Transformer-based Surrogate Downscaling Model for Nested Numerical Weather Prediction Grids

High-resolution numerical weather prediction remains computationally expensive. We develop a Transformer-based surrogatedownscaling model for the WRF model, trained on a publicly available multiresolution WRF dataset with three nested domains(9 km, 3 km, and 1 km) over the southwestern Atlantic. The surrogate targets 10 m wind speed and follows an encoder-Transformer-decoder architecture to replace the inner nests (D02 and D03) with a learned mapping from the coarse outerdomain and surface fields to high-resolution winds. We evaluate three downscaling pairs (D01→D02, D02→D03, D01→D03)under 5-fold cross-validation and two input configurations: single-channel input (wind speed) and three-channel input (windspeed, near-surface air temperature, and surface pressure). Results are compared against bicubic interpolation with meanabsolute error and a Sobel gradient metrics over the full-grid, land, and sea. Across all domain pairs, regions, and input settings,the surrogate outperforms bicubic interpolation in at least 82.7% of timesteps. The D01→D02 pair, with higher target land ratio,exhibits the strongest gains: all timesteps improve on land and in the full-grid for Sobel, and land MAE improves in all timestepsfor the three-variable configuration and in all but one timestep for the single-variable model. Improvements are still consistentsfor predominantly oceanic targets and for the more challenging 9 km→1 km mapping. These results indicate that the surrogatecan effectively emulate high-resolution WRF nests for near-surface wind.