level: research
aircast-sr is a foundation model that performs atmospheric super-resolution, taking global ai weather forecasts at 0.25 degree (~28 km) and downscaling them to 1 km horizontal resolution. it generates hourly forecasts for 67 hours, covering eight coupled surface variables at once. the model uses a three-dimensional u-net inside a latent consistency model diffusion framework. training uses patch-based samples over the contiguous united states, with graphcast forecasts as input and noaa's analysis of record for calibration as the target.
the model achieves near-zero bias across all variables and lead times. this means the downscaled forecasts closely match the high-resolution reference data without systematic over- or under-prediction. by operating in a latent space, the diffusion process becomes more efficient, enabling faster generation of fine-scale details. the approach is designed to handle the complexity of multiple interacting surface variables, such as temperature, wind, and precipitation, in a unified way.
operational weather prediction at kilometer scales is usually too expensive for traditional numerical models. aircast-sr aims to make such detailed forecasts more accessible for applications in energy, agriculture, and disaster management. the model's ability to produce rapid, high-resolution forecasts from coarser ai predictions could support better planning and response in these sectors. the work builds on advances in both ai weather forecasting and generative modeling.
why it matters: it enables cheap, high-resolution weather forecasts for critical sectors like energy and disaster response, using ai to bypass expensive traditional models.