🌍 Earth Science¶

🧪 ICML2026 · 2 paper notes

Scaling Laws of Global Weather Models: This paper presents the first cross-model scaling law analysis of five mainstream data-driven weather models (Aurora, AIFS, Pangu, GraphCast, SFNO) under a unified training/evaluation protocol. It finds that weather models favor "width over depth," compute budgets should prioritize more training data over larger models, and scaling behaviors vary significantly across meteorological variables—distinct patterns from NLP/Vision scaling laws.
(Sparse) Attention to the Details: Preserving Spectral Fidelity in ML-based Weather Forecasting Models: MOSAIC addresses two types of spectral degradation in ML weather forecasting models (spectral damping from deterministic averaging and high-frequency aliasing from coarsened latent spaces) by combining "probabilistic perturbation + mesh-aligned block-sparse attention on HEALPix spherical grids." With only 214M parameters at 1.5° resolution, it matches or exceeds models with 6× higher resolution, generating a 24-member 10-day forecast in 12 seconds on a single H100.