What problem does it solve?
Weather forecasts come from numerical weather prediction: a physics model integrates the equations of the atmosphere from a starting state built from satellite, aircraft, ship and ground observations. It needs a supercomputer, and every extra ensemble member costs more computing time. Forecast users such as grid operators, airlines and emergency managers want more members, more frequent updates and a quicker delivery of the data.
A machine learning model learns how the weather evolved in the past and applies that to the current starting state. ECMWF describes the AIFS as trained on how the weather has evolved and fed every six hours with the same initial conditions as its physics based model. ECMWF and NOAA, the two weather services on this page, run AI models in operations next to their physics based models. The evidence is mixed on skill. ECMWF reports gains on many measures at a coarser grid, and an independent preprint on wind speed forecasts found the raw physics based ensemble more skillful at every horizon.
How does it work?
- Build the starting state. The weather service combines a short range forecast with observations in a data assimilation step. This step still uses the physics based system.
- Run the machine learning model. The model takes the starting state and steps the atmosphere forward, for example in six hour steps, for the next days. An ensemble version runs several dozen perturbed forecasts to show the range of outcomes.
- Run beside the physics based model. The service publishes the AI forecasts next to the physics based forecasts, often as open data, and some services combine both in a hybrid ensemble.
- Verify and compare. Skill scores against observations and against the physics based model are tracked for each variable, region and lead time, and forecasters keep the choice of which product to trust.
- Serve users. Energy companies, aviation, emergency services and others use the forecast data in their own models, such as renewable generation forecasts and storm outage prediction.
- Audience
- Back office
- Autonomy
- Copilot
- Adoption
- Early adopters
- Channels
- API and system to system
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Forecast accuracy improvement | Too few to pool | Not pooled: up to 20% | 0plus 1 up to | 1 organization |
Value drivers: Speed and cycle time, Lower cost to serve, Risk and loss reduction.
Which organizations already do this?
2 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
This library is not a list of Blits.ai customers. It documents public AI deployments from any vendor, each with its source.
European Centre for Medium-Range Weather Forecasts
Europe · Government and public sector · 2025
ECMWF, the intergovernmental forecasting centre of 35 Member and Co-operating States, took its Artificial Intelligence Forecasting System (AIFS) into operations on 25 February 2025 as a single deterministic forecast, and added the ensemble version, AIFS ENS, on 1 July 2025. Both run side by side with the physics based Integrated Forecasting System (IFS) and use the IFS analysis as their starting point.
ECMWF reports gains of up to 20% on some measures and about 1,000 times less energy to make a forecast, at a coarser grid (28 km for AIFS Single, 31 km for AIFS ENS against 9 km for the IFS). An independent comparison of 10 metre wind speed ensemble forecasts found the raw IFS ensemble substantially more skillful than the raw AIFS ensemble at every horizon.
- Forecast accuracy improvement: up to 20%, Gains against the leading physics based models on many measures, including tropical cyclone tracks, for AIFS Single
"The AIFS outperforms state-of-the-art physics-based models for many measures, including tropical cyclone tracks, with gains of up to 20%."
Claimed by: organization
National Oceanic and Atmospheric Administration
United States · Government and public sector · 2025
The National Weather Service's National Centers for Environmental Prediction (NCEP) made three AI based global models operational on 17 December 2025 after an evaluation period that began on 9 December. The Artificial Intelligence Global Forecast System (AIGFS) runs a deterministic forecast, the Artificial Intelligence Global Ensemble Forecast System (AIGEFS) runs 31 members, and the Hybrid Global Ensemble Forecast System (HGEFS) combines AIGEFS with the physics based GEFSv12 in a 62 member ensemble. The models are based on Google DeepMind's GraphCast and were developed with NOAA research laboratories and the Earth Prediction Innovation Center.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Decades of global weather reanalysis data for training and a recent archive for testing
- Operational initial conditions from a data assimilation system
- Station and satellite observations for verification
Systems to integrate
- Operational data assimilation and production schedule
- Verification and forecast product systems
- Open data or API distribution to users, such as grid operators and aviation
Complexity: High
A weather service needs decades of reanalysis data to train the model, accelerators for training, the physics based assimilation system that produces the starting state, a verification system that scores every variable against observations, and 24 hour operational support. Most organizations that use weather forecasts consume the output and do not build the model.
- 1
Choose the products to run
Decide whether to run an open AI model from a weather service or to train your own, and which variables, lead times and regions your users need.
- 2
Verify against observations
Score the AI forecast per variable and lead time against stations for your region, not only against global averages. Compare with the physics based forecast and with the post processed version of each.
- 3
Run in parallel
Publish the AI forecast next to the physics based forecast for a full season before it supports a warning or a trading decision.
- 4
Add post processing
Correct systematic errors with statistical post processing for the variables your users care about, such as wind speed at turbine height.
- 5
Review after model updates
Repeat the verification after each model version or after a change to the starting state.
Guardrails
- Forecasters keep the choice of product, and warnings rest on the full forecast suite
- Skill is reported per variable, region and lead time, with the grid resolution of each model stated
- The AI model never runs without the physics based starting state and a fallback to the physics based forecast
- Model versions are logged so a forecast can be reproduced
KPIs to instrument
- Skill scores per variable, region and lead time against observations, for the AI and the physics based model
- Skill of tropical cyclone track and intensity forecasts
- Time from the start of a run to the delivery of the forecast
- Energy and cost of computing per forecast
Human in the loop
Duty forecasters review the model output and decide on warnings and guidance. The service's science leadership decides which products are operational and when a model version is replaced.
Common failure modes
- Lower resolution hides local weather
- ECMWF's AIFS runs at 28 or 31 km against 9 km for its physics based model, so it can miss local extremes. Check the variables with local structure, such as precipitation and wind.
- Skill differs by variable
- An independent preprint on 10 metre wind speed found the raw physics based ensemble substantially more skillful than the raw AIFS ensemble at all horizons, with post processing narrowing the gap. A gain on one measure does not carry over to every variable.
- Extremes are checked only on average error
- Average skill scores can hide the cases that matter for warnings. Verify tropical cyclone intensity and extreme precipitation separately from track and average error.
What are the risks and rules?
EU AI Act
Minimal risk
A weather forecasting model is not an Annex III use and does not interact with people, so it carries no Article 50 duty. Its output informs warnings, energy trading and aviation, but it is not a safety component of critical infrastructure under Annex III point 2. A user that feeds the forecast into an automated control of critical infrastructure or into a system that falls under Annex III carries the high risk duties for that system, not for the forecast.
Rules that apply
Guidance
- Service Change Notice 25-89, Implementation of AIGFS, AIGEFS and HGEFS (NOAA National Weather Service, North America). Shows how a national weather service announces an AI model, with an evaluation period, data locations and the physics based model it is combined with.
- AI and physics-based weather forecasting: A comparative study (arXiv, Global). Independent comparison of operational IFS and AIFS 10 metre wind speed ensemble forecasts, July to November 2025.
Controls to put in place
- Verification against observations per variable, region and lead time, published with each model version
- A fallback to the physics based forecast if the AI model fails or drifts
- Versioned model and training data records
- Documented rules for which product supports a public warning
Frequently asked questions
- Do AI weather models replace physics based forecasts?
Not at the services on this page. ECMWF runs its AIFS next to the physics based IFS and says it sees them as complementary, and NOAA combines its AI ensemble with the physics based GEFSv12 in a 62 member hybrid ensemble. ECMWF's AIFS starts from the same initial conditions as its physics based model.
- How much better are AI weather forecasts?
ECMWF reports that the AIFS outperforms physics based models for many measures, including tropical cyclone tracks, with gains of up to 20%. An independent preprint on 10 metre wind speed found the raw physics based ensemble more skillful than the raw AIFS ensemble at all horizons, so the answer depends on the variable and the lead time.
- How much computing does an AI weather forecast save?
ECMWF states that the AIFS brings increased speed and a reduction of approximately 1,000 times in energy use for making a forecast. That covers the forecast step, and the data assimilation that produces the starting state still runs on the physics based system.
- Is a weather forecasting model regulated under the EU AI Act?
Usually at the minimal tier, because forecasting is not an Annex III use. A company that feeds forecasts into an automated control of critical infrastructure should assess that system on its own terms.
How to cite this page
Blits.ai AI Use Case Library, "AI numerical weather forecasting at weather services", last verified 18 November 2026, https://www.blits.ai/ai-use-cases/ai-numerical-weather-forecasting. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 18 November 2026: First published