Butterfly Effect and Predictability of Global AI Weather Models in Unusual Tropical Cyclone Track Forecast

In the past few years, the rapid breakthroughs of artificial intelligence (AI)-driven weather prediction have opened up a new direction to solve the weather prediction problems. Yet, recent publications have raised doubts about the ability of global AI weather models to capture the butterfly effect in atmospheric systems, a limitation that implies these data-driven models may tend to underestimate uncertainties in ensemble predictions. This conclusion, however, remains a subject of controversy within the research community. A counterargument is that claiming the absence of butterfly effect in AI weather models may erroneously imply infinite predictability of atmospheric circulation in such models.

Studying the butterfly effect in AI weather models is not only critical for evaluating the ability of AI weather prediction techniques, but also provides a theoretical basis for AI-based ensemble forecasting and weather modications. In this presentation, I will share our latest findings about the existence of the butterfly effect in current global AI weather models. Specifically, our results show that tropical cyclone track forecasts generated by Pangu-Weather, one of the state-of-the-art global AI weather models, can be sensitive to initial perturbations under certain conditions. This presentation will particularly focus on a case study of Super Typhoon Khanun (2023), which is characterized by its unusual zigzagging track. Based on a series of probabilistic prediction experiments, I will demonstrate the impacts of initial perturbations on Pangu-Weather’s forecast results of Typhoon Khanun. Then, the differences in the butterfly effect between numerical weather prediction (NWP) models and data-driven AI weather models will be presented. Finally, I will discuss the implications of our findings for ensemble AI weather forecasting and potential interventions in tropical cyclones.