Neural Weather Prediction

Numerical weather prediction integrates the primitive equations on a grid, and running one deterministic global forecast at operational resolution takes on the order of an hour on hundreds of supercomputer nodes. GraphCast produces a 10-day global forecast at 0.25 degrees, covering hundreds of variable-level combinations, in under a minute on a single TPU v4 device, and beat ECMWF's operational deterministic system HRES on 90 percent of 1,380 verification targets (Lam et al., Science 382(6677), 1416–1421, 2023, arXiv:2212.12794).

That result is what turned learned forecasting from a curiosity into the mainstream of operational meteorology inside two years.

What the model actually learns

The training data is not observations. It is ERA5, ECMWF's reanalysis: a physically consistent gridded reconstruction of the atmosphere from 1979 onward, produced by assimilating historical observations into a numerical model. This matters more than any architectural detail. The learned model is trained to reproduce the output of a physics model that has been fitted to observations, which is why it inherits both the reanalysis's quality and its biases.

The task is an autoregressive one-step map. Given the atmospheric state at \(t\) and \(t - 6\text{h}\), predict the state at \(t + 6\text{h}\); roll it forward to reach longer leads. GraphCast encodes the lat-lon grid onto a multi-mesh icosahedral graph, runs message passing on that mesh, and decodes back, so that a single message-passing step propagates information across long distances. It is trained with a curriculum that increases the number of autoregressive steps, which forces the model to remain stable when fed its own output.

Other architectures reach comparable skill differently. Pangu-Weather uses a 3D Earth-specific transformer with separate models for different lead times to limit error accumulation (Bi et al., Nature 619, 533–538, 2023). FourCastNet used adaptive Fourier neural operators and was the first to show the approach was viable at 0.25 degrees (Pathak et al., 2022, arXiv:2202.11214).

Deterministic skill hides a real problem

A model trained with a mean-squared-error objective learns the conditional mean. The conditional mean of the atmosphere is smooth, so single-forecast learned models produce blurry fields: they score well on RMSE and understate gradients, fronts, and extremes. That is a genuine defect for anyone forecasting a storm rather than a monthly average.

The fix was to make the model generative. GenCast is a conditional diffusion model producing an ensemble of 15-day forecasts at 0.25 degrees, generating each member in about 8 minutes on a single TPU v5. It exceeded ECMWF's operational ensemble ENS on 97.2 percent of 1,320 evaluation targets, and improved prediction of extremes, tropical cyclone tracks and wind-power-relevant quantities (Price et al., Nature, 2025, arXiv:2312.15796).

Beating an ensemble is the harder result. Ensembles are how operational forecasting expresses uncertainty, and matching one requires calibrated spread, not just a good mean.

The dependency that is easy to miss

Learned forecasters are not standalone. They are initialised from an analysis produced by a conventional data assimilation system, which is itself the expensive part of the operational pipeline: ingesting satellite radiances, radiosondes, aircraft and buoy data into a consistent state estimate. Remove ECMWF's assimilation and GraphCast has nothing to start from.

This is why the honest framing is that machine learning has replaced the forecast model, not the forecast system. ECMWF's own AIFS, which became an operational product in 2025, sits inside the existing assimilation infrastructure for exactly this reason.

When it breaks

Out-of-distribution climate. The model interpolates within the training distribution. A warming climate moves the distribution, and there is no physical constraint forcing the model to extrapolate correctly. Conservation of mass and energy are not enforced; they are approximately learned.

Blurring at long leads. Even generative models lose small-scale structure as lead time grows, and RMSE-optimised ones lose it immediately. Scores that reward smoothness reward the wrong thing for extreme-event forecasting.

Verification is against reanalysis, not reality. Most headline comparisons score against ERA5. That favours a model trained on ERA5 over a physics model that was not, and scoring against station observations narrows the gap. This is a live methodological dispute in the field, not a settled point.

Rare events are rare in training too. A category-5 landfall or a record heatwave has few analogues in 45 years of reanalysis. Physics-based models get these partly right because the equations do not care how often something has happened before.