AirCast-SR: A Foundation Model for Kilometer-Scale Atmospheric Super-Resolution via Latent Consistency Diffusion

In three linesAirCast-SR is an atmospheric super-resolution foundation model that downscales global AI weather forecasts from 28 km to 1 km horizontal resolution. Built on a 3D U-Net conditioned within a Latent Consistency Model diffusion framework, trained on GraphCast forecasts and NOAA data, it produces 67-hour forecasts with near-zero bias and demonstrates zero-shot global transferability to India and Germany.

Summary generated by Claude — human-verified