Regionalization of FuXi-ENS global predictions in South Korea through dynamical downscaling and model output statistics

While machine learning (ML) weather models are emerging as promising tools for predicting global weather conditions, their coarse resolution and systematic biases restrict proper application in regional contexts. This study assesses the added value of dynamical downscaling and the effectiveness of model output statistics in post-processing to enhance July temperature predictions in South Korea from FuXi-ENS, a state-of-the-art ML-based global forecast, with a one-month lead time. To improve the performance of FuXi-ENS in this region characterized by complex geographic features, dynamical downscaling is conducted using the Weather Research and Forecasting modeling system optimized for the target region. A Joint Gaussian model is then applied to post-process the downscaled predictions and is benchmarked against widely used quantile mapping under the framework of leave-one-year-out cross-validation. Despite significant improvements in the spatial representation of temperature compared to the FuXi-ENS ensemble, the downscaled predictions still exhibit large errors in temporal evolution, often underperforming relative to reference climatological forecasts. This study clearly demonstrates that further improvements based on model output statistics could enhance the accuracy of these predictions. Consequently, it substantiates the synergetic integration of dynamical downscaling with statistical post-processing to transform ML-based global predictions into actionable regional information.

Acknowledgments
This work was supported by Korea Environmental Industry & Technology Institute through Water Management Program for Drought Project, funded by Korea Ministry of Environment (2022003610003).