AirCast-SR downscales global weather to 1 km using diffusion AI

AirCast-SR is a new foundation model for kilometer-scale atmospheric super-resolution developed by a team including Luitel, Singh, and 12 other researchers from universities and organizations. It takes global AI weather forecasts from GraphCast at 0.25° resolution and downscales them to 1 km horizontal resolution at hourly temporal steps, producing 67-hour forecasts of eight coupled surface variables simultaneously. The model uses a three-dimensional U-Net conditioned within a Latent Consistency Model (LCM) diffusion framework, trained on patch-based samples over the contiguous United States using NOAA's Analysis of Record for Calibration (AORC) as the target. This approach achieves near-zero bias across all variables and lead times, and radial power spectral density analysis confirms preservation of fine-scale atmospheric structure at wavelengths of 10 km to 100 km where coarser models lose spectral power.

The model was validated across three CONUS case studies spanning winter, summer, and spring seasons. Critically, AirCast-SR demonstrates zero-shot global transferability over India and Germany using independent surface station observations without any retraining or fine-tuning. As an open-weights foundation model, it establishes a new paradigm for affordable kilometer-scale AI weather prediction, providing a platform for regional fine-tuning, distillation, and downstream applications in energy, agriculture, disaster management, and climate services. This addresses the computational prohibitions of traditional numerical weather prediction at kilometer scales, democratizing access to fine-grained spatiotemporal weather data.