On the phenomena-oriented validation of spatial neural-network based surface wind downscaling over the Arctic seas

Abstract

Neural-network based downscaling offers a computationally efficient alternative to high-resolution dynamical modeling, yet its validation remains challenging because conventional pointwise metrics provide limited insight into the physical realism of reconstructed atmospheric fields. This study introduces and demonstrates a phenomena-oriented validation framework for evaluating neural-network based spatial downscaling of surface wind fields over the Arctic seas. We implement a deep neural network with skip connections, trained to downscale ERA5 reanalysis data using high-resolution WRF simulations as a reference. The core methodological advance is a rigorous validation approach that moves beyond conventional pointwise metrics. It employs the tracking of polar mesocyclones and the Novaya Zemlya bora to assess the representation of mesoscale atmospheric dynamics. The downscaling model demonstrates a 50-fold greater computational efficiency than non-hydrostatic modeling with the reservation of the need for the collection of training dataset computed with high-resolution WRF model for a limited time period. Validation results show that our deep learning model successfully captures mesoscale spatial variability, closely matches the distributions of vortex intensity characteristics of the high-resolution reference, and produces physically consistent wind fields that yield realistic significant wave heights when used to force a wave model. This work establishes that a phenomena-based assessment is crucial for validating the physical realism of statistical downscaling methods in complex marine environments.