Machine-learning systems that predict global weather from past data, rivalling traditional supercomputer forecasts.
For the curious
Weather forecasts used to need enormous supercomputers crunching physics for hours. Now AI has watched decades of weather and learned the patterns — and it can guess tomorrow's weather almost as well, in less than a minute!
For students
Traditional forecasting solves fluid-physics equations on a global grid. GraphCast instead learns from ~40 years of reanalysis data, predicting temperature, wind, humidity and cyclone tracks for 10 days in under a minute. In comparisons it outscored the leading conventional system on most standard measures — while using far less energy to run.
Going deeper
GraphCast uses graph neural networks over the globe; Huawei's Pangu-Weather and ECMWF's AIFS followed. ECMWF now runs AI forecasts operationally alongside physics models. Limits remain: training on the past makes unprecedented extremes hard, small-scale physics (thunderstorms) is blurred, and trust requires hybrid physics-AI systems.
For experts
The frontier moved fast: neural weather models now drive ensemble systems and medium-range hurricane intensity gains, and generative downscaling (e.g., NeuralGCM hybrids) targets kilometre-scale detail. For a citizen network like Weather Pulse, the near-term opportunity is AI post-processing that fuses official grids with ground-truth community observations.
Common misconceptions
- AI hasn't replaced meteorologists — it is a new instrument in the orchestra, run beside physics models and human expertise.
Sources
- Google DeepMind
- ECMWF
- Wikipedia