational waves for nuclear equation-of-state inference

arXiv:2511.08010 · astro-ph.HE, astro-ph.IM, gr-qc · Submitted 2025-11-11 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Astrophysics Radio. Generated commentary on the latest astrophysics papers.

Vera: I'm Vera.

Jocelyn: And I'm Jocelyn, here with Subrahmanyan.

Vera: We're looking at the paper: "Entropy-based classification of core-collapse supernova gravitational waves for nuclear equation-of-state inference".

Jocelyn: Here's the gist of the paper. Probing supernovae through gravitational wave entropy studies an entropy-based framework to analyze gravitational-wave signals from core-collapse supernovae.

Vera: This research is important because it identifies the most effective combination of signal representation, entropy measure, feature selection method, and classifier for distinguishing between different progenitor masses in core-collapse supernova gravitational waves.

Jocelyn: Signal Representation and Entropy Features The study employs a probabilistic framework to analyze gravitational-wave signals through four main stages: signal transformation, entropy-based feature extraction, optimal feature selection, and machine learning classification.

Vera: Signals can be represented in the time, frequency and time-frequency domains.

Title and authors: Tom: Okay. Basically, they’re building this whole system to take those gravitational waves from core-collapse supernovae and use four steps to classify them. They start by looking at the signal in different ways—time domain, frequency domain, and time-frequency domain. Then they extract features using different entropy measures, like Shannon or Rényi. After that, they use an AI classifier to decide which mass class the supernova came from.

Vera: So it’s a whole pipeline, right? Signal transformation to feature extraction to classification. What do you think is the big picture of all that?

Jocelyn: The core idea is finding the single best combination of all those parts—the right way to represent the signal, the specific entropy measure they should use, how they select which features are most important, and which classifier to run—to get a mass classification that works really well.

Subrahmanyan: I’m interested in what this means for us theoretically. If we can reliably sort out the progenitor mass using these signals, it gives us direct observational constraints on how matter behaves when it gets incredibly dense. It links the wave physics directly to our models of nuclear structure, which is super important for cosmology.

Vera: Exactly. They found that their best setup—the one that worked best—involved using the continuous wavelet transform for the signal representation and Rényi entropy with an alpha parameter of two as the feature measure, combined with a method called Relief-F for selecting the most useful features, and finally a Support Vector Machine to do the actual sorting.

Jocelyn: And they got about seventy-seven percent accuracy in that setup when looking at signals around two hundred signal-to-noise ratios. That’s a solid number for this kind of complex data analysis.

Subrahmanyan: Seventy-seven percent is good, but we need to keep an eye on the caveats. The paper points out that the continuous wavelet transform provided the clearest separation between different mass classes across the range of signal strengths they tested, which suggests that time series representation wasn't as good for this task.

Vera: That makes sense from a data perspective; if you’re trying to distinguish between two things, you want the representation that makes those differences most obvious in the first place. The paper also showed that Rényi entropy with alpha equals two was the most accurate measure among all four entropy types they tested.

Jocelyn: So, we have a very specific recipe now for an AI system: use CWT as your signal input, Rényi entropy alpha two for features, Relief-F to pick the best ones out of those features, and an SVM to make the final call.

Subrahmanyan: And if we think about this in a larger sense, this paper provides a practical tool. It shows how we can use machine learning to analyze these subtle gravitational wave signals and turn them into something that constrains fundamental physics about the stars themselves without needing massive computational simulations for every single possibility.

Vera: Right. It’s moving us closer to using these detections not just as events, but as data points that tell us about the physical conditions inside those dying stars. What we need to watch next is how this works when we start looking at more complex signals, maybe involving things like neutrino physics or other instabilities mentioned in the paper later on.

The paper's summary: Tom: So we’re looking at how they can make this whole classification system even better, based on what they found in the first part of their work. Vera, Jocelyn, what are the main upgrades they suggest?

Vera: Well, they’ve already identified a pretty solid starting point with CWT and Rényi entropy alpha two, but now they’re looking at how to make the whole process more flexible. They want to test different signal representations—like comparing the time series versus the short-time Fourier transform—to see which one actually gives us a better distinction between different progenitor masses.

Jocelyn: That makes sense, because if you can't even get a clear separation using one representation, then any classifier you put in will struggle to do its job well. They’re testing Shannon entropy and Tsallis entropy too, just to see if they give the same results or if Rényi is truly superior.

Subrahmanyan: From a theoretical standpoint, this means we can’t just rely on one signal processing method; we need a more robust framework that can handle the different physical processes happening during the supernova collapse. It pushes us toward a model where the physics itself dictates which mathematical representation is most useful for extracting information.

Vera: Right. And they also tweaked how they select which features to use—they found that Relief-F was pretty good, but they’re comparing it against LASSO and RFE with cross-validation to see if there’s a way to get an even more refined subset of entropy features.

Jocelyn: The idea is that by finding a better feature selection method, we can feed the classifier much cleaner data. If you give the SVM a highly curated set of entropy features, it should perform even better than it did with the first set they tested.

Subrahmanyan: If this refinement leads to an accuracy boost beyond seventy-seven percent, that’s significant because it means our observational tools are getting sharper at probing nuclear physics through these gravitational wave signals. It suggests that the complexity of the source isn't just noise; it’s structured information we need to decode.

Vera: Exactly. They aren't just settling on one configuration; they are showing that there’s a search space here—a whole set of combinations—and they’re systematically improving it. It turns this from a single study into a methodology for how to approach these kinds of astrophysical problems with AI.

Jocelyn: And the implication for follow-up observations is that we should expect future analyses to try these alternative representations and entropy measures, not just stick to the one they found first. We need diversity in our data processing methods.

Subrahmanyan: I think the bigger picture here is about validation. If we can prove that this methodology works across different signal types, it gives us confidence that we’re seeing real physical signals from these events and not just mathematical artifacts of the gravitational wave modeling itself.

The paper's improvements: Tom: So to wrap things up, this paper on Entropy-based classification of core-collapse supernova gravitational waves for nuclear equation-of-state inference basically shows that we can build a whole AI pipeline—using the continuous wavelet transform, Rényi entropy alpha two, Relief-F feature selection, and an SVM classifier—to sort those messy gravitational wave signals into different progenitor mass classes with about seventy-seven percent accuracy.

Jocelyn: Right. It’s a pretty concrete setup for making sense of core-collapse events using statistical features rather than just looking at the raw waveforms directly. The implication is that we get a new way to constrain the nuclear equation of state by classifying these supernovae better.

Subrahmanyan: I think this method offers a practical observational handle on theoretical physics; it gives us an empirical way to test our models of dense matter under extreme gravity, which is exactly what we need when trying to understand how stars collapse.

Vera: And the caveat is that while this setup is very accurate at high signal-to-noise ratios, it relies on the CWT representation being the best one, which they showed was true for separating mass classes.

Jocelyn: So moving forward, the next logical step is testing all those different entropy measures and representations they mentioned to see if we can find an even more robust method for this classification.

Subrahmanyan: I’m looking forward to seeing how this framework integrates with other cosmological probes; if we can connect these gravitational wave results to large-scale structure measurements, that would be a huge confirmation of the whole picture.

Vera: Exactly. We need to keep iterating on this pipeline to make sure we aren't just optimizing for one specific scenario, but finding the most stable classification technique possible across different signal conditions.

Jocelyn: It’s a good tool for pulsar timing array work because it tells us exactly what kind of signal we should be looking for when we see something new in the gravitational wave data.

Subrahmanyan: If future observations show results that contradict this classification, it’ll be a strong signal that we need to look at more complex physics than just these standard core-collapse models suggest.

Conclusion: Jocelyn: So, to wrap up this paper on "Entropy-based classification of core-collapse supernova gravitational waves for nuclear equation-of-state inference," they built an AI system using the continuous wavelet transform, Rényi entropy alpha two, Relief-F feature selection, and a Support Vector Machine to sort those gravitational wave signals into different progenitor mass classes with about seventy seven percent accuracy.

Vera: Right. It’s a concrete setup for making sense of core-collapse events using statistical features instead of just looking at the raw waveforms directly. The implication is that we get a new way to constrain the nuclear equation of state by classifying these supernovae better.

Subrahmanyan: I think this method offers a practical observational handle on theoretical physics; it gives us an empirical way to test our models of dense matter under extreme gravity, which is exactly what we need when trying to understand how stars collapse.

Jocelyn: It’s a good tool for pulsar timing array work because it tells us exactly what kind of signal we should be looking for when we see something new in the gravitational wave data.

Vera: And the caveat is that while this setup is very accurate at high signal-to-noise ratios, it relies on the continuous wavelet transform being the best one for separating those mass classes.

Subrahmanyan: If future observations show results that contradict this classification, it’ll be a strong signal that we need to look at more complex physics than just these standard core-collapse models suggest.

Jocelyn: So moving forward, the next logical step is testing all those different entropy measures and representations they mentioned to see if we can find an even more robust method for this classification.

Vera: Exactly. We need to keep iterating on this pipeline to make sure we aren't just optimizing for one specific scenario, but finding the most stable classification technique possible across different signal conditions.

Subrahmanyan: I’m looking forward to seeing how this framework integrates with other cosmological probes; if we can connect these gravitational wave results to large-scale structure measurements, that would be a huge confirmation of the whole picture.

Jocelyn: It seems like a concrete tool we can use to probe the interior structure of stars that are about to explode. The next step is figuring out how much better this gets when we move beyond just rotating bounce waveforms, for example.

Vera: I agree; it gives us a way to test different physical scenarios against the data we collect from gravitational waves. It’s an interesting approach to constraining nuclear physics without needing to run computationally intensive simulations for every single case.

Subrahmanyan: So, we can use this AI system setup—CWT, Rényi entropy alpha two, Relief-F selection, and SVM—to rapidly sort through the incoming stream of gravitational wave data from core-collapse events. It’s about using these statistical tools to make sense of the chaos of stellar collapse and connect it back to the fundamental physics governing those dense objects.

Jocelyn: That’s what we had here on this paper, "Entropy-based classification of core-collapse supernova gravitational waves for nuclear equation-of-state inference." Thanks for tuning in with us today.

Department of Electronics and Astrophysics, Al-Farabi Kazakh National University · Energetic Cosmos Laboratory, Nazarbayev University · Department of Physics, Nazarbayev University · Institute of Experimental and Theoretical Physics · Departamento de Astronomía y Astrofísica, Universitat de València · Observatori Astronòmic, Universitat de València · Department of Statistics, University of Auckland · School of Artificial Intelligence and Data Science, Astana IT University

astro-ph.HE, astro-ph.IM, gr-qc

Submitted: 2025-11-11

Updated: 2026-10-31

Comments: Accepted for publication in CQG

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 65/100

The gist: Probing supernovae through gravitational wave entropy studies an entropy-based framework to analyze gravitational-wave signals from core-collapse supernovae.

Key concepts

Signal Representation
This refers to how the raw gravitational wave data is transformed into a format suitable for analysis. The study compared time-domain, frequency-domain, and time-frequency representations like the Short-Time Fourier Transform (STFT) and Continuous Wavelet Transform (CWT). The CWT was found to be the best choice for separating different supernova mass classes.
Entropy Measures
Entropy measures quantify the uncertainty or randomness within a signal's features. Four types were tested: Shannon, exponential, Rényi, and Tsallis. Rényi entropy with $\alpha=2$ proved to be the most effective measure for extracting discriminative information about the supernova progenitor masses from the CWT representation.
Feature Selection (Relief-F)
This is a method used to choose the most important characteristics (features) from a large set of extracted entropy values. Relief-F, a distance-based approach, was found to be superior in this study compared to methods like LASSO and RFE-CV because it selected a more refined subset of features that best distinguished between different supernova mass classes.
SVM Classifier
Support Vector Machines (SVM) were used as the machine learning model to classify the gravitational wave signals into different progenitor mass categories. The SVM classifier performed optimally, achieving a classification accuracy of around 77% in the final, optimized configuration tested against other classifiers.

Terminology

Summary

Probing supernovae through gravitational wave entropy studies an entropy-based framework to analyze gravitational-wave signals from core-collapse supernovae. p. 0

Where the paper says it

abylkairov@nu. kz Abstract: We study an entropy-based framework to analyze gravitational-wave signals from core-collapse supernovae.

This research is important because it identifies the most effective combination of signal representation, entropy measure, feature selection method, and classifier for distinguishing between different progenitor masses in core-collapse supernova gravitational waves. p. 0
Where the paper says it

We find that the combination of Rényi entropy from the wavelet domain and the Relief-F selection method yields the most effective discrimination among different gravitational-wave signals. Keywords: core-collapse supernovae, gravitational waves, entropy feature, feature selection 1.

Signal Representation and Entropy Features

The study employs a probabilistic framework to analyze gravitational-wave signals through four main stages: signal transformation, entropy-based feature extraction, optimal feature selection, and machine learning classification. Signals can be represented in the time, frequency and time-frequency domains. p. 2

Where the paper says it

The overall workflow of the proposed methodology is illustrated in Figure 1. Signal Representation Signals can be represented in the time, frequency and time-frequency domains.

To capture how frequencies change over time, the authors employ two common time-frequency representations: the short-time Fourier transform (STFT) and the continuous wavelet transform (CWT). p. 2
Where the paper says it

Signal Representation Signals can be represented in the time, frequency and time-frequency domains. To capture how frequencies change over time, we employ two common time-frequency representations: the short-time Fourier transform (STFT), which uses fixed-size sliding windows, and the continuous wavelet transform (CWT), which uses scale-dependent sliding windows to achieve higher time resolution at high frequencies and higher frequency resolution at low frequencies.

The CWT provides simultaneous localization in both the time and frequency domains. p. 4
Where the paper says it

Version November 12, 2025 submitted to Entropy 5 of 19 The CWT provides simultaneous localization in both the time and frequency domains [91,92]. For a segment, the CWT is defined as Wx(s, τ) = 1 √ s Z ∞ −∞ x(t) ψ ∗ t − τ s dt, (4) where ψ is the mother wavelet, s denotes scale.

The authors compute four types of entropy for each representation: Shannon, exponential, Rényi, and Tsallis. p. 0
Where the paper says it

We use waveforms generated by numerical simulations and analyze them in both the time domain and the time-frequency domain using short-time Fourier and continuous wavelet transforms. From each representation, we compute four entropy measures – Shannon, exponential, Rényi, and Tsallis – and apply three feature selection methods to identify the most informative features.

Feature Selection Procedures

The researchers identify the most informative characteristics using three feature-selection algorithms: Least Absolute Shrinkage and Selection Operator (LASSO), Recursive Feature Elimination with CrossValidation (RFE-CV), and Relief-F. LASSO applies an L1 regularization penalty t. p. 4

Where the paper says it

In this work, we employ three representative algorithms: LASSO, RFE-CV, and the Relief-F algorithm. LASSO applies an L1 regularization penalty t

RFE-CV iteratively eliminates the least relevant features while optimizing performance through cross-validation. p. 5
Where the paper says it

01] suppress irrelevant coefficients and produce sparse, interpretable models [99–101]. RFE-CV iteratively eliminates the least relevant features while optimizing performance through cross-validation [102].

Relief-F, a distance-based approach, evaluates features by their ability to distinguish neighboring samples from different classes. p. 5
Where the paper says it

RFE-CV iteratively eliminates the least relevant features while optimizing performance through cross-validation [102]. Relief-F, a distance-based approach, evaluates features by their ability to distinguish neighboring samples from different classes, providing robustness against noise and missing data [103,104].

Classification Results and Optimal Configuration

The optimal configuration identified consists of the CWT as the signal representation, Rényi entropy with parameter α = 2 as the entropy feature, Relief-F as the feature selection method, and SVM as the classification algorithm. This combination provides the highest classification accuracy and most stable performance across varying SNR levels. p. 7

Where the paper says it

In Figure 3, the average Rényi entropy (α = 2) as a function of SNR for the three signal representations: CWT, STFT, and time series. Among them, the CWT representation provides the clearest visual separation between different mass classes across the SNR range, suggesting that it preserves more discriminative information under varying noise levels.

The optimal configuration achieves a classification accuracy of ∼ 77% with this setup. p. 9
Where the paper says it

As shown in Table 4, the SVM achieves the highest accuracy among all classifiers, reaching 77. 4% in the default configuration.

Key Findings on Representation and Entropy

The CWT representation provides the clearest visual separation between different mass classes across the SNR range. The time series representation shows moderate class distinction, while the STFT-based entropy exhibits no clear separation between masses. p. 7

Where the paper says it

Among them, the CWT representation provides the clearest visual separation between different mass classes across the SNR range, suggesting that it preserves more discriminative information under varying noise levels. The time series representation shows moderate class distinction, while the STFT-based entropy exhibits no clear separation between masses.

The CWT representation shows the most stable and consistent entropy behavior, making it the most suitable representation for subsequent classification. p. 7
Where the paper says it

In contrast, the CWT adapts its resolution across frequencies, retaining both low- and high-frequency components that vary with progenitor mass. Overall, the CWT representation shows the most stable and consistent entropy behavior, making it the most suitable representation for subsequent classification.

Rényi entropy with α = 2 derived from the CWT achieves the highest accuracy among all tested measures. p. 8
Where the paper says it

To complement these findings, Table 3 reports the classification accuracies obtained using different entropy measures derived from CWT, STFT, and time series representations. The CWT consistently achieves the highest accuracy, exceeding ∼ 0.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved AI system can classify core-collapse supernova gravitational-wave signals by using a combination of CWT representation, Rényi entropy with alpha = 2 as the entropy feature, Relief-F feature selection method, and an SVM classifier to achieve classification accuracy of around 77 percent at SNR = 200.

  2. The system can be adapted to analyze different signal representations (time series and STFT) and compare their performance using entropy measures like Shannon, Exponential, Rényi with alpha = 2, and Tsallis with q = 1.5 to determine which representation provides the most discriminative information for mass classification.

  3. The AI system can be tuned to optimize feature selection by employing the Relief-F algorithm, which led to a modest improvement in accuracy compared to LASSO and RFE-CV, suggesting it selects a more refined and discriminative subset of entropy features.

  4. The system can utilize an SVM classifier as the primary model due to its performance, reaching 77.4% in the default configuration, which is superior to other tested classifiers like k-NN or Random Forest for this specific task at high SNR.

Abstract

Gravitational waves from core-collapse supernovae provide a unique probe of the physical processes occurring during stellar collapse and of the properties of dense nuclear matter. In this work we investigate the classification of supernova gravitational-wave signals associated with different nuclear equations of state using an entropy-based machine learning framework. The analysis combines signal decomposition, feature extraction, and feature selection techniques to identify informative characteristics of the waveforms and distinguish between different equation of state models. We evaluate the performance of several machine learning classifiers using a compact set of selected features and assess their robustness in the presence of detector noise. We find that entropy-based representations capture the underlying differences between waveform classes and enable accurate classification while maintaining better performance compared to raw data under noisy conditions. These results demonstrate the potential of complexity-based feature extraction as an efficient and interpretable approach for gravitational wave studies of core-collapse supernovae.

Sources

Related papers