AI and the World's Hard Problems · Physics · arXiv · 2026-10-23
CURTAIN: Towards a Foundation Model for Neutrino Telescopes
Neutrino telescopes record over a trillion photon detections a year, but no self-supervised learning task had met all three goals: improving with more pretraining data, transferring to a different detector, and beating models trained without pretraining. The authors introduce CURTAIN, a temporal masking task that meets all three. Pretraining the DeepIce encoder with it cut median angular error by 45% and median energy error by 31% compared with training from scratch, and reduced angular error by 36% on an unseen detector geometry.