AI Learned to Forecast Weather

THE BREAK

Numerical weather prediction spent decades advancing the atmosphere by solving discretized physical equations. Machine-learned weather models learned a different operation: given a global atmospheric state, predict the next state directly from examples of how states evolved in the past.

The distinction is larger than a faster solver. During inference, systems such as GraphCast and ECMWF's AIFS do not step through the same fluid-dynamics code as a traditional forecast model. They approximate the state transition itself. The atmosphere became a learned dynamical system.

Weather is unusually useful for testing this idea because the target arrives quickly. A ten-day forecast is judged ten days later. The field has decades of standardized verification, operational centres that compare models continuously, and an atmosphere with little interest in helping a benchmark look good. If a learned system works here, the result is difficult 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 into the model repeatedly to produce a forecast out to ten days. The system does not learn a ten-day answer in one jump. It learns a short transition and composes it many times.

The model operates on a graph representation of the globe at roughly 0.25-degree spacing and predicts hundreds of atmospheric and surface variables. In the Science evaluation, GraphCast outperformed ECMWF's high-resolution deterministic system on more than 90% of 1,380 tested variable and lead-time combinations and produced a ten-day global forecast in under a minute on specialized hardware.

The speed matters because forecasting is not one calculation. Centres run several cycles every day, generate many products, rerun models during development and increasingly need ensembles rather than a single future. A model that is merely a little cheaper is convenient. A model that changes the cost by orders of magnitude changes which forecasting strategies are affordable.

The recursive design also exposes the main technical danger. A small error at one six-hour step becomes part of the input to the next. Learned weather models therefore have to remain stable when operating on their own previous predictions, not only when 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

Deterministic skill is only part of operational weather forecasting. The atmosphere admits several plausible futures from an uncertain starting state, so forecast centres run ensembles. GenCast, published in Nature in 2024, brought that logic into a learned generative model rather than treating probability as an afterthought.

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

The important output is not fifty slightly different maps. It is a sampled joint distribution of plausible atmospheric evolutions. That preserves relationships through time and space, so a member describing a storm track also carries the winds, temperatures and pressures associated with that same trajectory. Probabilities built from incoherent pieces are much easier to compute and much less useful.

GenCast also illustrates why fast inference matters more for ensembles than for deterministic forecasts. Google reported that a fifteen-day member could be generated in minutes on a TPU and members can run in parallel. When uncertainty requires many futures, a cheap transition model turns probability from an expensive luxury into a routine product.

AIFS IN OPERATIONS

Then the research model acquired a duty roster

The more consequential transition 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 systems moved to version 2 alongside an upgrade of ECMWF's physics-based Integrated Forecasting System.

Operational means more than a model appears on a chart. The forecasts run on schedule, are monitored, distributed to users, archived, supported and expected to arrive every day. AIFS products are generated four times daily and extend to fifteen days. ECMWF also makes real-time output and model resources openly available.

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

The result 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 failure surface. Meteorology has reacted to a supposed paradigm war by putting both paradigms on the schedule.

THE STACK

The learned forecast still begins inside a physical system

Most leading learned weather models were trained on reanalysis rather than raw atmospheric observations. Reanalysis is the reconstruction described in the previous volume: observations are combined with a physical forecast model to estimate a complete atmospheric state. The learned model is therefore trained on an atmosphere that has already passed through physics and data assimilation.

Operational initialization follows the same pattern. AIFS begins from ECMWF's analysis, which is produced by a physics-based data assimilation system. The learned transition is fast because another part of the stack has already performed the difficult work of turning scattered observations into a coherent starting state.

This also explains why learned and physical models remain complementary. Physical models represent conservation laws and coupled processes explicitly, can be extended with new process knowledge, and provide the assimilation machinery that builds initial conditions. Learned models can reproduce large parts of atmospheric 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 a philosophical argument. It may be the system that assigns each component the job it performs best, notices when the components disagree, and preserves uncertainty when responsibility passes from one to another.

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.