In an era when extreme weather events are becoming more frequent and intense, researchers are on a quest for faster and more accurate prediction methods. Enter artificial intelligence (AI), which is bringing new possibilities.
In May of this year, Microsoft unveiled the weather prediction tool Aurora. Paris Perdikaris, a Microsoft researcher involved in the project, said that AI tools are good at recognizing patterns. To train Aurora, Microsoft provided over 1 million hours of climate data, which is 16 times that of the latest GPT model.
Now, Aurora can predict global air pollution conditions for the next five days and weather conditions for the next ten days 5,000 times faster than traditional methods. The Weather Company’s previous AI weather model could predict storm intensity but was too coarse. After collaborating with Nvidia this year, computing power has been enhanced, and AI predictions are more accurate and detailed.
A team from Villanova University is targeting storm climate phenomena and determining their impact by identifying storm scales and shapes through AI. With the help of machine learning, the prediction warning time has been advanced from 15 minutes before occurrence to one hour, providing residents with more response time.
“Speed” is a significant advantage of such AI tools. Over the past 50 years, general circulation models (GCMs) have been the mainstream for weather prediction. They require a large amount of climate data and supercomputer operations. They are accurate but consume a lot of time and energy, and inaccurate data may lead to calculation errors. The new AI weather prediction tool may run on a laptop, but its accuracy remains to be seen.
Microsoft said that Aurora will be open to the public in the coming months and hopes that climate researchers will test it. Perdikaris predicts that AI can be integrated into workflows in the next 2 to 5 years.
Unlike Microsoft Aurora’s “pure machine learning” model, Google DeepMind is trying a more comprehensive model. Last month, in a new paper, Google DeepMind pointed out that its new model “NeuralGCM” is more accurate than pure machine learning models and existing models in climate predictions for 1 to 10 days.
Aaron Hill, an assistant professor of meteorology at the University of Oklahoma, said that the innovation of this model lies in integrating AI while retaining some fluid dynamics calculations. Use traditional methods for large-scale atmospheric change predictions, and integrate AI predictions for cloud formation within a range of less than 25 kilometers or regional microclimates. Google researcher Stephan Hoyer said that selectively integrating AI can correct areas where errors may accumulate on a small scale. NeuralGCM can reduce the demand for computing power while maintaining prediction accuracy.
Hill believes that such AI tools reduce the computing power burden and have the potential to build and operate long-term and large-scale climate models. In the face of the climate crisis, commodity traders, agricultural planners, and the insurance industry are all willing to pay for faster and more accurate weather prediction models. This field is developing rapidly, but people are still waiting and watching.
