← All research storiesWeather
AI weather models tracked hurricanes better than a top physics forecast, but still underestimated their strongest winds
Testing three AI weather models on 178 strong tropical cyclones, researchers found they placed storms better than Europe’s leading physics-based forecast but missed peak wind strength.
What the scientists did
Tropical cyclones, called hurricanes or typhoons depending on where they form, are among the hardest weather to forecast, especially how strong they will get. New AI weather models learn patterns from decades of past weather data and can make forecasts much faster than traditional computer models that solve physics equations.
The team compared three AI models (GenCast, FourCastNet 3 and FGN) with IFS-ENS, a leading physics-based forecast from the European Centre for Medium-Range Weather Forecasts. They ran 15-day forecasts for 178 strong storms from every ocean basin, from 2021–2023 and from 2025, producing 26,541 individual forecast runs. Then they scored where each model put the storm, how strong it made it, and whether its range of possible outcomes was realistic.
Words to know
Tropical cyclone
A large rotating storm that forms over warm ocean water.
- Hurricanes (Atlantic and eastern Pacific)
- Typhoons (western Pacific)
- Cyclones (Indian Ocean)
Ensemble forecast
Running a forecast many times with slightly different starting points to see the range of possible outcomes.
- 50 slightly different forecasts of one storm
- “Spaghetti plots” showing many possible hurricane paths
Intensity
How strong a storm is, usually measured by its top sustained wind speed or its lowest air pressure.
- Category 1 to 5 on the Saffir–Simpson scale
- Maximum wind speed in knots
What they found
For 2025 storms, the AI model FGN gave better probability-based location forecasts than IFS-ENS for about 90% of the 41 storms scored, and an operational version of GenCast beat it for 78%. The AI models also followed storm tracks more often: about 79% of the time for FGN versus 66% for IFS-ENS in 2025.
But the AI models had weak spots. They tended to underestimate peak winds, and their ensembles were too narrow, meaning they were sometimes overconfident. Even so, they kept a realistic link between wind speed and air pressure, close to the reference data.
Why it matters to you
If you live near a coast, better track forecasts mean better information about where a hurricane is likely to hit, which shapes evacuations, school closures and when to get ready. This study suggests AI is already strong at the “where” but still weaker at the “how strong,” so forecasters will likely keep combining AI with traditional models and human judgment.
Where this could lead
Sharper hurricane track forecasts
Forecast centers could lean more on AI ensembles when predicting storm paths.
1–3 yearsBetter report cards for AI forecasts
The scoring methods in this study could help test new AI weather models before they are trusted for warnings.
1–3 yearsThese are possibilities the research points toward, not promises. Most early findings take years of testing before they reach everyday life, and some never do.
Why scientists care
Extreme events are the real test of any weather model, because that is when forecasts save lives and money. Scientists want to know whether AI models only copy typical weather or actually capture the physics of rare, violent storms, and this study suggests they capture at least some of it.
Keep in mind
This is a preprint that has not yet been peer-reviewed. It looks back at past storms rather than forecasting live ones, covers only a few seasons, and some comparisons used different storms, model versions and reference data. Several authors work for the insurance company AXA.
How far along is the science?
- Peer-reviewedReviewed by independent experts for a journal before publication. Peer review is a quality check, not a grade of evidence strength: confidence still depends on study design, limitations, replication and the wider evidence.
- PreprintA manuscript shared publicly, often before or alongside journal review. Check its separate journal and review status (each story says what we know). Results may change.
- Conference talkPresented to other scientists at a meeting. Usually short and early; a full paper may come later.
- Agency reportPublished by a science agency such as NASA, NOAA or NIH. Reviewed internally, but not by an outside journal.
- Press releaseAn announcement from a university or company. Useful, but always check the research it describes.
How soon could it reach you?
- Could reach you soonCould show up in products, services or advice soon. Our rough editorial guess is about 1 to 3 years, not a forecast from the research.
- A few years outPromising, but needs more testing or engineering first. Our rough editorial guess is several years, with a lot of uncertainty.
- Long-term scienceBasic research that builds knowledge. Real-world uses, if any, are far off.
- Small stepA small update to earlier work. Useful to researchers, but not a change you would notice yet.
Every story links to its original source: the paper itself, or the conference program when no paper is out yet. Explanations are written with AI, usually from the paper’s abstract (each story says what it was written from), then checked against the source. We only use journals, recognized preprint servers, science agencies, scientific conferences and established science news outlets.