SNOWstorm (v1.0) – a deep-learning based model for near-surface winds and drifting snow in mountain environments

Abstract. Wind-driven redistribution of snow and resulting heterogeneous snow accumulation poses a major uncertainty in mountain hydrology and distributed glacier mass balance models as it is often neglected. High-quality information on the fine-scale wind structure is crucial to predict snow redistribution, but past approaches either relied on highly simplified assumptions or on computationally expensive numerical simulations, inhibiting the application for long-term studies.

To bridge this gap, we introduce SNOWstorm – the snow drift sublimation and transport model. It is designed as a deep-learning based emulator model, that is trained on data from high-resolution (∆x = 50 m) numerical simulations in semi-idealized conditions, to be applicable over a wide range of atmospheric conditions and for a wide range of mountain regions. The model can be driven with input of standard atmospheric variables from coarse- to meso-scale numerical models and predicts near-surface wind fields, and rates of wind-driven snow mass change, drifting snow sublimation and snow transport. Validation experiments show that the model reproduces major terrain-induced flow features as well as patterns of snow redistribution. In a first real-world application study in the European Alps, SNOWstorm predicts wind fields and drifting snow patterns comparable to nested numerical large-eddy simulations, though at more than five orders of magnitude less computational expense. The model thus shows the potential to be used in future studies on multi-seasonal influence of snow redistribution on glacier mass balance in various climatic settings.