1) Training: the model learns from decades of ERA5 reanalysis (about 40 years of hourly global data), mapping the atmospheric state at time t to the state at t+dt (typically dt = 6 h). 2) Representation: an encoder moves lat/lon grid fields into an internal representation — GraphCast uses a graph multi-mesh (GNN), Pangu-Weather a 3D transformer, and FourCastNet an adaptive Fourier neural operator (AFNO). 3) Increment prediction: the neural core predicts the state change and a decoder maps it back to the grid. 4) Autoregressive rollout: the model output is fed back as input to build a multi-step forecast (up to 10-15 days). 5) Ensemble: generative models such as GenCast, instead of a single forecast, sample many scenarios from a conditional distribution via a diffusion process, producing a probabilistic forecast.
Traditional numerical weather prediction (NWP) is highly accurate but requires hours of supercomputer time and is expensive for generating large ensembles. AI models address the cost and inference-time problem: they learn atmospheric dynamics from historical data and produce forecasts of comparable or higher accuracy orders of magnitude faster and cheaper, enabling low-cost ensemble forecasting and rapid response to extreme events.
Maps the global atmospheric state from a lat/lon grid into the model's internal representation (graph multi-mesh, transformer tokens, or Fourier modes).
Official
Propagates the atmospheric state across space and time. GraphCast uses a graph network on a multi-mesh, Pangu-Weather a 3D transformer (3DEST), and FourCastNet an adaptive Fourier neural operator (AFNO).
Official
Maps the internal representation back to the lat/lon grid and predicts the state increment (residual) for the next time step.
Official
Iteratively feeds the model output back as input to build a multi-step forecast (e.g. every 6 h up to 10-15 days).
In generative models (GenCast) it samples many scenarios from a conditional distribution via a diffusion process, producing a probabilistic ensemble.
Official
Feeding the model's own predictions back as input causes error growth and field smoothing at longer forecast horizons.
Models optimising mean squared error (GraphCast, Pangu) produce blurred, averaged forecasts that underestimate extremes.
Models are trained on and initialised from ERA5 reanalysis; the lack of native raw-observation assimilation limits purely operational use.
The first global data-driven model at NWP resolution (0.25 degree), based on an adaptive Fourier neural operator (AFNO); showed that neural networks can forecast global weather orders of magnitude faster than NWP.
A graph neural network (GNN) model forecasting up to 10 days in under a minute on a TPU v4; more accurate than ECMWF's operational HRES on 90% of 1380 tested variables. Published in Science.
A 3D transformer (3DEST) model trained on 43 years of ERA5; presented as the first AI method to outperform traditional NWP from one hour to one week. Published in Nature.
A probabilistic diffusion model generating an ensemble of 50+ forecasts up to 15 days; more accurate than ECMWF's ENS on 97.2% of targets. Shifted the paradigm from deterministic to generative/probabilistic forecasting. Published in Nature.
Microsoft publishes Aurora, a foundation model of the Earth system (Nature 2025); ECMWF operationalises its machine-learning AIFS system; Google integrates the WeatherNext line into Search, Maps and Gemini and offers it via BigQuery, Earth Engine and Google Maps Platform.
Global grid cell size.
Maximum forecast horizon.
Interval between successive autoregressive steps.
Number of sampled scenarios in probabilistic models.
Atmospheric and surface variables across multiple pressure levels.
Deterministic models (GraphCast, Pangu) are a dense forward pass; generative models (GenCast) add conditional diffusion sampling for ensembles.
Spatial computation within a single step is fully parallel on GPU/TPU, but the autoregressive rollout is sequential in time — each step depends on the previous step's output.
Training and inference of weather models (FourCastNet, Aurora) rely on large NVIDIA GPUs (A100/H100).
GraphCast and GenCast run on Google TPUs; GraphCast produces a 10-day forecast in under a minute on TPU v4, and GenCast a 15-day forecast in about 8 minutes on TPU v5.
