Can Artificial Intelligence-Based Weather Prediction Models Simulate the Butterfly Effect?

We investigate error growth from small-amplitude initial condition perturbations, simulated with a recent artificial intelligence-based weather prediction model. From past simulations with standard physically-based numerical models as well as from theoretical considerations it is expected that such small-amplitude initial condition perturbations would grow very fast initially. This fast growth then sets a fixed and fundamental limit to the predictability of weather, a phenomenon known as the butterfly effect. We find however, that the AI-based model completely fails to reproduce the rapid initial growth rates and hence would incorrectly suggest an unlimited predictability of the atmosphere. In contrast, if the initial perturbations are large and comparable to current uncertainties in the estimation of the initial state, the AI-based model basically agrees with physically-based simulations, although some deficits are still present. Even if perfect observations and models were available, the time interval for which weather forecasts can be accurate is limited. This limit is related to fundamental physical characteristics of the earth's atmosphere, which make small errors grow very fast and spread out, a feature known as the butterfly effect. In this article, we test if an artificial intelligence-based weather prediction model is able to reproduce this butterfly effect. Therefore, we computed several weather forecasts that differed only very slightly in their starting conditions. We find, that in contrast to standard weather forecasting models, the initial difference grow only slowly in the AI-based model and there is no indication of a butterfly effect at all. This provides an example of how machine learning models can fail to reproduce a fundamental physical principle, even though they can accurately mimic many observed behaviors. Current artificial-intelligence-based models cannot simulate the butterfly effect and incorrectly suggest unlimited atmospheric predictabilityTheir error growth rate and structure remain similar to synoptic-scale error growth regardless of the amplitude of the initial perturbationSynoptic-scale error growth from current levels of initial condition uncertainty appears mostly realistic, except for a short initial decay