Neural network atmospheric bias correction on heterogeneous data with fine-scale dynamics preservation

Abstract

High-resolution regional numerical weather prediction (NWP) models are essential tools for capturing atmospheric dynamics, including critical mesoscale weather phenomena, yet they often suffer from systematic biases caused by errors in initial conditions, parameterization schemes, and numerical integration. One established way to address these errors is statistical bias correction, which adjusts model outputs toward reference datasets to reduce systematic deviations. Such methods commonly rely on observational datasets that can be sparse in space and time or relatively coarse. However, naively tuning forecasts to these datasets may provoke loss of vital small-scale features present in the original high-resolution predictions. To address this, we propose a statistical bias-correction method designed to improve atmospheric forecast accuracy while preserving the small-scale dynamics of the original predictions. We introduce a novel loss-function term that minimizes the difference between small-scale features in the original and corrected forecasts. A convolutional U-net with a Transformer in the latent space is trained on European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis Version 5 (ERA5) reanalysis data, along with meteorological station observations and scatterometer measurements over the Kara and Barents Seas. The proposed Bidirectional Encoder Representations from Transformers and U-net model achieves accuracy comparable with state-of-the-art neural network methods, while offering superior performance in perceptual metrics describing mesoscale dynamics. The method shows season-dependent skill, with the strongest improvements observed during the sea-ice season. Incorporating specific observational datasets during training leads to consistent error reductions when evaluated against the corresponding sources. The proposed approach is readily applicable in operational weather forecasting.