Weather forecasting has traditionally required supercomputers, but in a potentially transformational breakthrough scientists at Cambridge University have developed Aardvark Weather, an AI system capable of generating accurate forecasts without relying on traditional physics-based simulations. Martin Cooper MBCS reports.
Described in a new Nature paper, Aardvark Weather replaces nearly every step of the current forecasting pipeline with a single deep learning system.
Today’s weather forecasts depend on a process known as numerical weather prediction (NWP), which uses equations that combine raw weather observation data with previous forecasts to predict how the atmosphere will evolve. The process requires vast amounts of computing power; even a single forecast from agencies like the ECMWF or the US National Weather Service can take thousands of CPU hours on a supercomputer.
Aardvark takes a radically different approach, treating forecasting as a pattern recognition problem rather than a physics calculation — an approach that’s faster, cheaper and potentially more flexible. The model can generate global forecasts from raw observations in about one second using just four high-end graphics processing units (GPUs).
How Aardvark Weather works
The model is designed around three main components: an encoder, a processor and a decoder. The encoder takes in raw observational data, such as weather balloon readings, and produces a structured picture of the current state of the atmosphere. Aardvark’s processor then takes over. Rather than simulate the laws of thermodynamics and fluid motion, it uses deep neural networks to predict the future state directly from the current one based on its learned understanding of weather patterns. It generates forecasts incrementally, using its own previous outputs to continue predicting further into the future. Finally, the decoder translates the model’s gridded forecasts into local predictions.
Crucially, all three components are trained together in what’s known as an end-to-end learning system. This means that during training, errors in the final forecast help adjust every part of the model. This unified approach enables Aardvark to learn not only how to forecast weather but also how to comprehend the structure and nuances of the data itself — a capability that traditional systems often struggle with.
Comparing systems
When tested against established forecasting models, Aardvark performed impressively. It matched or outperformed GFS on many key variables and timeframes, particularly for temperature and wind forecasts extending up to 10 days. Although ECMWF still leads on some metrics — especially at higher resolutions — Aardvark comes surprisingly close, all while requiring a fraction of the computing power.
The pros and pitfalls of Aardvark
One of Aardvark’s most significant advantages is accessibility. Because it doesn’t need a supercomputer or a full-scale weather centre to run, it opens the door for smaller countries, private companies and research groups to generate high quality forecasts. It also allows for customisation — for instance, a renewable energy company could optimise it for forecasting wind at turbine height.
For you
Be part of something bigger, join BCS, The Chartered Institute for IT.
The model’s flexibility could also help modernise forecasting in parts of the world that global systems have historically underserved. The researchers report that Aardvark performs exceptionally well in data-sparse regions, such as sub-Saharan Africa or the Pacific Islands, where traditional models often struggle.
Still, there are limits. The current version of Aardvark works at a coarser spatial resolution than leading NWP models and can lose accuracy in extreme weather events that are rare in the training data. Some weather features — especially small-scale, localised phenomena — also remain challenging for any global AI model to predict accurately.
Nevertheless, the researchers argue that Aardvark signals the start of a new era in weather forecasting, representing a fundamental shift towards data driven models. As climate change drives more extreme and unpredictable weather worldwide, the ability to deliver fast, accurate and affordable forecasts is more urgent than ever. Aardvark may not yet replace traditional weather models; but it suggests a future where forecasting is faster, more flexible, and more accessible to those who need it most.
Take it further
Interested in this and similar topics? Explore BCS' books and courses:
