Classifying Core-Collapse Supernova Gravitational Waves using Supervised Contrastive Learning

Abstract

The detection and reconstruction of gravitational waves from core-collapse supernovae (CCSN) present significant challenges due to the highly stochastic nature of the signals and the complexity of detector noise. In this work, we introduce a deep learning framework utilizing a ResNet-50 encoder pre-trained via supervised contrastive learning to classify CCSN signals and distinguish them from instrumental noise artifacts. Our approach explicitly optimizes the feature space to maximize intra-class compactness and inter-class separability. Using a simulated four-detector network (LIGO Hanford, LIGO Livingston, Virgo, and KAGRA) and realistic datasets injecting magnetorotational and neutrino-driven waveforms, we demonstrate that the contrastive learning paradigm establishes a superior metric structure within the embedding space, significantly enhancing detection efficiency. At a false positive rate of $10^{-4}$, our method achieves a true positive rate (TPR) of nearly $100\%$ for both rotational and neutrino-driven signals within a distance range of $10$--$200$~kpc, while maintaining a TPR of approximately $80\%$ at $1200$~kpc. In contrast, traditional end-to-end methods yield a TPR below $20\%$ for rotational signals at distances $\geq 200$~kpc, and fail to exceed $60\%$ for neutrino-driven signals even at a close proximity of $10$~kpc.