AI weather forecasting
Uses the attention mechanism to integrate long-range and short-range effects
In the image above, the blue lines are traditional predictions for the path of hurricane Beryl. The black line is the prediction by GraphCast, an AI program. The yellow line is the path Beryl took.
This is quite impressive. Here’s how the Times describes the current approach and the AI approach.
Current approach. Modelers build a virtual planet crisscrossed by millions of data voids and fill the empty spaces with current weather observations. This requires that forecaseters blend data from many sources into a guess at what the atmosphere is doing right now. The knotty equations of fluid mechanics then turn the blended observations into predictions. Despite the enormous power of supercomputers, the number crunching can take an hour or more. And of course, as the weather changes, the forecasts must be updated.
AI approach. Instead of relying on current readings and millions of calculations, an A.I. agent draws on what it has learned about the cause-and-effect relationships that govern the planet’s weather. The method works with great success because A.I. excels at pattern recognition. It can rapidly sort through mountains of information and spot intricacies that humans cannot discern. Doing so has led to breakthroughs in speech recognition, drug discovery, computer vision and cancer detection. In weather forecasting, A.I. learns about atmospheric forces by scanning repositories of real-world observations. It then identifies the subtle patterns and uses that knowledge to predict the weather, doing so with remarkable speed and accuracy. A.I. program are trained on four decades of global weather observations compiled by the European forecasting center. GraphCast can produce in seconds a 10-day forecast that would take a supercomputer more than an hour. In a series of tests, GraphCast outperformed the best forecasting model of the European Center for Medium-Range Weather Forecasts, considered the world’s top weather agency, more than 90 percent of the time.
Click the image above to get to the NYT article for more details.
GraphCast was developed by Google’s DeepMind, the same company that developed AlphaFold, which predicts how proteins will fold. This is the same company that developed AlphaZero, the system that beat the world champion Go player.
None of these systems are Large Language Models (LLMs)—although they use an attention mechanism similar to those that enable LLMs to function as well as they do in generating text. Nvidia has its own version of GraphCast. As explained here, its version uses the attention mechanism to “capture long-range dependencies within the data.”