When it comes to weather forecasts, supercomputers are usually mentioned. This is because a supercomputer is needed to calculate various variables related to the weather. Research has shown that weather forecasting is now possible without expensive supercomputers, thanks to artificial intelligence (AI). It is now expected that even developing countries will be able to conduct important weather forecasting for agriculture and fisheries using desktop computers.
Researchers from the University of Cambridge, the Alan Turing Institute, Microsoft Research in the United States, and the European Centre for Medium-Range Weather Forecasts (ECMWF) unveiled an AI-based weather prediction system called "Aardvark Weather" on the 21st. The research findings were published that day in the international academic journal "Nature."
The existing weather forecasting method known as "Numerical Weather Prediction (NWP)" collects data from satellites, weather stations, weather balloons, and other sources around the world and calculates complex physical equations. This process requires vast computational resources, necessitating supercomputers, and it took several hours to days to produce a forecast. Recently, companies like Google and DeepMind have developed methods to replace some of the calculations of the existing approach using AI, but there were limitations to processing the entire process with AI.
Aardvark Weather is the first system to process raw data collected from satellites, weather stations, weather balloons, ships, and aircraft, and to generate forecasts entirely using AI. The research team noted that it maintains high accuracy while being much faster and requiring fewer resources than traditional methods.
According to the research team, this system can generate forecasts in just one second on a standard desktop computer. The team stated, "The required input data is only about 10% of what is needed in the existing method, but it can achieve accuracy comparable to the latest NWP models," adding, "It is expected to provide accurate forecasts for eight days and enable detailed forecasting at specific regional levels."
Of course, there are limitations. Aardvark Weather uses a grid model with a resolution divided by latitude and longitude at intervals of 1.5 degrees, while ECMWF's latest AI model employs a much finer grid model with a 0.3-degree interval. Aardvark Weather may be somewhat disadvantaged in capturing sudden localized heavy rains or complex weather phenomena. The AI forecasting system developed by DeepMind operates on supercomputers.
Just having more observational information does not naturally solve these issues. AI learns from large amounts of information and acquires patterns on its own. However, weather forecasting AI cannot rely solely on information. A physical model that explains weather phenomena is still necessary. Richard Turner, a professor at the University of Cambridge who led the research, noted, "AI models are ultimately trained using traditional physics-based models, so training them solely on observational data has not yet been successful."
Professor Turner said, "If meteorologists develop more precise physical models in the future, they can train AI on them to build faster and more efficient forecasting systems," adding, "This could be applied in various fields, such as regional temperature predictions for African agriculture or wind speed forecasts for renewable energy production in Europe."
References
Nature (2025), DOI: https://doi.org/10.1038/s41586-025-08897-0