Evaluating AI Weather Forecast Models in Detecting Cold Surges over South Korea: Examinig Atmospheric Circulation and Temperature Variability

Abstract

This study evaluates the performance of Artificial Intelligence (AI)-based weather prediction models, GraphCast and Pangu-Weather, by comparing their cold surge forecasts over the Korean Peninsula against those of the physics-based numerical weather prediction models including the Integrated Forecasting System (IFS) operated by the European Centre for Medium-Range Weather Forecasts (ECMWF) and the Korea Integrated Model (KIM). The results show that the AI-based models generally reproduce 2-m temperature variations during cold surge events more accurately than IFS and KIM at short to medium lead times, exhibiting closer agreement with ECMWF Reanalysis version 5 (ERA5) and station observations, while IFS displays a systematic cold bias. To examine large-scale dynamical characteristics, the relationship between forecast lead time and the wave-train index and blocking index is analyzed. For wave-train-type cold surges, all models produce wave-train index values that are generally comparable across lead times, indicating a similar representation of wave-train anomaly strength with limited lead time sensitivity. For blocking-type cold surges, AI-based models yield blocking index values that are closer to those of ERA5, whereas IFS shows a clearer correlation between forecast lead time and the predicted blocking index. Spatial analyses of 300-hPa geopotential height (Z300) anomalies indicate that both AI-based and physics based models reasonably reproduce the overall structure of wave-train and blocking patterns, although differences remain in anomaly intensity and spatial placement. These differences become more evident at longer lead times, reflecting uncertainties in the evolution and placement of large-scale circulation anomalies. In particular, root-mean-square error (RMSE) of Z300 is generally lower for IFS, indicating a more accurate representation of the absolute magnitude and lead time evolution of upper-level anomalies associated with blocking-type cold surges. Overall, the results highlight complementary strengths between AI-based and physics based prediction systems, suggesting that their combined use may improve forecasts of extreme cold surges over South Korea.