Machine-learned weather forecasting

Assembled from 23 blocks · 1 evidence · 28 related

1. Story
2. Global weather prediction by learned models trained on decades of reanalysis data, rather than by numerically integrating the equations of atmospheric motion — and the clearest case of a machine-learning method being taken into operational service by a public scientific institution. DeepMind's GenCast, published in Nature in December 2024, is reported to have beaten the European Centre for Medium-Range Weather Forecasts' ensemble on 97.2% of 1,320 evaluation targets while producing a fifteen-day global forecast in about eight minutes on a single accelerator, against hours on a supercomputer. The institutional step came on 25 February 2025, when ECMWF made its own Artificial Intelligence Forecasting System operational, reporting gains of up to 20% on measures including tropical cyclone track error. This is a genuinely different kind of evidence from a benchmark score: an intergovernmental forecasting agency accepted operational responsibility for the output, and it runs alongside — not instead of — the physics-based system that generates the data it was trained on.
3. Knowledge
4. Deep learning
5. Machine-learned weather forecasting
6. European Centre for Medium-Range Weather Forecasts
7. Connections
8. Deep learning
9. European Centre for Medium-Range Weather Forecasts
10. GenCast is published with a like-for-like comparison against an operational ensemble
11. ECMWF makes a machine-learned forecast model operational
12. The AI accelerator
13. ImageNet
14. The transformer architecture
15. AlphaFold
16. Machine-learned weather forecasting
17. Machine learning in medical imaging
18. A deep convolutional network wins the ImageNet challenge
19. AlphaGo defeats Lee Sedol in Seoul
20. Machine-learned weather forecasting
21. ECMWF makes a machine-learned forecast model operational
22. Evidence
23. ECMWF's Artificial Intelligence Forecasting System became operational on 25 February 2025, outperforming the physics-based system on many measures including tropical cyclone tracks with gains of up to 20%; GenCast, published in Nature in December 2024, is reported to have beaten the ECMWF ensemble on 97.2% of 1,320 evaluation targets and to produce a fifteen-day global forecast in about eight minutes on a single accelerator. V55 verification basis: neither ecmwf.int nor nature.com was fetched. The operational date and the up-to-20% figure were carried in search results attributing them to ECMWF's own news item; the 97.2%, 1,320-target and eight-minute figures were carried with attribution to the Nature paper and to reporting of it.