Toward More Realistic Machine-Learning Inference of the Dense-Matter Equation of State from Supernova Gravitational Waves

Abstract

Gravitational waves from core-collapse supernovae offer a unique probe of the equation of state (EOS) of dense nuclear matter. For rapidly rotating stars, previous machine-learning studies demonstrated promising EOS classification accuracy. However, these analyses relied on several simplifying assumptions. In this work, we relax three key assumptions. First, we include real detector noise. Second, we expand the analysis from a single progenitor model to four models spanning 12 to 40 solar masses, and for each mass we consider multiple rotational configurations, from slow to rapid. Third, we introduce uncertainty in the core bounce time of up to 20 ms, rather than assuming it is known precisely. We find that none of these effects significantly degrades EOS classification performance. Instead, the larger dataset associated with multiple progenitor models and noise realizations improves training and classification accuracy. This study is a step in a broader effort to progressively incorporate more realistic conditions into gravitational-wave inference for core-collapse supernovae.

Figures (7)

• Figure 1: Radial profiles of density (solid lines) and specific entropy (dashed lines) for the 12 (blue), 15 (orange), 27 (green), and 40 (red) $M_\odot$ progenitor models prior to core collapse.
• Figure 2: Example of a gravitational wave signal injected into O4a detector noise with $\mathrm{SNR}=20$.
• Figure 3: Accuracy as a function of SNR for real and simulated detector noises. Results are shown for two waveform representations in the time and frequency domains. At $\mathrm{SNR}=200$, the time series classifier achieves $92.3\%\pm2.9\%$ accuracy for simulated noise and $91.6\%\pm3.1\%$ for real noise, while the frequency domain classifier achieves $87.2\%\pm2.9\%$ for simulated noise and $85.3\%\pm2.9\%$ for real noise. The quoted uncertainties correspond to the $1\sigma$ standard deviation of the accuracies obtained from the 50 independent train--test splits described above.
• Figure 4: EOS classification accuracy as a function of SNR. Blue triangles represent classification results in the time domain, red circles indicate classification in the frequency domain, and green rectangles show frequency-domain classification using an expanded dataset. The left, middle, and right panels correspond to the case with 0, 10, and 20 ms bounce time uncertainty cases. The title of each sub-panel indicates the assumed bounce time uncertainty $\Delta t_\mathrm{b}$.
• Figure 5: Examples of two GW signals with two different bounce time assignments. The bounce time uncertainty is $\Delta t_\mathrm{b}=20$ ms. Red and blue lines show the same signal with different assigned bounce times, illustrating that variations arise from both shifted bounce times and intrinsic waveform differences. The gray shaded region indicates the bounce time possible position interval.