AI weather models
Neural-network forecast models trained on decades of weather data now run operationally at ECMWF and beat physics-based models on many measures.
Open in the interactive tree →Google DeepMind's GraphCast (Science, November 2023) beat the best traditional global model, ECMWF's HRES, on more than 90% of 1,380 test variables and lead times, at a tiny fraction of the compute. A 10-day forecast that needs hours on a supercomputer takes under a minute on a single TPU.
As of October 2026
ECMWF began operational forecasts with its AI system AIFS Single on 25 February 2025 (gains of up to 20% for tropical-cyclone tracks) and with the ensemble AIFS ENS on 1 July 2025, which it says beats its physics-based ensemble on many measures by up to 20% while using about 1,000 times less energy, though at 31 km instead of 9 km resolution. Both models were upgraded to version 2 on 12 May 2026, adding the first operational data-driven wave forecasts, snow cover and a 10 hPa stratosphere level. Google retired its WeatherNext Graph datasets on 29 July 2026 in favour of WeatherNext 3.
Open steps
- Weeks-2-to-6 forecast skill High AI leverageAI weather models now match physics models at weeks 2-6; the open step is a clear, reliable lead that bridges weather and seasonal forecasts.
- Learned convection closures High AI leverageTrain neural closure models for thunderstorm parameterization from cloud-resolving data, to improve coarse-grid tropical forecasts.
- Single-model ensemble-like uncertainty Medium AI leverageTrain networks to generate ensemble-like uncertainty from a single large model, reducing compute while keeping error bars.
Where AI could help
High AI leverage. Training and checking neural forecast models is already operational work for AI; the limits are resolution, observations and unprecedented extremes.
- Train higher-resolution and longer-range ensembles at a fraction of physics-model cost
- Learn from raw observations directly, reducing reliance on physics-based data assimilation
- Add new outputs such as waves, snow cover and ocean variables
- Calibrate rare extremes and tropical cyclones where AI skill is least certain
Shown so far
- In July 2025, ECMWF put its ensemble AI model AIFS ENS into operation: 51 members, gains of up to 20% over its physics-based ensemble on many measures, about 1,000 times less energy, at 31 km instead of 9 km resolution. source
- In summer 2025, India's agriculture ministry sent AI-based monsoon-onset forecasts (NeuralGCM blended with ECMWF's AIFS) by SMS to 38 million farmers, up to a month ahead. source
Prerequisites
- Numerical weather prediction1950
- Weather Radar1957Radar rainfall archives train AI nowcasting and forecast models
- Earth observation by satellite1960
- Deep Learning2012