Insights on long-term stability in auto-regressive AI weather models
Please login to view abstract download link
AI weather models and weather-based foundation models have demonstrated impressive skills in short- to medium-range forecasts. While most weather models become unstable on longer time scales, a wide variety of AI climate emulators have been proposed, raising questions about the fundamental differences between these approaches. In this work, we compare state-of-the-art models when producing rollouts on annual time scales. We quantify and characterize different types of instability: smoothing, visual artifacts, drift, and loss of seasonality. This analysis highlights the previously unreported stability of the Aurora foundation model and the Earth System Foundation Model (ESFM) for rollouts longer than 35 years. We also show that stability emerges when models are able to constrain small-scale noise and the temporal dynamics of the physical processes they predict is commensurate with their accuracy. These results open perspectives to apply weather models on long time scales and provide insights to design stable auto-regressive models for other types of complex time-dependent systems.
