Weather forecasts can shape how much time communities have to prepare for dangerous conditions. Google DeepMind’s GraphCast offers meteorologists a faster way to predict weather, including the paths of cyclones and extreme temperatures. Its results point to a larger role for machine learning in forecasting, while leaving conventional models important for some tasks.
A longer lead time for dangerous weather
GraphCast can forecast conditions up to 10 days ahead. In research published in Science, it outperformed the European Centre for Medium-Range Weather Forecasts (ECMWF) model in more than 90% of over 1,300 test areas. For the troposphere, the lowest part of the atmosphere where most weather occurs, it performed better on more than 99% of weather variables, including rain and air temperature.
One example came in September, when GraphCast predicted that Hurricane Lee would make landfall in Nova Scotia nine days in advance. Traditional forecasting models identified Nova Scotia only six days in advance. That difference could give meteorologists and others more time to consider the storm’s likely path and prepare for its effects.
Earlier warnings matter because extreme weather can be difficult to anticipate, and climate change has made events less predictable. A forecast does not prevent a disaster, but more time and accuracy can support better preparation and may help save lives.
How the AI model makes a forecast
Conventional weather prediction relies on large computer simulations. They work through physics-based equations and weather variables such as temperature, precipitation, pressure, wind, humidity and cloudiness. Running these calculations can take substantial computing energy and time.
GraphCast takes a different route. It uses four decades of historical weather data to learn patterns, then produces its calculations in under a minute. Its graph neural network represents Earth’s surface as more than a million grid points. At each point, it predicts conditions including temperature, wind speed and direction, mean sea-level pressure and humidity.
Rather than calculating each weather variable through the same series of physics-based equations, the model uses relationships it has learned across the data to estimate what comes next. This speed is especially relevant when forecasters need to assess a developing threat and update expectations over time.
AI joins a changing forecast toolkit
GraphCast is part of a broader shift in weather forecasting. Other machine-learning models mentioned in the research include Huawei’s Pangu-Weather and Nvidia’s FourcastNet. GraphCast improves on competing models such as Pangu-Weather and can predict more weather variables, according to Rémi Lam, a staff research scientist at Google DeepMind.
The ECMWF is already using GraphCast. Peter Dueben, the center’s head of Earth system modeling, said its arrival showed that machine-learning models had become too capable to ignore. Other agencies, including the Swedish Meteorological and Hydrological Institute, have also used the graph neural network architecture proposed by Google DeepMind to build models of their own.
The change is not simply a matter of replacing one forecasting method with another. Meteorologists have long worked with complex physical simulations, and machine-learning models offer a different way to use historical data. Agencies can draw on both approaches as they assess forecasts and the limits of each model.
Where GraphCast still falls short
GraphCast does not lead conventional models in every area. Dueben points to precipitation as one area where it still lags. That gap is a reason meteorologists will need to use traditional forecasting alongside machine-learning systems to produce better predictions.
Google DeepMind is making GraphCast open source, which could let more researchers and weather agencies examine and build on the model. Aditya Grover, a UCLA computer science professor who developed ClimaX, described the approach as a moment of reckoning for weather prediction because it demonstrated what historical data can support.
The model’s promise, then, rests on both its speed and its place in a wider forecasting system. Earlier and more accurate warnings could improve preparation for some extreme events, while continued use of conventional methods can help cover areas where GraphCast remains less capable.