Key takeaways
- In published tests, AI weather models beat the best traditional systems on most measures: Google’s GraphCast beat it on 90% of 1,380 targets, and GenCast beat the European ensemble on 97.2% of 1,320 targets.
- Europe’s weather centre, ECMWF, put its own AI model into operation in February 2025 and says it improves tropical cyclone tracks by up to 20% with about 1,000 times less energy per forecast.
- In 2025 a Google DeepMind hurricane model was, per a former National Hurricane Center branch chief, “the best guidance we saw this year.” The Hurricane Center says AI will be “a component of the hurricane forecast process going forward.”
- Limits: “days earlier” is not proven across all storms, the models are a “black box,” and traditional physics-based models still matter.
This article is general information based on published research, ECMWF, and NPR reporting. Always follow your local National Weather Service forecasts and warnings.
How traditional forecasts work
Traditional weather forecasting, called numerical weather prediction, uses physics. It takes today’s observations and runs equations for how wind, moisture and heat move through the atmosphere to forecast what happens next. NPR describes these models as based on “equations that calculate how wind, moisture and heat move around in the atmosphere,” and gives the US model GFS as an example. ECMWF says the physics-based model it runs uses a 9-km grid.
How AI forecasts work
AI weather models skip the equations. Instead, as a hurricane scientist told NPR, they “go back and look at historical records and tease out patterns and relationships about what has happened in the past in very subtle ways that a person could not extract on their own.” ECMWF says its AI model uses the same starting conditions as its physics model, which combine a previous short-term forecast with around 60 million quality-controlled observations from satellites, planes, boats, buoys and ground stations. The AI then learns from how past weather evolved.
The research results
Three peer-reviewed studies from 2023 to 2025 show the pace of change:
- GraphCast (Science, 2023): forecasts hundreds of weather variables for the next 10 days at 0.25° resolution “in under 1 minute.” It “significantly outperforms the most accurate operational deterministic systems on 90% of 1380 verification targets,” and supports better tropical cyclone tracking, atmospheric river and extreme temperature forecasts.
- Pangu-Weather (Nature, 2023): produced “stronger deterministic forecast results on reanalysis data in all tested variables” than ECMWF’s operational system, and tracked tropical cyclones more accurately when started from reanalysis data (a cleaned-up historical dataset, not live data).
- GenCast (Nature, 2025): generates a set of 15-day forecasts in 8 minutes, with greater skill than ECMWF’s ensemble (ENS) on 97.2% of 1,320 targets, and “better predicts extreme weather, tropical cyclone tracks and wind power production.”
An important detail: several tests used historical data as the starting point, not the messy live data a real forecast needs. That is one reason what happens in real operations matters.
From research to real forecasts
Europe. ECMWF took its Artificial Intelligence Forecasting System (AIFS) into operations on February 25, 2025, running alongside its physics model. It says AIFS “outperforms state-of-the-art physics-based models for many measures, including tropical cyclone tracks, with gains of up to 20%,” and cuts energy use for a forecast by around 1,000 times. It is an open model, with a 28-km grid.
United States. NPR reported on the 2025 Atlantic hurricane season that Google DeepMind’s AI hurricane model, developed with the National Hurricane Center and Colorado State University, was the standout. A week before Hurricane Melissa hit Jamaica as a Category 5 storm with 185 mph winds, forecast models disagreed on its path. The DeepMind model “accurately predict[ed] Melissa’s path and its Category 5 intensity,” according to NPR. James Franklin, a former National Hurricane Center branch chief who analyzed the season, said it “was the best guidance we saw this year,” and a chart he produced showed it with the lowest overall track error. The NHC’s science operations officer said: “I think it’s clear at this point that AI will be a component of the hurricane forecast process going forward.”
Does it give warnings “days earlier”?
Be careful here. The sources show:
- AI models are more accurate on average in many tests, and in one high-profile case (Melissa) an AI model got the path and intensity right when others disagreed a week before landfall.
- A researcher at Colorado State told NPR that in the past, AI models did well on a hurricane’s track but not its intensity, and that the Google model did well on intensity because it also used historical data on how past hurricanes developed. She expects forecasters will be able to forecast tracks and intensity earlier than before. That is an expectation, not a proven result for all storms.
- One strong season and one famous storm are not the same as a track record across many years.
The limits
- “Black box.” Franklin told NPR: “A lot of data goes in. You get a forecast that comes out. But you don’t really know how it came up with that.”
- Intensity. Intensity forecasts have been a weak spot for AI models, according to the CSU scientist, though 2025 results improved.
- Physics still matters. ECMWF runs AI “side by side” with its physics model, and NPR says additional AI models are in development by NOAA and ECMWF. The trend is combining them, not replacing the old models overnight.
What this means for you
- Follow official forecasts. The National Weather Service and National Hurricane Center issue the warnings. Their forecasts now use AI guidance among many tools.
- Don’t treat a viral model run as a forecast. Single model maps on social media can differ widely from official forecasts.
- Have a plan before the warning. Know your evacuation zone and keep supplies ready.
Frequently asked questions
Is AI better than traditional weather forecasting? In published tests, AI models beat the best traditional systems on most measures. In practice, forecasters use both.
Did AI predict Hurricane Melissa? NPR reported that Google DeepMind’s model accurately predicted Melissa’s path and Category 5 intensity a week before landfall, when other models disagreed.
Will AI give earlier warnings? Scientists expect forecasts further into the future to improve, but that isn’t proven for all storms.
Is AI weather forecasting in use now? Yes. ECMWF’s AI model has been operational since February 2025, and the National Hurricane Center used Google DeepMind’s model in 2025.
Information checked on: November 13, 2026. This field changes quickly, so check current sources.
Sources
- Learning skillful medium-range global weather forecasting. Science, 2023
- Accurate medium-range global weather forecasting with 3D neural networks. Nature, 2023
- Probabilistic weather forecasting with machine learning. Nature, 2025
- ECMWF, “ECMWF’s AI forecasts become operational” (February 25, 2025)
- Greg Allen, NPR, “As the 2025 Atlantic hurricane season ends, the future of forecasting is AI” (November 29, 2025)
