The great hope for AI is that it can solve difficult problems, reducing costs and making solutions widely accessible. Aarvark is a weather forecasting system that can be run on a single desktop computer and gives results as good as those from America’s Global Forecast System (GFS).

Aarvark is being developed by the UK's Alan Turing Institute with partners including Google DeepMind, Cambridge University, the European Centre for Medium-Range Weather Forecasts and Microsoft. A paper published in Nature reports promising results that could make sophisticated forecasting accessible to those with limited resources. The project extends research by Huawei, Google, and Microsoft and demonstrates that one step of the weather prediction process known as the numerical solver, which calculates how weather evolves over time, can be replaced with AI to produce faster and more accurate predictions. 

Whereas traditional forecasts can take hours to produce on a supercomputer, Aardvark once trained, can create forecasts within minutes and can be run on a desktop computer. 

Aardvark takes in multimodal data from satellites, weather stations and weather balloons, and produces a ten-day global forecast. Its encoder module contains approximately 31 million parameters and requires 13 hours to train. The processor module contains approximately 54 million parameters, and requires 8 hours to train on ERA5,  and 3 hours to fine-tune using the output of the encoder module as input. Each of the eleven decoder modules contains approximately 2 million parameters and takes approximately 30 minutes to train. End-to-end finetuning of the encoder, processor and decoder modules takes 2 hours. The total time to train the model is therefore approximately 100 GPU hours and all the model training for the paper was performed on a single virtual machine with four NVIDIA A100 GPUs.

The visual below compares an Aardvark windspeed forecast with weather than subsequently happened (ground truth) showing that Aardvark can accurately predict wind speeds around the world, including the formation of 2018 tropical cyclone Berguitta which caused £100 million in damage in Mauritius & Réunion.

The results reported in nature are already impressive, revealing that Aardvark matches or outperforms GFS across most lead times. However, according to Professor Richard Turner and Dr Scott Hosking, writing on the Alan Turing Institute blog, Aardvark is only using about 10% of the available data to make its forecasts, meaning that further improvements in accuracy should be possible. 

They also point out that:

A streamlined system like this could also play a significant role in democratising access to advanced forecasting tools, empowering developing or data-sparse countries to build capacity and create bespoke weather forecasting systems that previously would have required large teams to operate, deploy and maintain. 

More Information

Project Aardvark - Alan Turing Institute blog

Allen, A., Markou, S., Tebbutt, W. et al. End-to-end data-driven weather prediction. Nature (2025). 

https://doi.org/10.1038/s41586-025-08897-0

Related Articles

 Alan Turing Institute Starts Work

DeepMind Takes On The Rain

Data Analytics for Weathering Storms

Google Maps Weather Layer API

To be informed about new articles on I Programmer, sign up for our weekly newsletter, subscribe to the RSS feed and follow us on Twitter, Facebook or Linkedin.

Comments

Make a Comment or View Existing Comments Using Disqus

or email your comment to: comments@i-programmer.info