Extreme weather events are occurring with increasing frequency and intensity, a trend many people have observed. Researchers are diligently seeking faster and more accurate ways to predict these events, with AI presenting a new possibility. In May of this year, Microsoft not only updated its Copilot but also launched a weather forecasting tool named Aurora. These AI tools excel at pattern recognition, as noted by Microsoft researcher Paris Perdikaris. To train Aurora, Microsoft fed it more than a million hours of climate data—about 16 times the amount used for the latest GPT model. Currently, Aurora can predict global air pollution levels for the next five days and weather conditions for the next ten days at a speed 5,000 times faster than traditional methods.
Previously, AI weather models like The Weather Company’s could estimate the intensity of impending storms, but the predictions lacked the precision needed for detailed wind speed and rainfall forecasts. After a partnership with NVIDIA this year, enhanced computing power has allowed for quicker and more precise AI predictions. A team from Villanova University has focused on storm patterns, aiming to evaluate their severity based on size and shape, which helps determine the likelihood of phenomena such as tornadoes or hail. Traditionally, students manually analyzed storm images, but AI now performs this task: "You train it again and again, and its performance improves significantly."
With the help of machine learning, the warning time for predictions has increased from 15 minutes before an event to one hour, giving residents more time to prepare. While you can't move a farm, you can bring livestock indoors. The key advantage of these AI tools lies in speed. Over the last 50 years, traditional weather forecasting has relied on General Circulation Models (GCMs), which are complex mathematical frameworks requiring vast amounts of climate data and powerful supercomputers. While these models are accurate, they consume considerable time and energy.
Additionally, the data collected from weather stations, balloons, buoys, and satellites can be inaccurate, with small errors potentially leading to significant miscalculations. In contrast, new AI weather prediction tools can operate on standard laptops, although their accuracy is still under evaluation. Microsoft indicates that Aurora will be publicly available in the coming months and hopes climate researchers will utilize it for testing. Perdikaris anticipates that researchers will ultimately decide whether to incorporate models like Aurora into their workflows, predicting AI integration could happen within the next 2 to 5 years.
Unlike Microsoft's approach, Google DeepMind is exploring a more integrated model. Last month, it published a paper indicating that its new model, "neuralGCM," offers more accurate climate forecasts for 1 to 10 days than both pure machine learning models and currently used methods. Assistant Professor Aaron Hill from the University of Oklahoma highlights the model's innovation in combining AI with fluid dynamics computations for larger atmospheric changes, while employing AI for finer-scale predictions, like cloud formation within a 25-kilometer radius. "We selectively integrate AI to correct small-scale errors," says Google researcher Stephan Hoyer.
According to the paper, NeuralGCM maintains prediction accuracy while significantly reducing computational demands. Although AI's advantages currently manifest in small-scale computations, Hill argues that the most meaningful aspect of these tools is their potential to build and run long-term, large-scale climate models without overwhelming computational resources. "Simulating global climate over extended periods is computationally intensive. We need not abandon our accumulated atmospheric knowledge from the last 100 years; instead, we can integrate it with AI and machine learning."
In the face of the climate crisis, various sectors—including commodity trading, agricultural planning, and insurance—are keen on faster and more accurate weather prediction models, fueling rapid development in this field. The demand for AI solutions exists, but many stakeholders are still in a wait-and-see mode.
Ranking
