Assessing Subseasonal Predictions Above Lower Troposphere of GraphCast

Advancements in computing and data availability have spurred the development of AI weather emulators. GraphCast, one of the best current existing AI emulators developed by Google DeepMind, outperforms state-of-the-art numerical forecasting models in global-mean variables at the surface and lower troposphere. However, their skill in predicting upper atmospheric states and representing dynamical relationships across various physical fields remains unclear. In this study, we evaluate GraphCast’s predictions of physical fields in the upper troposphere and stratosphere and its ability to capture the stratosphere-troposphere coupling, focusing on sudden stratospheric warmings (SSWs), one type of extreme circulation event in the stratosphere, which are considered an important source of subseasonal to seasonal predictability to surface weather. We generate predictions for the recent SSW events using GraphCast initialized with ERA5 reanalysis data. The results show that GraphCast’s prediction skill diminishes rapidly with altitude, failing to reproduce SSW events. To understand this failure dynamically, we compute the Eliassen-Palm flux (EP-flux) and analyze wave patterns in the lower troposphere to assess interactions between the upward propagating planetary waves and the zonal-mean flow in the stratosphere. Our study highlights the potential issues of the GraphCast model and emphasizes the need for a comprehensive evaluation of AI emulators throughout the atmospheric column and dynamical interactions across physical fields, which will provide feedback to improve the AI emulators and extend their applications to advance our dynamical understanding of the weather and climate systems.