Weather Forecasting AI
Weather forecasting AI is the use of machine-learning models, trained on decades of past atmospheric states, to predict future weather directly, in place of or alongside the physics-based simulation that has underpinned forecasting for decades. Its significance is speed: a trained model can produce a global multi-day forecast in about a minute on a single accelerator, where a conventional forecast requires a supercomputer, and in several published comparisons the learned models match or exceed the accuracy of leading physical systems on standard scores.
Numerical weather prediction (NWP), the contrast case, solves the equations of fluid dynamics and thermodynamics on a global grid, starting from an estimate of the current state of the atmosphere. The idea goes back to Lewis Fry Richardson's 1922 book on forecasting by calculation, and operational NWP became practical with electronic computers in the 1950s. A modern system has two expensive parts: data assimilation, which blends millions of observations from satellites, aircraft, balloons and stations with a short prior forecast to produce an initial state, and the forward model, which integrates that state in time while representing unresolved processes such as clouds and convection by parameterizations. Centers such as the European Centre for Medium-Range Weather Forecasts (ECMWF) run this at high resolution and also run ensembles of perturbed forecasts to quantify uncertainty. The physics gives NWP interpretability and a sound basis for unusual situations, at heavy computational cost.
FourCastNet, from NVIDIA and collaborators (Pathak et al., 2022), was an early demonstration that a neural network could forecast global weather at 0.25 degree resolution. It used Adaptive Fourier Neural Operators, an architecture that performs global mixing in the frequency domain, and was trained on ECMWF's ERA5 reanalysis, the consistent reconstruction of past weather covering decades. It was competitive with NWP for some variables, such as wind and near-surface fields, but weaker for others such as precipitation, and it showed that data-driven forecasts were feasible.
Pangu-Weather, by Huawei researchers (Bi et al., Nature, 2023), used a three-dimensional Earth-specific transformer and reported deterministic forecasts that beat ECMWF's operational high-resolution system on several metrics. It trained separate models for different lead times, from one hour to a week, to limit error accumulation, and it also tested tropical cyclone track forecasting. It was also a prominent example of an industrial lab, not a weather service, delivering a leading forecasting model.
GraphCast, from Google DeepMind (Lam et al., 2023, published in Science), represents the globe as a graph and applies a graph neural network to step forward six hours at a time, producing a ten-day forecast at 0.25 degree resolution. It was trained on ERA5 data from 1979 to 2017, and the authors reported that it outperformed ECMWF's high-resolution deterministic system, HRES, on about 90 percent of the 1,380 verification targets they evaluated, while generating a forecast in under a minute on one Google TPU. It also showed useful skill in predicting cyclone tracks and atmospheric rivers. Later work included DeepMind's GenCast, a probabilistic ensemble model, and ECMWF developed its own machine-learning system, the Artificial Intelligence Forecasting System, and began operating it alongside its physics-based forecasts.
The caveats are substantial, and a frequent misconception is that these models replace physics-based forecasting. Autoregressive deterministic models tend to smooth fine detail as lead time grows because averaging minimizes error; they can struggle with rare extremes underrepresented in the training record and with a climate that is shifting beyond it; and they usually start from an initial state produced by NWP data assimilation, so they depend on the system they compete with. Their value for ensembles is large, since fast models let centers run far more members to quantify risk, but verifying performance on high-impact events, not just average scores, is the central question.
Read weather forecasting AI as a learned shortcut to what physics-based simulation computes the slow way: it can match the leading systems on standard scores in a fraction of the time, which makes larger ensembles and new users possible, while its dependence on assimilated initial conditions and its softer treatment of extremes keep physical models in the loop.