Day 4

Detailed paper information

Back to list

Paper title Stochastic downscaling of meteorological fields with deep neural networks
  1. Michael Langguth Forschungszentrum Juelich GmbH Speaker
  2. Bing Gong Forschungszentrum Jülich GmbH
  3. Amirpasha Mozaffari Juelich Supercomputing Center
  4. Yan Ji Forschungszentrum Jülich GmbH
  5. Martin G. Schultz Juelich Supercomputing Center
Form of presentation Poster
  • C1. AI and Data Analytics
    • C1.07 ML4Earth: Machine Learning for Earth Sciences
Abstract text Weather forecasts at high spatio-temporal resolution are of great relevance for industry and society. However, contemporary global NWP models deploy grids with a spacing of about 10 km which is too coarse to capture relevant variability in the presence of complex topography. To overcome the limitations of coarse-grained model output, statistical downscaling with deep neural networks is attaining increasing attention.

In this study, a powerful generative adversarial network (GAN) for downscaling the 2m temperature is presented. The generator of the GAN model is built upon a U-net architecture and furthermore equipped with a recurrent layer to obtain a temporarily coherent downscaling product. As an exemplary case study, coarsened 2m temperature fields from the ERA5 reanalysis dataset are downscaled to the same horizontal resolution (0.1°) as the Integrated Forecasting System (IFS) model which runs operationally at the European Centre for Medium-Range Weather Forecasts (ECMWF). We choose Central Europe including the Alps as a proper target region for our downscaling experiment.
Our GAN model is evaluated in terms of several evaluation metrics which measure the error on grid point-level as well as the goodness of the downscaled product in terms of the spatial variability and the produced probability distribution function. Furthermore, we demonstrate how different input quantities help the model to create an improved downscaling product. These quantities comprise dynamic variables such as wind and temperature on different pressure levels, but also static fields such as the surface elevation and the land-sea mask. Incorporating the selected input variables ensures that our neural network for downscaling is capable of capturing challenging situations with the presence of temperature inversions over complex terrain.

The results motivate further development of the deep neural network including a further increase in the spatial resolution of the target product as well as applications to other meteorological variables such as wind or precipitation.