A U-LSGAN Network for Downscaling Mesoscale Outputs to LES Simulation

Abstract Subkilometer-grid numerical simulations are highly desired for many weather-sensitive applications. A U-Net-based least squares generative adversarial network (U-LSGAN) model that integrates generative adversarial networks with least-squared loss in a U-Net framework is developed to downscale 1 km × 1 km mesoscale model outputs to 200 m × 200 m large-eddy simulation (LES) grids for Lushan and surrounding regions in eastern China, an area characterized by steep mountains and complex land cover. Three key meteorological variables—2-m temperature (T2), 2-m relative humidity (RH2), and 10-m wind speed (WSPD10)—are downscaled, and the impacts of auxiliary inputs including terrain height, land uses, accumulated precipitation, and surface pressure are evaluated. The model uses mean absolute error (MAE), smooth L1 (smooth MAE), and reduced-weight adversarial losses to optimize the downscaling process between coarse-resolution inputs and fine-scale targets, enabling accurate reproduction of subkilometer weather patterns. Results based on 2522 test samples show that U-LSGAN can effectively capture the magnitude and fine-scale structures of all three variables with power spectra closely matching the LES model outputs. In comparison to bilinear interpolation, U-LSGAN achieves dramatic improvements, with MAEs of RH2, T2, and WSPD10 reduced by 36.05%, 43.16%, and 52.02%, respectively. Significance Statement Many industrial applications, such as air traffic control, wind turbine siting, and wind power forecasting, as well as hazardous gas or aerosol accident response, and cloud-seeding weather modification operations require weather analyses and forecasts at very fine scale (grids at intervals of tens to hundreds of meters). Although such refined weather modeling can be done with large-eddy simulation (LES) models, it is typically very computing costly and cannot provide timely service to meet urgent decision needs. Based on 2522 test samples, it is demonstrated that the deep learning model developed in this study can accurately and efficiently downscale 1-km grid surface mesoscale weather fields to 200-m grid LES outputs within seconds and it works well for both flat and complex terrain regions.