Diagnosing Forecast Error Propagation and Large‐Scale Dynamics of Weather Extremes With an AI Weather Model

Abstract Artificial intelligence (AI) weather models can generate fast and accurate weather forecasts, yet they still struggle to predict regional extremes. Exploring error propagation pathways and associated large‐scale dynamics is essential for understanding extreme weather drivers and AI model performance. We propose a true‐state constraint (TSC) method that can diagnose error propagation by adjusting forecasts in an upstream region toward reanalysis data and tracking the responses elsewhere. Applied to GraphCast, a leading global AI model, this approach effectively captures critical, misrepresented atmospheric processes that are responsible for the development of two extreme events. The model’s responses to the adjustments are consistent with the behaviors in dynamical nudging experiments. Furthermore, we propose a climatology constraint method that, when combined with the TSC method, can quantify the remote impacts of evolving atmospheric circulation anomalies. This work underscores the potential of AI weather models as powerful research tools for mechanisms driving weather extremes.