A new study published on arXiv analyzes four probabilistic machine learning weather prediction models: NeuralGCM-ENS, FourCastNet 3, AIFS-ENS, and GenCast. The research compares their kinetic energy (KE) spectra and transfer mechanisms against the physics-based IFS ENS model. While NeuralGCM-ENS shows promise in reproducing KE transfer, the other machine learning models struggle with accurate upscale transfer, with AIFS-ENS and GenCast accumulating KE at high wavenumbers due to uncorrelated stochastic perturbations. All models exhibit upscale error growth, but they fail to replicate the rapid initial spread of ensemble forecasts at small scales characteristic of the butterfly effect, indicating potential misrepresentation of kinetic energy transfer despite producing skillful forecasts. AI
IMPACT Highlights potential limitations in machine learning weather models' ability to accurately represent atmospheric physics, despite skillful forecasting.
RANK_REASON Research paper published on arXiv detailing findings about machine learning weather prediction models. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
