Assessment of Atmospheric Dynamics Based on Neural-Network Downscaling of Near-Surface Wind Speed Fields over the Barents and Kara Seas

Abstract

This study examines the use of a deep learning approach for spatial downscaling (increasing spatial resolution) of near-surface wind over the Barents and Kara Seas using deep feedforward artificial neural networks, aiming to enhance spatial resolution while reducing computational costs compared to non-hydrostatic modeling. Low-resolution input data are obtained from the global atmospheric reanalysis ERA5, while high-resolution reference data are provided by the Weather Research and Forecasting (WRF) model. The results of neural network–based downscaling are compared with those from bilinear interpolation. The proposed model improves the distribution of mesoscale structure lifecycle parameters, bringing them closer to the high-resolution simulation data, and outperforms the latter in computational speed by a factor of 50. The wave height calculated using boundary conditions from the neural network model instead of the non-hydrostatic simulation shows similar values. The developed neural network model also exhibits less than 3% deviation from high-resolution dynamic modeling in terms of the number of mesoscale structures.