This paper introduces a new machine-learning method to detect gravitational waves from core-collapse supernovae (CCSNe). The method, called a contrastive self-supervised convolutional autoencoder (CS-CAE), is designed to reduce the effect of random detector noise and to be less dependent on specific waveform templates. In tests with simulated data for the planned Einstein Telescope (ET), the model reaches an effective sensitive distance of roughly 120 kiloparsecs and separates CCSNe signals from stationary noise and short glitches better than a conventional autoencoder baseline.

The authors build CS-CAE by combining three ideas. First, they use a convolutional autoencoder (CAE), a neural network that compresses a time series into a short “latent” code and then reconstructs it. Second, they add a noise-centered latent regularizer that encourages noisy inputs to have latent codes near the origin when there is no signal. Third, they add a projection head trained with a contrastive objective (an InfoNCE-style loss). In training, the contrastive step pairs independent noisy realizations of the same underlying CCSN waveform and pulls their latent codes together while pushing other codes apart. This teaches the encoder to focus on signal-consistent features and ignore incidental noise.

Training data were simulated for three ET interferometer channels and whitened between 10 and 2048 Hz. The simulated CCSN waveforms come from a phenomenological source model meant to reproduce common time–frequency features seen in multi-dimensional simulations: upward-drifting high-frequency proto-neutron-star (PNS) oscillations, low-frequency activity from standing accretion shock instability (SASI) and convection, possible rapid-rotation bounce-like bursts, and broad stochastic components. Detector noise was generated as Gaussian noise consistent with the ET-D amplitude spectral density.