Speaker
Description
Simulation-based inference (SBI) has become the default for intractable-likelihood problems across fundamental physics, but is limited by model misspecification: a density estimator trained on one simulator becomes inaccurate or miscalibrated when applied to data drawn from a different forward model or with unseen instrumental systematics. We study this in 21cm cosmology, where the brightness-temperature field of the Epoch of Reionization is the key target for the Square Kilometre Array. Competing simulators disagree and the true instrument response is unknown in advance. We treat SKATR, a Vision Transformer pretrained with a self-supervised Joint Embedding Predictive Architecture (JEPA) as a foundation model for the field: it is trained once, label-free, on a large set of cheap semi-numerical lightcones, then frozen and used purely as a summarizer, with a lightweight conditional-flow-matching head performing the downstream parameter inference. Despite never seeing the target simulator, its parameters, or any noise, the frozen encoder transfers under a joint domain shift, to a more physically accurate hydrodynamical radiative-transfer simulator and to realistic interferometer noise, matching or exceeding a supervised network trained directly on the noisy target while remaining the only pipeline whose posteriors stay calibrated. In a controlled ablation that holds architecture fixed and varies only the pretraining objective, we find that the self-supervised objective, not data scale, is what buys this robustness. These results position self-supervised pretraining as a general recipe for trustworthy, physics- and systematics-agnostic inference.