Probing Supernovae through gravitational wave entropy

TL;DR

The paper tackles classifying CCSN gravitational-wave signals by converting signals into time-domain and two time-frequency representations, then computing multiple entropy-based features. Using a four-step pipeline (transformation, entropy extraction, feature selection, and ML classification), the authors systematically compare representations, entropy measures, feature selectors, and classifiers, finding that wavelet-domain entropy—specifically Rényi entropy with $\alpha=2$—coupled with Relief-F feature selection and an SVM classifier, delivers the strongest discrimination across progenitor masses (≈77% accuracy). The results emphasize that wavelet-based, non-stationary features better capture CCSN GW information than time-series or fixed-resolution STFT features, and that entropy-based ML approaches can effectively probe underlying explosion physics. These findings provide a promising avenue for using entropy-augmented GW analyses to distinguish stellar progenitors and, more broadly, to explore CCSN dynamics in current and future detectors, while highlighting limitations related to the simplified rotating-bounce test signals and Gaussian-noise assumptions.

Abstract

We study an entropy-based framework to analyze gravitational-wave signals from core-collapse supernovae. 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. We then train machine-learning classifiers on these features to compare the performance of different methodological combinations. 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.

Figures (7)

• Figure S1: Flowchart outlining the main steps of the proposed methodology. The signal representation stage (Time series, STFT, CWT) is described in Section \ref{['sec:representations']}, the entropy feature extraction in Section \ref{['subsec:entropy']}, the feature selection methods (LASSO, Relief-F, RFE-CV) and the classification procedure in Section \ref{['subsec:fsandclass']}.
• Figure S2: Gravitational wave strain as a function of time. Left: clean signals generated for four different source masses. Right: example of a single signal for the SFHo EOS with $T/|W|=0.07$ and a 40 $M_\odot$ mass after the addition of noise corresponding to an SNR of 200.
• Figure S3: Rényi entropy ($\alpha = 2$) as a function of SNR for CWT (a), STFT (b), and time series representations (c).
• Figure S4: Various types of entropy as a function of SNR estimated with CWT. (a) Shannon entropy; (b) Exponential entropy; (c) Rényi entropy; (d) Tsallis entropy.
• Figure S5: Dependence of accuracy on $\alpha$ and $q$ parameters of Rényi and Tsallis entropy at $\mathrm{SNR} = 200$. The remaining parameters correspond to those of the optimal model configuration (Table \ref{['tab:default_model']}). For $q = 1$ and $\alpha = 1$, both Rényi and Tsallis entropies reduce to the Shannon entropy, highlighted by the blue square.