AI Learned to Forecast Weather

THE BREAK

For seventy years, a weather forecast meant pushing the atmosphere forward by solving the equations of fluid dynamics on a grid, the arrangement ENIAC first made work in 1950. Machine-learned weather models do something else.

Given the state of the global atmosphere, they predict the next state directly, from examples of how states have followed one another in the past.

The difference is larger than a faster solver. When systems such as GraphCast and ECMWF's AIFS make a forecast, they do not step through fluid-dynamics code at all. They approximate the transition from one state to the next. The atmosphere became a learned dynamical system, which would have surprised Bjerknes and possibly annoyed him.

Weather is an unusually good place to test the idea because the answer arrives quickly. A ten-day forecast is marked ten days later. The field has decades of standardised verification, forecast centres that compare models every day, and an atmosphere with no interest in helping a benchmark look good. If a learned system works here, it is hard to dismiss as a laboratory trick.

LEARNED DYNAMICS

GraphCast learned the six-hour step

GraphCast, published by Google DeepMind in 2023, takes two global weather states six hours apart and predicts the state six hours later. That prediction is fed back in, over and over, to reach ten days. The system never learns a ten-day answer in one jump. It learns one short step and repeats it forty times.

It works on a graph covering the globe at roughly 0.25-degree spacing and predicts hundreds of atmospheric and surface variables. In the evaluation published in Science, it beat ECMWF's high-resolution deterministic forecast on more than 90% of 1,380 combinations of variable and lead time, and produced a ten-day global forecast in under a minute on specialised hardware. ENIAC's ratio of computing time to forecast time, one to one in 1950, had become better than one to fourteen thousand.

The speed matters because forecasting is not one calculation. Centres run several cycles a day, generate many products, rerun models during development and increasingly need ensembles rather than single runs. A model that is a little cheaper is convenient. A model that is orders of magnitude cheaper changes which forecasting strategies anyone can afford.

The repeated step also exposes the main danger. A small error at one six-hour step becomes part of the input to the next. A learned weather model has to stay stable while feeding on its own predictions, not just while predicting from perfect historical states. A model can be accurate for one step and unusable after forty.

GENCAST

A useful future is a distribution, not a line

A single best forecast is only part of the job. Volume 3 explained why forecast centres run ensembles. GenCast, published in Nature in 2024, brought that logic inside a learned generative model rather than bolting probability on afterwards.

GenCast generates random but physically coherent global weather trajectories at 0.25-degree resolution out to fifteen days, and its published evaluation used ensembles of 50 or more members. Against ECMWF's ensemble, it showed better skill on 97.2% of 1,320 evaluated targets, and improved on several tests involving extremes, cyclone tracks and conditions relevant to wind power.

The important output is not fifty slightly different maps. It is a sample from the joint distribution of plausible atmospheres. Each member keeps its relationships intact through time and space, so a member describing a particular storm track also carries the winds, temperatures and pressures that go with that track. Probabilities assembled from incoherent pieces are much easier to compute and much less useful.

GenCast also shows why speed matters more for ensembles than for single forecasts. Google reported that one fifteen-day member could be generated in minutes on a single TPU, with members running in parallel. When honesty about uncertainty requires many futures, a cheap step turns probability from an expensive luxury into a routine product.

AIFS IN OPERATIONS

Then the research model acquired a duty roster

The more consequential change happened at ECMWF. AIFS Single became operational on 25 February 2025. AIFS ENS followed on 1 July 2025 with 51 ensemble members. On 12 May 2026 both moved to version 2, alongside an upgrade of ECMWF's physics-based Integrated Forecasting System.

Operational means more than appearing on a chart. The forecasts run on schedule, four times a day out to fifteen days, and are monitored, distributed, archived, supported and expected every single day, including the day the model has a bad idea. ECMWF also makes real-time output and model resources openly available.

ECMWF reported that its 2025 AI ensemble produced forecasts more than ten times faster than the physics-based ensemble, using about 1,000 times less energy for the forecast calculation, with skill gains for several variables. It also ran at coarser resolution than the physics-based ensemble and did not replace the coupled Earth-system capabilities of the IFS.

The outcome is less dramatic than replacement and more important. ECMWF runs both. The learned model is not a protest against physics. It is another operational component, with a different cost, resolution, skill profile and way of failing. Meteorology responded to a supposed war between paradigms by putting both of them on the rota.

THE STACK

The learned forecast still begins inside a physical system

The leading learned weather models were trained on ERA5, the reanalysis described in Volume 3, rather than on raw observations. What matters here is that GraphCast and its relatives learned from a reconstruction of the atmosphere that had already passed through physics and data assimilation. They are students of a model's opinion of the past, and a very well-informed opinion at that.

Day-to-day operation follows the same pattern. AIFS starts each forecast from ECMWF's analysis, produced by a physics-based assimilation system. The learned step is fast because another part of the stack has already done the hard work of turning scattered observations into a coherent starting state.

This is why learned and physical models remain partners. Physical models represent conservation laws and coupled processes explicitly, can be extended with new knowledge, and supply the machinery that builds starting states. Learned models can reproduce much of the atmosphere's evolution at extraordinary speed and skill once that state exists.

For planetary intelligence the lesson is architectural. The best system may not be the model that wins the philosophical argument. It may be the one that gives each component the job it does best, notices when the components disagree, and keeps the uncertainty attached when responsibility passes from one to the next.

Weather forecasting did not become less physical when AI arrived. It became a stack in which some of the physics moved out of the forecast step and remained indispensable everywhere around it.