AI System Aardvark Weather Will Replace Supercomputers in Weather Forecasting

Scientists at the University of Cambridge have unveiled Aardvark Weather, an AI weather forecasting system that can generate both global and local forecasts in minutes using a regular desktop computer. Compared to traditional approaches, Aardvark provides forecasts tens of times faster and requires thousands of times less computing resources.

Image source: Brian McGowan / Unsplash

Aardvark Weather was developed at the University of Cambridge in collaboration with the Alan Turing Institute, Microsoft Research and the European Centre for Medium-Range Weather Forecasts (ECMWF). The speed and efficiency of weather models are critical: traditional methods rely on resource-intensive supercomputers and the work of many people, while forecasting often takes several hours. In this context, Aardvark demonstrates a fundamentally new approach, eliminating the need for cumbersome infrastructure.

Recent developments by tech giants like Huawei, Google, and Microsoft have shown that AI can significantly improve individual components of meteorological modeling. Particularly important have been advances in algorithms that model atmospheric dynamics over time. The integration of AI into these processes has already led to faster calculations and increased accuracy of weather forecasts.

A prime example is Google, which is developing AI-based models and offering two such solutions to corporate clients of its cloud platform. These models, created by researchers at Google DeepMind, use historical weather data and can predict weather conditions 10 to 15 days in advance.

Unlike hybrid approaches, Aardvark completely replaces the traditional multi-stage forecasting scheme with a single machine learning (ML) model. Using a standard desktop computer, the system processes data from various sources, including satellites and weather stations, and generates forecasts within minutes.

Professor Richard Turner, from the University of Cambridge’s Department of Engineering, who led the research, said: “Aardvark reimagines existing weather forecasting methods, making forecasts faster, cheaper, more flexible and more accurate than ever before. Aardvark is thousands of times faster than any previous weather forecasting method.” Aardvark will eventually be able to forecast eight days in advance with a high degree of accuracy, three days longer than current systems.

Despite using only a fraction of the data used by traditional models, Aardvark outperforms the US National Forecast System (GFS) on several key metrics, while remaining comparable in accuracy to the US National Weather Service (NWS), which typically uses a combination of models and expert analysis.

Image source: NASA

The system’s end-to-end machine learning approach could be easily adapted to other climate modelling tasks, from predicting hurricanes, tornadoes and wildfires to assessing air quality, ocean current dynamics and predicting sea ice, said Anna Allen, the study’s first author from the University of Cambridge’s Department of Computer Science and Technology.

One of Aardvark’s most notable features is its architectural simplicity and high flexibility. Because the system is trained directly on weather data, it can be quickly reconfigured to generate customized forecasts tailored to specific industries or regions. This could be, for example, a temperature forecast for agriculture in Africa or a wind speed forecast for European renewable energy companies. This approach is fundamentally different from traditional systems, which require years of work by large teams of specialists to set up.

This adaptability could be a game changer for countries with limited access to computing power and weather data. As Dr Scott Hosking of the Alan Turing Institute points out, moving weather forecasting from supercomputers to desktop machines makes highly accurate forecasts accessible to developing countries and data-poor regions.

The next stage of the project will be the formation of a new research group based at the Alan Turing Institute. Its task will be to study the possibilities of implementing the technology in countries of the Global South, as well as integrating Aardvark into larger environmental forecasting initiatives.