Aardvark Weather

Aardvark Weather is a machine learning system from the University of Cambridge, with collaborators at the Alan Turing Institute, Microsoft Research, and ECMWF, that replaces the entire traditional numerical weather prediction pipeline with a single neural model. Published in Nature in March 2025, it ingests raw observations from satellites, stations, and sensors and outputs global gridded and local station forecasts directly.

The striking claim is efficiency: forecasts that once needed supercomputers and large support teams can be produced in minutes on a desktop computer, using a fraction of the input data and orders of magnitude less compute, while matching or beating operational baselines on multiple variables with skillful local forecasts out to ten days. Because it learns directly from data, it can be quickly adapted into bespoke models for specific industries or regions, which the researchers frame as a path to high-quality forecasting for data-sparse developing nations. It remains a research system rather than an operational product.