Residual neural networks to classify the high frequency emission in core-collapse supernova gravitational waves

TL;DR

This work presents a ResNet50-based approach to classify the morphology of the High Frequency Feature (HFF) in core-collapse supernova gravitational waves by analyzing time-frequency Morlet scalograms as RGB images. It trains on phenomenological waveforms injected into real LIGO-Virgo O3b noise and tests on numerical CCSN waveforms across distances, achieving near-perfect accuracy at 1 kpc/5 kpc for some detectors and notable degradation at 10 kpc due to SNR limitations. The study demonstrates the feasibility of fast, morphology-based HFF characterization in real interferometric data, highlights the critical role of training-data SNR distributions, and shows that regression-based alternatives underperform compared with end-to-end classification for this task. The results support using TF-image classification as a practical step toward early-stage HFF morphology identification in CCSN GW data, with implications for rapid follow-up and multi-messenger studies.

Abstract

We present a new methodology to explore the morphology of the High Frequency Feature (HFF), i.e., the dominant, rising-frequency GW emission from a proto-neutron star in core-collapse supernovae (CCSNe). We used a residual neural network (ResNet50) to perform multi-class classification of image samples constructed from time-frequency Morlet wavelet scalograms. We defined a three-class problem by categorizing the HFF slope as Steep, Moderate, or Low, according to physically informed ranges. The ResNet50 model was optimized with phenomenological waveforms injected into real noise from the LIGO-Virgo O3b observing run and then tested with numerically simulated CCSN waveforms embedded in the same real noise. At galactic distances of 1 kpc and 5 kpc with H1 and L1 data and 1 kpc with V1 data, we obtained highly accurate results (test accuracies from 0.8933 to 0.9867), which show the feasibility of our methodology. For further distances, we observed declines in test accuracy until 0.8000 with H1 and L1 data at 10 kpc and until 0.5933 with V1 data at 10 kpc, which we attribute to limitations in the input datasets. Our methodology is sufficiently general to enable early-stage characterization of the HFF in real interferometric data.

Figures (15)

• Figure 1: Examples of phenomenological waveforms in the time domain (upper panel) and their Morlet wavelet scalograms (bottom panel). They belong to classes 1, 2, and 3. For HFF slope $2108$ Hz/s, $f_0 = 126.63$ Hz and $f_1 = 3416.08$ Hz; for slope $1505$ Hz/s, $f_0 = 108.14$ Hz and $f_1 = 1774.62$ Hz; and for slope $935$ Hz/s, $f_0 = 105.28$ Hz and $f_1 = 1148.24$ Hz. The time duration of the waveforms varies because of the nature of their model, which include a random force. Waveforms are shown in the absence of noise; however, as detailed in subsection \ref{['sec:dataset_gen']}, they were injected into real LIGO and Virgo interferometric noise to generate the ResNet50 optimization dataset.
• Figure 2: Numerical simulated CCSN waveforms simulations used in this work, both in the time domain (upper panel) and their Morlet wavelet scalograms (bottom panel). Andresen et al. 2019 $m15nr$ waveform comes from a 3D CCSN simulation. Morozova et al. 2018 $M13\_SFHo$ and Cerdá-Durán et al. 2013 $fiducial$ al. waveforms come from 2D CCSN simulations. These waveforms are astrophysically realistic. Then, as detailed in subsection \ref{['sec:results_genrel_wf']}, we used their HFF emission, injected into real LIGO and Virgo interferometric noise, to test the ResNet50 model.
• Figure 3: HFF slope estimation for CCSNe numerical waveforms using linear regression. To unambiguously perform this fit, we previously removed wavelet contributions of the background (defined as pixels with intensity less than the arithmetic mean of all pixels) and features other than the HFF. This estimation was performed to assign, to each waveform, one of the previously defined target classes, as shown in Table \ref{['tab:HFFslopes']}.
• Figure 4: Population exploration of samples used to optimize the ResNet50 model. These contain phenomenological waveforms injected into LIGO-Virgo real O3b noise. Leveraging that these waveforms (by the nature of their model) do not codify the distance, amplitudes were chosen to create a high-SNR dataset to facilitate the learning process.. The SNR values distribution is shown in the bottom panel, in which class 1 samples are shifted to lower SNR values in comparison to SNR distributions for class 2 and 3 samples. Moreover, we have a significant overlap in the SNR distributions of class 2 and class 3 in the range of approximately 40-70, indicating that SNR alone could not be a reliable discriminator between these classes. On the other hand, the upper left and upper right panels show distributions (in the absence of noise) of HFF slope and waveform duration, respectively. Phenomenological waveforms were generated by a stochastic model, and their duration varies from $0.3$ s to just under $1.0$ s.
• Figure 5: Correlation between HFF slopes and SNR values of samples of noise plus phenomenological waveform. The inverse relationship for class 1 samples arises because these reach the upper frequency limit more quickly, resulting in shorter durations and lower accumulated SNR for a given amplitude.