There is a frequent need in climate and weather forecasting to generate forecasts at higher spatial resolutions. At a model spatial resolution of 4 km or less, the Weather Research and Forecasting (WRF) model is considered convection permitting (C-P), generating vastly improved precipitation forecasts. However, C-P scale climate modeling is very expensive, even at a regional scale. Argonne is implementing a generative adversarial network (GAN) with an encoder-decoder architecture that allows the generator to learn precipitation features in higher-resolution data in order to reproduce those features in coarse-resolution data. This methodology will have wide application in generating C-P scale forecasts for weather, climate, and earth system models.

Precipitation is one of the most challenging components in earth system models. While we develop a model for precipitation, we use other variables (e.g., topography, sea level pressure, precipitable water, temperature) as conditions to assist the learning between coarse and fine resolutions. We design the encoder-decoder architecture to get a larger receptive field using CNN. Our preliminary results show that this approach achieves reasonable results when using 50-km precipitation to generate 12-km precipitation.