Fast and accurate learned multiresolution dynamical downscaling for precipitation

Abstract. This study develops a neural-network-based approach for emulating high-resolution modeled precipitation data with comparable statistical properties but at greatly reduced computational cost. The key idea is to use combination of low- and simulations (that differ not only in spatial resolution also geospatial patterns) train neural network map from the former latter. Specifically, we define two types CNNs, one that stacks variables directly encodes each variable before stacking, CNN type both conventional loss function, such as mean square error (MSE), conditional generative adversarial (CGAN), total four variants. We compare new CNN-derived results generated original simulations, bilinear interpolater state-of-the-art CNN-based super-resolution (SR) technique. Results show SR technique produces similar those interpolator smoother temporal distributions smaller variabilities extremes than simulations. While CNNs trained by MSE generate better over some regions do, their predictions are still biased CGAN more realistic physically reasonable results, capturing variability time space intense long-lasting storms. proposed downscaling can downscale 50 12 km 14 min 30 years once (training takes 4 h using 1 GPU), while dynamical would take month 600 CPU cores contiguous United States.