AI weather forecasting means using software that learns from past weather data to predict what the atmosphere will do next. It does not replace the physics-based models that forecasters have used for decades. Instead, the large weather centres now run AI models beside those physics models, compare the results, and use both when they prepare the forecast you see on your phone.
A physics-based model starts with the current state of the air, then works forward using equations for air movement, heat transfer, and rain formation. An AI model is trained on large sets of past weather records, learns how one weather state changes into the next, and applies those patterns to the latest observations. That difference explains both the speed gains and the limits below.
What AI weather forecasting means
ECMWF, the European Centre for Medium-Range Weather Forecasts, says its AI Forecasting System, called AIFS, uses the same starting conditions as its physics-based system. Those starting conditions combine a short-term forecast with around 60 million observations from satellites, planes, boats, sea buoys, and ground stations, updated every six hours.
From that starting point, the two systems part company. The physics system, with a grid spacing of 9 km, integrates the laws of physics in its code. The AI system, with a grid spacing of 28 km in its first version, uses a model trained on past weather. A smaller grid number means smaller squares, so the model can show smaller features.
NOAA, the United States National Oceanic and Atmospheric Administration, followed a similar path. Its first AI suite grew out of Project EAGLE, a joint effort between the National Weather Service, NOAA research laboratories, the Environmental Modeling Center, and the Earth Prediction Innovation Center. The team used Google DeepMind's GraphCast model as a starting point and fine-tuned it with NOAA's Global Data Assimilation System analyses, the agency's best estimates of the current atmosphere.
That history matters. An AI model is only as good as the records it learned from, and those records come from observations and earlier physics-based work. When a rare event has few past examples, the model has less to learn from.
What has improved so far
The first clear gain is speed. ECMWF says AIFS brings increased speed and uses approximately 1,000 times less energy for making a forecast. NOAA reports a similar pattern in its own suite. Its main AI model, the Artificial Intelligence Global Forecast System, or AIGFS, produces a single 16-day forecast using only 0.3% of the computing resources of the traditional Global Forecast System, and finishes in approximately 40 minutes.
Speed matters because a forecast is only useful if it arrives in time to act on. A faster run leaves more time to check the result, to run a second version, and to prepare a warning. It also lowers the computing bill, which is why centres can run more versions of the same forecast.
The second gain is in storm tracks. ECMWF reported on 25 February 2025, when AIFS became operational, that it outperforms state-of-the-art physics-based models for many measures, including tropical cyclone tracks, with gains of up to 20%. A tropical cyclone track is the path a storm is expected to follow. A better track helps a coastal town know sooner whether it sits in the likely path, and helps a shipping firm or an airline plan a safer route.
NOAA reported the same pattern. Its AIGFS shows a significant reduction in tropical cyclone track errors at longer lead times, so the forecast of where a storm will go stays useful further ahead. Its ensemble system, AIGEFS, which runs 31 slightly different versions to show a range of outcomes, extends useful forecast skill by 18 to 24 hours compared with the traditional ensemble system, while using only 9% of its computing resources.
The third gain is in how ensembles are combined. NOAA also runs a hybrid system, HGEFS, which joins the 31 AI members with the 31 physics members into a 62-member set. In initial testing, that combined set consistently outperformed both the AI-only and physics-only sets across most major checks. Two different methods, each with its own blind spots, can give a steadier answer together than either gives alone.
ECMWF moved in the same direction. After launching the single-forecast version, AIFS Single, in February 2025, it launched an ensemble version in July 2025. An ensemble is a group of forecasts with small differences in starting conditions. When most members agree, confidence is higher. When they spread apart, that spread is itself useful information.
Where AI forecasts still fall short
The same reports that praise storm tracks are careful about storm strength. NOAA said its first AI version, v1.0, shows a degradation in tropical cyclone intensity forecasts, even as track forecasts improve. Intensity means how strong a storm is, including its wind speed. A forecast that places a storm correctly but understates its strength can still mislead a reader who only checks the path.
That weakness is now being worked on in public. On 27 July 2026, the National Weather Service put an upgraded version, AIGFS v1.1, into use. Its notice says the changes improve tropical cyclone intensity forecasts, improve forecast precipitation, and reduce blurring at longer lead times. Blurring means a forecast that smooths out sharp detail as it looks further ahead, so a heavy rain band looks like a wide, light patch instead. The upgrade came from changes to the training method, including fine-tuning on four years of Global Data Assimilation System analysis.
A second limit is local detail. The World Meteorological Organization, known as WMO, noted in October 2025 that considerable challenges remain in AI systems' capability to support forecasts of local high-impact weather and hydrological processes, including floods and river flows. A global model with a 28 km grid cannot, on its own, show a single valley thunderstorm or a street-level flood. Local forecasters still need local observations, radar, and ground knowledge to turn a global forecast into a useful local warning.
A third limit is the rare event. AI models learn from the past, so an event with few past examples is harder for them to handle. That is why WMO's message is firm: AI must complement, not replace, existing forecasting methods and infrastructure. The physics models, the observation network, and the human forecaster all stay in the chain.
A fourth limit is trust in the result. An AI model can produce a confident-looking map without showing its working in the way a physics equation can be inspected. Forecasters therefore check AI output against physics output and observations. A reader can do a simpler version: compare two sources, look at the timing, and treat a single model run as one view, not a verdict.
How to read a forecast with care
Start with the warning, not the icon. A weather app icon shows one outcome for one place at one time. A warning from a national meteorological service, or from a local authority, is the decision that matters for safety. If the app and the warning disagree, follow the warning.
Next, look at the range. Many forecast pages now show a spread of possible temperatures or rainfall amounts. A narrow spread means the models mostly agree. A wide spread means the outcome is still open, and a plan that depends on a single number is risky. For a delivery or an outdoor event, a wide spread is a signal to prepare a second option, not a reason to cancel at once.
Then check the place and the time in detail. A forecast for a city can hide big differences between a hillside and a valley, or between a coast and an inland town. Read the hourly view, not only the daily summary, when the decision involves travel, stock, or staff safety.
For a small business, turn the forecast into a short list before the season starts. Decide in advance what you will do if heavy rain is forecast for a delivery day, strong wind for an outdoor stall, or extreme heat for staff working outside. Write down who decides and how customers will be told. A better forecast helps most when the response is already agreed.
Finally, keep a simple record. Note the forecast you acted on, the warning you followed, and what actually happened. Over a season, that record shows which source you trust for your area.
What to do next
If you run a small business, pick one weather-sensitive task this week and write its rain plan in three lines: the trigger, the action, and the person in charge. A trigger could be a heavy rain warning for your town. An action could be moving stock off the floor, confirming a delivery by phone, or shifting an outdoor job to the morning.
If you only want a better daily habit, choose one trusted source for warnings in your country and one forecast app for planning, in that order. Check the warning first, then the app for timing. When the two disagree, wait for the next update rather than picking the version you prefer.
If you work in farming, transport, energy, or construction, ask your forecast provider which models sit behind its product, and whether it shows an ensemble spread.
Frequently asked questions
Does AI weather forecasting replace human forecasters?
No. ECMWF and NOAA both run AI models beside physics-based models, and WMO says AI must complement, not replace, existing methods. Human forecasters check the models, add local knowledge, and issue warnings.
Why can two forecast apps disagree on the same day?
Because they may use different models, different starting data, or different ways of turning a global forecast into a local one. A disagreement is a sign of uncertainty, not proof that one app is broken. Check the warning source and the ensemble spread before you decide.
Is a faster forecast always a more accurate forecast?
No. Speed helps a forecast arrive in time and allows more versions to be run, but accuracy still depends on the observations, the training data, and the limits described above, including storm strength and local detail.
AI weather forecasting has improved the speed, cost, and storm-track skill of global forecasts, and that improvement is now part of daily operations at major centres. It has not removed uncertainty, and it has not replaced the need for local warnings, careful reading, and a plan made in advance.
If you would like help turning clear, checked explanations like this into content, a website, or a workflow for your own business, explore our services or get in touch through our contact page for a free consultation.
What to watch next
This article should be updated when NOAA issues a further AIGFS upgrade, when ECMWF reports a new AIFS version, or when WMO publishes progress on Early Warnings for All. At that point, the new figures, the date checked, and the Sources section should all be updated together, and any claim that no longer holds should be removed or corrected.
Sources
- ECMWF — ECMWF's AI forecasts become operational — date checked: 2026-11-01 — https://www.ecmwf.int/en/about/media-centre/news/2025/ecmwfs-ai-forecasts-become-operational
- ECMWF — About our forecasts — date checked: 2026-11-01 — https://www.ecmwf.int/en/forecasts/documentation-and-support
- HPCWire (source: NOAA) — NOAA deploys new generation of AI-driven global weather models — date checked: 2026-11-01 — https://www.hpcwire.com/aiwire/2025/12/17/noaa-deploys-new-generation-of-ai-driven-global-weather-models/
- World Meteorological Organization — World Meteorological Congress endorses actions to promote AI for forecasts and warnings — date checked: 2026-11-01 — https://wmo.int/news/media-centre/world-meteorological-congress-endorses-actions-promote-ai-forecasts-and-warnings
- National Weather Service — Service Change Notice 26-68: Implementation of the Artificial Intelligence Global Forecast System (AIGFS) upgrade (v1.1) — date checked: 2026-11-01 — https://weather.gov/media/notification/pdf_2026/scn26-68_AIGFS_v1.1.pdf
