Subseasonal Prediction of the Wintertime North Pacific Blocking in AI‐Based Weather Models

Abstract

Recent studies have demonstrated that AI‐based weather models exhibit reasonably encoded atmospheric dynamics and skill of short‐term forecasting comparable to physics‐based models. However, it remains unclear whether AI‐based weather models can simulate system‐oriented dynamics and provide skillful predictions across subseasonal timescales. Here, we assess the simulation of the North Pacific blocking and its subseasonal prediction skill in three weather models with a hierarchical implementation of AI: ECMWF IFS (full‐physics), NeuralGCM (hybrid), and Pangu‐Weather (data‐driven). While all models reproduce blocking frequency well at 1–2 week lead times, biases diverge when the lead time increases to 3–4 weeks. ECMWF IFS overpredicts blocking frequency, related to overestimations of both short‐lived blockings (5–6 days) and long‐lived blockings (>9 days). Pangu‐Weather tends to underpredict blocking frequency, largely due to an underestimation of blockings with 7–8 day durations. NeuralGCM has a dipolar structure of blocking frequency bias, exhibiting mixed characteristics between the ECMWF IFS and Pangu‐Weather models. Despite these biases, all three models successfully reproduced blocking evolutions through 3–4 weeks of simulation. Remarkably, they all reasonably captured the moist processes of blocking dynamics despite different model hierarchies. The subseasonal prediction remains skillful for up to 2 weeks across all models. In ECMWF IFS, this skill can be enhanced by leveraging interannual and subseasonal variability originating from the tropical Pacific. This enhancement of skill is also qualitatively observed in NeuralGCM and Pangu‐Weather, suggesting that AI‐based weather models have the potential to deliver physically reliable predictions of weather extremes across subseasonal timescales.