Can AI‐Based Weather Prediction Models Simulate the Butterfly Effect? The Role of Architecture and Implementation

Abstract

Simulations of numerical weather prediction models indicate that the atmosphere possesses an intrinsic limit of predictability. Initial perturbations of tiny amplitude grow quickly in areas of convection and latent heat release, then spread out and move upscale, eventually affecting even the largest planetary scales after about 2 weeks. In this study, we investigate the ability of several state‐of‐the‐art AI‐based weather prediction models to reproduce this phenomenon, which is sometimes referred to as the “butterfly effect.” The AI results are compared to those of a conventional, physics‐based, weather prediction model run at various resolutions. Evaluating six key characteristics of this butterfly effect, we find that the behavior of the AI models can be separated into two groups. The first group did not reproduce any of the key characteristics, while the second group did reproduce some, in particular fast initial uncertainty growth and indication of an intrinsic limit. However, the behavior was physically inconsistent and based on the production of numerical noise, and for some models even dependent on whether the experiments were carried out on a CPU or a GPU. It seems likely that the inability of AI models to simulate the butterfly effect results from limitations in the analysis data used for training, since their size, design and architecture turned out to be largely irrelevant.