This year was a nonstop parade of extreme weather events. Unprecedented heat swept the globe. This summer was the Earth’s hottest since 1880. From flash floods in California and ice storms in Texas to devastating wildfires in Maui and Canada, weather-related events deeply affected lives and communities.
Every second counts when it comes to predicting these events. AI could help.
This week, Google DeepMind released an AI that delivers 10-day weather forecasts with unprecedented accuracy and speed. Called GraphCast, the model can churn through hundreds of weather-related datapoints for a given location and generate predictions in under a minute. When challenged with over a thousand potential weather patterns, the AI beat state-of-the-art systems roughly 90 percent of the time.
But GraphCast isn’t just about building a more accurate weather app for picking wardrobes.
Although not explicitly trained to detect extreme weather patterns, the AI picked up several atmospheric events linked to these patterns. Compared to previous methods, it more accurately tracked cyclone trajectories and detected atmospheric rivers—sinewy regions in the atmosphere associated with flooding.
GraphCast also predicted the onset of extreme temperatures well in advance of current methods. With 2024 set to be even warmer and extreme weather events on the rise, the AI’s predictions could give communities valuable time to prepare and potentially save lives.
“GraphCast is now the most accurate 10-day global weather forecasting system in the world, and can predict extreme weather events further into the future than was previously possible,” the authors wrote in a DeepMind blog post.
Rainy Days
Predicting weather patterns, even just a week ahead, is an old but extremely challenging problem. We base many decisions on these forecasts. Some are embedded in our everyday lives: Should I grab my umbrella today? Other decisions are life-or-death, like when to issue orders to evacuate or shelter in place.
Our current forecasting software is largely based on physical models of the Earth’s atmosphere. By examining the physics of weather systems, scientists have written a number of equations from decades of data, which are then fed into supercomputers to generate predictions.
A prominent example is the Integrated Forecasting System at the European Center for Medium-Range Weather Forecasts. The system uses sophisticated calculations based on our current understanding of weather patterns to churn out predictions every six hours, providing the world with some of the most accurate weather forecasts available.
This system “and modern weather forecasting more generally, are triumphs of science and engineering,” wrote the DeepMind team.
Over the years, physics-based methods have rapidly improved in accuracy, in part thanks to more powerful computers. But they remain time consuming and costly.
This isn’t surprising. Weather is one the most complex physical systems on Earth. You might have heard of the butterfly effect: A butterfly flaps its wings, and this tiny change in the atmosphere alters the trajectory of a tornado. While just a metaphor, it captures the complexity of weather prediction.
GraphCast took a different approach. Forget physics, let’s find patterns in past weather data alone.
An AI Meteorologist
GraphCast builds on a type of neural network that’s previously been used to predict other physics-based systems, such as fluid dynamics.
It has three parts. First, the encoder maps relevant information—say, temperature and altitude at a certain location—onto an intricate graph. Think of this as an abstract infographic that machines can easily understand.
The second part is the processor which learns to analyze and pass information to the final part, the decoder. The decoder then translates the results into a real-world weather-prediction map. Altogether, GraphCast can predict weather patterns for the next six hours.
But six hours isn’t 10 days. Here’s the kicker. The AI can learn from its own forecasts. GraphCast’s predictions are fed back into itself as input, allowing it to progressively predict weather further out in time. It’s a method that’s also used in traditional weather prediction systems, the team wrote.
GraphCast was trained on nearly four decades of historical weather data. Taking a divide-and-conquer strategy, the team split the planet into small patches, roughly 17 by 17 miles at the equator. This resulted in more than a million “points” covering the globe.
For each point, the AI was trained with data collected at two times—one current, the other six hours ago—and included dozens of variables from the Earth’s surface and atmosphere—like temperature, humidity, and wind speed and direction at many different altitudes
The training was computationally intensive and took a month to complete.
Once trained, however, the AI itself is highly efficient. It can produce a 10-day forecast with a single TPU in under a minute. Traditional methods using supercomputers take hours of computation, explained the team.
Ray of Light
To test its abilities, the team pitted GraphCast against the current gold standard for weather prediction.
The AI was more accurate nearly 90 percent of the time. It especially excelled when relying only on data from the troposphere—the layer of atmosphere closest to the Earth and critical for weather forecasting—beating the competition 99.7 percent of the time. GraphCast also outperformed Pangu-Weather, a top competing weather model that uses machine learning.
The team next tested GraphCast in several dangerous weather scenarios: tracking tropical cyclones, detecting atmospheric rivers, and predicting extreme heat and cold. Although not trained on specific “warning signs,” the AI raised the alarm earlier than traditional models.
The model also had help from classic meteorology. For example, the team added existing cyclone tracking software to GraphCast’s forecasts. The combination paid off. In September, the AI successfully predicted the trajectory of Hurricane Lee as it swept up the East Coast towards Nova Scotia. The system accurately predicted the storm’s landfall nine days in advance—three precious days faster than traditional forecasting methods.
GraphCast won’t replace traditional physics-based models. Rather, DeepMind hopes it can bolster them. The European Center for Medium-Range Weather Forecasts is already experimenting with the model to see how it could be integrated into their predictions. DeepMind is also working to improve the AI’s ability to handle uncertainty—a critical need given the weather’s increasingly unpredictable behavior.
GraphCast isn’t the only AI weatherman. DeepMind and Google researchers previously built two regional models that can accurately forecast short-term weather 90 minutes or 24 hours ahead. However, GraphCast can look further ahead. When used with standard weather software, the combination could influence decisions on weather emergencies or guide climate policies. At the least, we might feel more confident about the decision to bring that umbrella to work.
“We believe this marks a turning point in weather forecasting,” the authors wrote.
Image Credit: Google DeepMind
