Applying Deep-Learning Technologies for Km-Scale Data Assimilation and LES Downscaling

Taking the advantage of the output archive of 3-km, 1-km, and 200m-grid WRF-FDDA (weather research and forecasting with four-dimensional data assimilation) system running over a complex terrain area in the Mideast of China, a data assimilation method based on an Auto-encoder network (AE) and a super-resolution technique based on generative adversarial network (GAN) are developed. The novel AE-based data assimilation method realizes surface weather analysis by mapping observation data to the latent space of a well-trained AE network, and then decoded to the gridded analysis. This approach is named latent space data assimilation with observation only (LSDA-OOA). By training the network with a year of the WRF output data, LSDA-OOA outperforms the traditional four-dimensional data assimilation approach. On the other hand, using the same numerical weather prediction model dataset, a U-LSGAN (Least-Square GAN with Unet) super-resolution network is designed and trained for downscaling the 1-km and 3-km WRF model outputs to ultra-high resolution 200-m WRF-LES model output. The results show an outstanding ability of U-LSGAN in recovering most detailed structures of the LES model fields when using the 1-km WRF output as the input and the model also exhibits a reasonably good ability when using the 3-km WRF output as the input. In general, this work demonstrates that the modern AI deep learning technologies possess a great potential in km data assimilation and sub-kilometer weather simulation and thus should be exploited to enhances the analysis and forecasting of regional and local weather circulations.