Welcome to Fourier-Land: Deep-Learning based downscaling of near-surface winds and drifting snow using WRF simulations over synthetic topographies

Redistribution of snow by the wind has been shown to greatly influence local snow accumulation in alpine terrain. Due to the small-scale nature of this process, previous studies either concentrated on short case studies over small areas or relied on highly simplified wind fields. To bridge the gap towards an assessment of the importance of snow drift over alpine glaciers on seasonal scales we present a new approach using simulations with the Weather Research and Forecasting (WRF) model and deep learning as a computationally efficient downscaling tool for near-surface winds and snow redistribution over complex topography. We created a training data set of high-resolution (dx=50 m) WRF simulations coupled to a novel drifting-snow module that is representative for winter-time alpine environments. The idealized setup allows us to control the degrees of freedom that the final model has to learn. We developed a new technique to create synthetic topographies with similar spectral information as real terrain employing inverse Fourier transforms of scaled fields of random noise. Initial conditions for the WRF simulations are taken to represent the distribution of atmospheric and snow conditions over a winter season. This training data set we feed into a U-Net shape architecture using convolutional neural networks. Here we present first results using a training data set with a reduced number of degrees of freedom as a prove of concept. Future developments will involve adding more complexity to the initial conditions as well as applying it to real-world settings. For this we will couple the model to a glacier mass balance model and run it with real-world atmospheric fields in order to asses the overall importance of drifting snow for alpine glaciers.