Explainer
How weather forecasts work: observations, models and AI
A weather forecast starts with observations, blends them with a short-range model forecast into a best estimate of today's atmosphere, then runs physics equations forward on supercomputers.[1][2] Because the atmosphere is chaotic, centres such as ECMWF run ensembles of 51 forecasts to show the range of outcomes, and since 2025 AI models such as ECMWF's AIFS have run operationally alongside the physics models.[3][4][5]
This explainer covers how everyday global weather forecasts are produced. For the AI side in more depth, read the AI weather models explainer next; for specific hazards, see the hurricane and wildfire explainers.
Step one: observing the atmosphere
Every forecast starts with measurements. Weather stations, aircraft, weather balloons (radiosondes), ships and buoys measure conditions where they are, while satellites and radar sense the atmosphere from a distance.[6] Satellites cover the whole planet, and the European Centre for Medium-Range Weather Forecasts (ECMWF) says they now contribute more to global forecast skill than any other kind of data.[7]
In-situ observations come from surface stations, aircraft, radiosondes, ships and buoys; remote sensing comes from satellites and ground-based radar.[6] Observations are unevenly distributed and carry errors, so on their own they cannot describe the global state at a given time.[8] ECMWF says satellite data now make the biggest contribution to global forecast skill, while in-situ data remain important, including for calibrating satellite instruments.[7]
Step two: building the starting point
Observations alone leave gaps, so forecasters combine them with a short-range forecast from the previous model run. This process, data assimilation, produces the “analysis”: the best estimate of the current state of the atmosphere, which becomes the forecast’s starting point.[1] The quality of a forecast depends heavily on the accuracy of that analysis.[1]
ECMWF assimilates atmospheric data with four-dimensional variational assimilation (4D-Var), which it introduced in 1997 and which quickly improved forecast quality.[9] Reanalysis datasets such as ECMWF’s ERA5, built from observations using numerical weather prediction, later became training data for AI weather models.[10]
Step three: running the physics
Traditional numerical weather prediction turns the physics equations of the atmosphere into algorithms run on supercomputers.[2] The US National Weather Service notes that such models only became useful once computers arrived, and that from the 1970s growing computing power and new observations, including satellite data, drove rapid improvement.[11]
Two global models are the best known. NOAA’s Global Forecast System (GFS) runs four times a day, forecasts up to 16 days ahead, and works at about 13 km resolution with 127 vertical layers.[12] ECMWF’s Integrated Forecasting System (IFS) has run both its high-resolution forecast and its medium-range ensemble at 9 km since an upgrade in June 2023.[13]
Ensembles: forecasting the uncertainty
The atmosphere is chaotic: a slight change in the starting conditions can lead to a significant change in the forecast.[3] Instead of relying on one run, an ensemble runs many simulations, each with slightly varied starting conditions and slightly perturbed models, to show the range of possible weather.[14]
ECMWF’s ensemble has 51 members: one unperturbed control forecast and 50 perturbed ones.[4] ECMWF began operational forecasting in 1979 and produced its first ensemble forecasts in 1992.[15] Over 30 years its ensemble resolution went from 210 km to 18 km, before the move to 9 km, and the lead time of its skilful ensemble forecasts grew by about one and a half days per decade on average.[16][13]
AI forecast models
Machine-learning models now forecast too. Google DeepMind’s GraphCast, published in 2023, beat ECMWF’s high-resolution forecast on more than 90% of 1,380 test variables and lead times, and makes a 10-day forecast in under a minute on one machine.[17][18] Its probabilistic counterpart GenCast beat ECMWF’s ensemble on 97.2% of 1,320 targets, and WeatherNext 2 followed in November 2025 (see weathernext).[19][20]
Operational centres have adopted AI alongside physics. ECMWF put its AI system, ecmwf-aifs, into operations on 25 February 2025 and its ensemble version on 1 July 2025, running beside the IFS.[5][21] ECMWF says the AIFS beats physics-based models on many measures, including tropical cyclone tracks by up to 20%, and uses about 1,000 times less energy per forecast.[22] NOAA deployed AI-driven global models, AIGFS and AIGEFS, plus a hybrid ensemble, on 17 December 2025.[23]
These models still depend on the traditional system: GraphCast was trained on four decades of ERA5 reanalysis, which is built from observations using numerical weather prediction.[10] For more, read how AI weather models work.
Forecasting hazards
The same machinery underpins hazard warnings. Tropical cyclones are very sensitive to small differences in atmospheric conditions, which makes them notoriously hard to forecast (see how hurricanes form).[24] The US National Weather Service issues Fire Weather Watches and Red Flag Warnings when critical fire weather is possible or imminent (see wildfires explained).[25] Related reading: El Niño and La Niña explained.
Read next
Next in the crash course
How AI weather models work
How machine-learning weather models learn from decades of past weather, why they are so fast, and how forecasters now use them.
Questions readers ask
Where does the data for a weather forecast come from?
From weather stations, aircraft, radiosondes, ships and buoys, and from remote sensing by satellites and radar. ECMWF says satellites now make the biggest contribution to global forecast skill.[6][7]
What is the difference between the European and American models?
ECMWF in Europe runs the Integrated Forecasting System, whose ensemble has run at 9 km since June 2023. NOAA's Global Forecast System runs four times a day at about 13 km and forecasts up to 16 days ahead.[13][12]
What is an ensemble forecast?
A set of forecasts run from slightly different starting conditions with slightly perturbed models. ECMWF's has 51 members, one control forecast and 50 perturbed ones, and the spread shows how uncertain the forecast is.[14][4]
Are forecasts getting better?
Yes. ECMWF says the lead time of its skilful ensemble forecasts has grown by about one and a half days per decade on average over 30 years.[16]
Do AI models replace physics-based forecasts?
Not so far. ECMWF's AIFS went operational in February 2025 running side by side with the physics-based IFS, and AI models such as GraphCast were trained on ERA5 reanalysis data produced with traditional numerical weather prediction.[5][10]
Why can't forecasts see weeks ahead in detail?
The atmosphere is chaotic, so a slight change in the starting conditions can lead to a significant change in the forecast. Ensembles exist to measure that sensitivity.[3]
Sources
· 25 claims from 8 publishers
Each numbered claim is a statement we checked against the sources listed with it. Tap a number in the text to see its sources in place.
- [1]
Data assimilation combines the latest observations with a short-range forecast to estimate the current state of the Earth system; this estimate, the analysis, provides the initial conditions for weather forecasts, and forecast quality depends on its accuracy. confirmedas of 2026-11-07
- Fact sheet: Earth system data assimilation · European Centre for Medium-Range Weather Forecasts · Primary source (retrieved 2026-11-07)
- [2]
Traditional numerical weather prediction translates physics equations into algorithms run on supercomputers. confirmedas of 2023-11-14
- GraphCast: AI model for faster and more accurate global weather forecasting · Google DeepMind · Primary source · 2023-11-14 (retrieved 2026-10-10)
- [3]
The atmosphere is a chaotic system, sensitive to initial conditions, so a slight change in inputs can lead to a significant change in the forecast. confirmedas of 2026-11-07
- Fact sheet: Ensemble weather forecasting · European Centre for Medium-Range Weather Forecasts · Primary source (retrieved 2026-11-07)
- [4]
ECMWF's ensemble produces 51 forecasts, one unperturbed control forecast and 50 with slight perturbations of initial conditions and models. confirmedas of 2026-11-07
- Fact sheet: Ensemble weather forecasting · European Centre for Medium-Range Weather Forecasts · Primary source (retrieved 2026-11-07)
- [5]
ECMWF took its Artificial Intelligence Forecasting System (AIFS) into operations on 25 February 2025, running side by side with its physics-based IFS. confirmedas of 2025-02-25
- ECMWF's AI forecasts become operational · ECMWF · Primary source · 2025-02-25 (retrieved 2026-10-10)
- [6]
Weather observations include in-situ measurements from weather stations, aircraft, radiosondes, ships and buoys, and remote sensing from satellites and ground-based radar. confirmedas of 2026-11-07
- Fact sheet: Earth system data assimilation · European Centre for Medium-Range Weather Forecasts · Primary source (retrieved 2026-11-07)
- [7]
ECMWF says satellite data now make the biggest contribution to global forecast skill, while in-situ data remain important, including for calibrating remote-sensing instruments. confirmedas of 2026-11-07
- Fact sheet: Earth system data assimilation · European Centre for Medium-Range Weather Forecasts · Primary source (retrieved 2026-11-07)
- [8]
ECMWF notes that observations are unevenly distributed and contain errors, so they cannot by themselves give a complete picture of the global state at a given time; data assimilation adjusts a short-range forecast to match them. confirmedas of 2026-11-07
- Fact sheet: Earth system data assimilation · European Centre for Medium-Range Weather Forecasts · Primary source (retrieved 2026-11-07)
- [9]
ECMWF uses four-dimensional variational data assimilation (4D-Var) for the atmosphere and implemented it in 1997, which quickly improved forecast quality. confirmedas of 2026-11-07
- Fact sheet: Earth system data assimilation · European Centre for Medium-Range Weather Forecasts · Primary source (retrieved 2026-11-07)
- [10]
GraphCast is a graph neural network trained on four decades of ECMWF's ERA5 reanalysis data, which is itself built from observations using traditional numerical weather prediction. confirmedas of 2023-11-14
- GraphCast: AI model for faster and more accurate global weather forecasting · Google DeepMind · Primary source · 2023-11-14 (retrieved 2026-10-10)
- [11]
The US National Weather Service says mathematical weather models only became useful with computers, and that from the 1970s more computing power and new observations, including satellite data, drove rapid gains in modelling skill. confirmedas of 2026-11-07
- About Models · US National Weather Service · Primary source (retrieved 2026-11-07)
- About Models · US National Weather Service · Primary source (retrieved 2026-11-07)
- [12]
NOAA's Global Forecast System (GFS) runs four times a day and produces forecasts up to 16 days ahead at about 13 km resolution with 127 vertical layers. confirmedas of 2026-11-07
- Global Forecast System (GFS) · NOAA NCEP Environmental Modeling Center · Primary source (retrieved 2026-11-07)
- Global Forecast System (GFS) · NOAA NCEP Environmental Modeling Center · Primary source (retrieved 2026-11-07)
- Global Forecast System (GFS) · NOAA NCEP Environmental Modeling Center · Primary source (retrieved 2026-11-07)
- [13]
ECMWF's Integrated Forecasting System upgrade to Cycle 48r1 on 27 June 2023 raised the horizontal resolution of its medium-range ensemble from 18 km to 9 km, the same as its high-resolution forecast. confirmedas of 2023-06-27
- Model upgrade increases skill and unifies medium-range resolutions · European Centre for Medium-Range Weather Forecasts · Primary source · 2023-06-27 (retrieved 2026-11-07)
- [14]
An ensemble forecast runs multiple simulations, each with slightly varied initial conditions and slightly perturbed models, to represent uncertainty and show the range of possible weather. confirmedas of 2026-11-07
- Fact sheet: Ensemble weather forecasting · European Centre for Medium-Range Weather Forecasts · Primary source (retrieved 2026-11-07)
- [15]
ECMWF began operational weather predictions in 1979 and produced its first ensemble forecasts in 1992. confirmedas of 2026-11-07
- Fact sheet: Ensemble weather forecasting · European Centre for Medium-Range Weather Forecasts · Primary source (retrieved 2026-11-07)
- Fact sheet: Ensemble weather forecasting · European Centre for Medium-Range Weather Forecasts · Primary source (retrieved 2026-11-07)
- [16]
ECMWF says that over 30 years its medium-range ensemble resolution rose from 210 km to 18 km, and the lead time of its skilful ensemble predictions increased by about one and a half days per decade on average. confirmedas of 2022-11-24
- 30 years of ensemble forecasting at ECMWF · European Centre for Medium-Range Weather Forecasts · Primary source (retrieved 2026-11-07)
- 30 years of ensemble forecasting at ECMWF · European Centre for Medium-Range Weather Forecasts · Primary source (retrieved 2026-11-07)
- [17]
Google DeepMind's GraphCast, described in Science in November 2023, gave more accurate predictions than ECMWF's HRES forecast on more than 90% of 1,380 test variables and lead times. confirmedas of 2023-11-14
- GraphCast: AI model for faster and more accurate global weather forecasting · Google DeepMind · Primary source · 2023-11-14 (retrieved 2026-10-10)
- [18]
GraphCast makes a 10-day forecast in less than a minute on a single Google TPU v4 machine, while a conventional HRES forecast can take hours on a supercomputer with hundreds of machines. confirmedas of 2023-11-14
- GraphCast: AI model for faster and more accurate global weather forecasting · Google DeepMind · Primary source · 2023-11-14 (retrieved 2026-10-10)
- [19]
GenCast, a probabilistic machine-learning weather model published in Nature in December 2024, had greater skill than ECMWF's ENS ensemble on 97.2% of 1,320 evaluated targets. confirmedas of 2024-12-04
- Probabilistic weather forecasting with machine learning · Nature · Primary source · 2024-12-04 (retrieved 2026-10-10)
- [20]
Google DeepMind's WeatherNext 2, announced in November 2025, uses a Functional Generative Network and surpasses the previous WeatherNext model on 99.9% of variables and lead times from 0 to 15 days. confirmedas of 2025-11-17
- WeatherNext 2: our most advanced weather forecasting model · Google DeepMind · Primary source · 2025-11-17 (retrieved 2026-10-10)
- [21]
ECMWF took the ensemble version of the AIFS into operations on 1 July 2025. confirmedas of 2025-07-01
- ECMWF's ensemble AI forecasts become operational · ECMWF · Primary source · 2025-07-01 (retrieved 2026-10-10)
- [22]
ECMWF says the AIFS outperforms physics-based models on many measures, including tropical cyclone tracks with gains of up to 20%, and uses about 1,000 times less energy per forecast. confirmedas of 2025-02-25
- ECMWF's AI forecasts become operational · ECMWF · Primary source · 2025-02-25 (retrieved 2026-10-10)
- ECMWF's AI forecasts become operational · ECMWF · Primary source · 2025-02-25 (retrieved 2026-10-10)
- [23]
On 17 December 2025 NOAA deployed operational AI-driven global models, AIGFS, AIGEFS and the hybrid HGEFS, which NOAA says deliver more accurate guidance using a fraction of the computing resources. confirmedas of 2025-12-17
- NOAA deploys new generation of AI-driven global weather models · World Meteorological Organization (news from members) · Primary source · 2025-12-17 (retrieved 2026-10-10)
- NOAA Deploys New AI Driven Global Weather Models · NOAA Earth Prediction Innovation Center · Primary source · 2025-12-17 (retrieved 2026-10-10)
- [24]
Google DeepMind notes that tropical cyclones are very sensitive to small differences in atmospheric conditions, which makes them notoriously difficult to forecast accurately. confirmedas of 2025-06-12
- Weather Lab: cyclone predictions with AI · Google DeepMind · Primary source · 2025-06-12 (retrieved 2026-10-10)
- [25]
The US National Weather Service issues a Fire Weather Watch when critical fire weather conditions are possible but not imminent, and a Red Flag Warning when critical fire weather is ongoing or expected shortly. confirmedas of 2026-11-07
- Fire Weather Watches and Red Flag Warnings · US National Weather Service · Primary source (retrieved 2026-11-07)
- Fire Weather Watches and Red Flag Warnings · US National Weather Service · Primary source (retrieved 2026-11-07)
Revision history (1)
- Page created.
Created Nov 7, 2026. Last reviewed by an editor on Nov 8, 2026. Next scheduled review: Mar 1, 2027.
Cite this page
"How weather forecasts work: observations, models and AI." ContentLora, updated Nov 7, 2026. https://contentlora.com/explain/how-weather-forecasts-work
Spotted an error? Suggest a correction or emailcorrections@contentlora.com.