Teaching AI to find supernova gravitational waves without a template

Gravitational waves from collapsing stars are faint and poorly modeled, making traditional template-matching unreliable. This approach trains a convolutional autoencoder using contrastive learning—pushing noisy versions of the same signal together in a compressed representation—so it learns what a supernova 'looks like' without relying on specific waveform models. Tested against the Einstein Telescope's sensitivity, it reaches ~120 kpc and handles detector glitches better than conventional unsupervised methods, matching supervised networks while generalizing to signal types it has never seen.