by ECMWFGB
Make weather forecasting faster, cheaper, and more accessible through AI while maintaining competitive accuracy.
Operational since Feb 2025. v2 (May 2026): wave and snow forecasts. Daily public outputs.
Updated 12 May 2026Checked 11 Oct0 updates this week
GNN treats 3D atmosphere as a graph where each grid point connects to neighbors. Learns how pressure, temperature, moisture influence each other across space.
Encoder-decoder with sliding window reads current state and outputs next 6-hour step. Repeated application predicts further ahead.
Learns from 40+ years of assimilated observations and verified forecasts—vast library of real weather evolutions.
Deterministic single best estimate; ensemble shows uncertainty ranges. Updated daily, outputs public within minutes.
| Spec | AIFS |
|---|---|
| Speed vs classical model | 10x fasterR (reported) |
| Energy vs classical model | 1,000x lowerR (reported) |
| Wave forecast variables (v2) | 11 variablesR (reported) |
| Update frequency | Every 12 hoursR (reported) |
R reported by the company
ECMWF
Weather forecasting, AI, neural networks, meteorology
European Centre for Medium-Range Weather Forecasts. Independent intergovernmental research institute based in Reading, UK. Operates AIFS, an AI weather forecasting system using graph neural networks that generates forecasts 10x faster and at 1,000x lower energy cost than classical models.
