Creators of a new AI weather program assert that it can achieve the same level of accuracy as conventional forecasts produced over hours or days on powerful supercomputers within just one second on a desktop. Since the 1950s, weather forecasting has predominantly depended on physics-based models that analyze data from satellites, balloons, and weather stations. These calculations, known as numerical weather prediction (NWP), demand substantial computing resources and energy, utilizing large, costly supercomputers. Recently, there have been efforts to enhance this process with AI. For instance, Google researchers developed an AI tool last year that could substitute complex code in sections of a weather model, significantly reducing the required computational power. DeepMind then advanced this approach by employing AI to generate the entire forecast, which was adopted by the European Centre for Medium-Range Weather Forecasts (ECMWF) with the introduction of the Artificial Intelligence Forecasting System last month. However, this gradual integration of AI in weather prediction hasn't eliminated all traditional computational methods, a gap that Richard Turner from the University of Cambridge and his team aim to address with their new model. Turner notes that previous efforts focused solely on forecasting and overlooked the vital initialisation step, in which data from global satellites, balloons, and weather stations is combined and organized into a structured grid. “That process consumes half the computational resources, ” he emphasizes. The researchers developed a system called Aardvark Weather, which replaces both the forecasting and initialisation stages for the first time. Aardvark utilizes only 10 percent of the input data that current systems need while producing results comparable to leading NWP forecasts, as highlighted in their study. Compiling a full forecast, typically requiring hours or days on a supercomputer for NWP, can now be accomplished in about one second on a standard desktop with Aardvark. Nonetheless, Aardvark operates on a grid model of the Earth’s surface with cells measuring 1. 5 degrees square, in contrast to the ECMWF's ERA5 model, which features cells as small as 0. 3 degrees. This larger grid size means Aardvark may miss intricate and unexpected weather patterns, according to David Schultz from the University of Manchester. “There are many unresolved factors that could disrupt your forecast, ” Schultz states.

“At this scale, they fail to represent extremes effectively. ” Turner contends that Aardvark may outperform some existing models when identifying rare events such as cyclones. However, he acknowledges that AI models like his still rely heavily on physics-based models for their training. “It absolutely fails if you remove the physics model training data and only use observational data, ” he admits. “We attempted to create a completely physics-free model, but the results weren't successful. ” Turner envisions a future where weather forecasting involves scientists refining ever-more precise physics-based models that, in turn, train AI systems to replicate their outputs more quickly and efficiently. Some, like Nikita Gourianov from the University of Oxford, are even more hopeful, predicting that AI could eventually produce weather forecasts that exceed NWP capabilities, solely using observational and historical weather data. “It’s a matter of both scale and ingenuity, ” he notes. “Being clever about data input and structuring the neural network is crucial. ”