SNOWstorm – A new emulator model for near-surface winds and drifting snow in glaciological applications

Local snow accumulation in alpine terrain is highly influenced by wind-driven redistribution of snow. Accurate knowledge of the small-scale flow field and the interactions between the snow and the atmosphere are therefore necessary to better simulate and understand glacier mass balance. To bridge the gap between an explicit treatment in high-resolution numerical simulations and computational feasibility for (multi-)seasonal assessments, we introduce SNOWstorm (the SNOW drift Sublimation and TranspORt Model), a deep-learning based model to predict high-resolution near-surface winds, snow redistribution and drifting snow sublimation from low-resolution atmospheric input and high-resolution topography. The model has a stacked U-Net shape architecture and is trained with data from large-eddy simulations (dx=50 m) in a semi-idealized environment. The numerical simulations for the training data set are performed with the Weather Research and Forecasting model (WRF) using a coupled drifting snow module. The surface topography and atmospheric conditions used in WRF reflect the variability seen in alpine terrain over a winter season.

Here we present the basic design of the model, possibilities for applications in the future, as well as first assessments of case studies coupling the model to real-world atmospheric input.