CURTAIN: Towards a Foundation Model for Neutrino Telescopes

Neutrino telescopes record around a trillion individual photon detections per year, yet there does not exist a SSL task which fully satisfies all three conditions: improving as the pretraining set grows, transferring to a different detector, and outperforming the best model trained without pretraining at all. We introduce CURTAIN, a temporal causal masking scheme and the first SSL task for neutrino telescopes to satisfy all three. Downstream performance improves with the number of pretraining events, from 100 thousand to 3.9 million. Against the same architecture trained from scratch, itself the strongest non-pretrained model, it cuts the median angular error by 51% and the median energy error by 25%. It also transfers to a detector geometry never seen in pretraining, improving direction estimation by 44%. At matched capacity and dataset size, CURTAIN-pretrained encoders beat those pretrained with the existing schemes on both tasks.