Speaker
Description
Global models and reanalysis products do not resolve small-scale details due to their coarse resolution. This is particularly relevant in regions where orography strongly determines atmospheric circulation. Dynamical downscaling with regional models such as WRF (Weather Research and Forecasting) can be a suitable solution for these cases, but this approach demands high computational cost and considerable time to complete the simulations. In recent years, however, there has been growing interest in using AI models based on diffusion architectures to perform statistical downscaling. This methodology not only reduces computational cost and inference time, but can also exploit the probabilistic nature of diffusion models to generate an ensemble of outputs instead of a single deterministic field, in line with the intrinsically probabilistic nature of downscaling.
Here, we present an AI emulator of atmospheric dynamical downscaling for the Canary Islands region, Spain. The model is based on a graph transformer encoder–processor–decoder architecture. This work uses high-resolution (3-km) WRF simulations for the period 1994–2023 to train a diffusion model built on the Anemoi framework, in order to produce high-resolution atmospheric fields from ERA5 data (~25-km resolution). Mean sea level pressure, both wind components, temperature and humidity are used as surface variables and constitute our main target variables to simulate; however, other variables at ten standard vertical levels are also included in the model, all with 6-hour temporal resolution.
Across all five surface variables, the emulator reduces the error relative to WRF compared with simple ERA5 interpolation, with the largest gains over the islands, where the added value of the 3-km simulation is concentrated. Individual ensemble members reproduce the spatial texture of WRF down to grid scale, a property that the ensemble mean loses. This approach enables the generation of high-resolution atmospheric fields that could be useful for renewable energy, risk management and precision agriculture applications, providing a plausible and realistic product in a fraction of the time required by dynamical downscaling.