Abstract The rapid development of artificial intelligence weather forecasting models (AIWFMs) has revolutionized weather prediction. GraphCast, a leading AIWFM, outperforms state‐of‐the‐art physics‐based numerical models at the surface and in the lower troposphere. However, it remains unclear how well GraphCast predicts stratospheric variability and represents stratosphere‐troposphere coupling, an important source of surface predictability on subseasonal to seasonal (S2S) timescales. In this study, we evaluate GraphCast's predictions of the stratosphere and its ability to capture stratosphere‐troposphere coupling, focusing on sudden stratospheric warmings (SSWs) in the Northern Hemisphere (NH). We initialized the model across different winter phases and at various lead times before SSW onsets to assess its predictions. Our results show that GraphCast predicts the stratospheric polar vortex strength up to 2 weeks, although errors increase rapidly with height and exhibit larger winter‐to‐winter variability than the physics‐based model. More importantly, GraphCast fails to predict any SSW events with a 1‐week lead time. The primary reason is its inability to accurately simulate wave‐mean flow interactions in the stratosphere, as it fails to satisfy the Transformed Eulerian Mean momentum budget, despite simulating the wave activity flux similar to reanalysis. Furthermore, GraphCast cannot accurately predict the downward propagation of stratospheric anomalies, resulting in larger surface errors after SSW events compared to normal winters. Our study highlights some issues with GraphCast in representing stratosphere‐troposphere coupling and obeying known dynamical relationships. We emphasize the need for a comprehensive evaluation of AIWFMs throughout the atmospheric column and across various dynamical interactions, which is essential to improve these models and extend their applications.
Zheng Wu (Thu,) studied this question.