Current AI weather models cannot simulate the butterfly effect.
10.01.2024
Even if we had perfect observations and models at our disposal, the time interval for which precise weather predictions are possible would remain limited. This limitation is related to fundamental physical properties of the Earth's atmosphere, causing small errors to grow and spread very quickly, a phenomenon known as the butterfly effect. In a recent study, we investigated whether a weather model based on artificial intelligence is able to reproduce this butterfly effect. We found that initial uncertainties in the AI-based model grow very slowly, showing no signs of a butterfly effect (see figure). The AI model would significantly overestimate the predictability of the atmosphere. This example illustrates that machine learning may struggle to reproduce basic physical principles, even though it can mimic observed behaviors precisely.
EOS magazine article:
https://eos.org/research-spotlights/ai-meets-its-match-the-butterfly-effect
Paper:
Selz, T., & Craig, G. C. (2023). Can artificial intelligence-based weather prediction models simulate the butterfly effect? Geophysical Research Letters, 50, e2023GL105747. https://doi.org/10.1029/2023GL105747