Aardvark Weather, a new AI model developed by researchers in the UK and Canada, can mark a turning point in global weather forecast by replacing traditional weather simulations with artificial intelligence to maximize cost efficiency and accuracy.
Researchers from the University of Cambridge, the Vector Institute at the University of Toronto and the Alan Turing Institute have unveiled the new findings in a recent report published in Nature.
In contrast to conventional prediction tools that simulate atmospheric physics through complex comparisons, earthvark is a “deep leather” model that generates worldwide predictions for wind, humidity, geopotential and temperature at multiple pressure levels.
It also provides local station forecasts for temperature of 2 meters and 10 meters of wind speed. Deep learning is a subset of machine learning that teaches computers to recognize patterns in large amounts of data.
“At the moment there are some computational expensive components in the forecast piping line,” Postdoctoral Fellow told the Vector Institute of the University of Toronto James Requeima Decrypt. “We have been able to replace many of these time -consuming parts with much lighter models that have been trained to perform the same tasks.”
By making those components more efficient, earthvark could perform predictions more often and at higher resolutions, which improves speed and accuracy.
As Requeima explained, the team designed components to replace each step in the forecast feeding line, in which raw observation data is converted into a weather forecast.
“We have found that as soon as this machine learning components are chained together, the overall performance improves considerably,” he said. “By refining the entire pipeline for the final task on which we focus, we can optimize each component, not only because of the isolated role, but also for how it contributes to the outcome that we give the most.”
The project also included researchers from Microsoft Research Cambridge, the European Center for Medium-Range Weather Prognoses (ECMWF) and the British Antarctic Survey.
Aardvark weather uses raw atmospheric data such as pressure, temperature and relative humidity measurements to produce worldwide and local predictions with high resolution.
The system is built around three neural components: an encoder, a processor and a decoder.
- Encoder: convert raw, unstructured observation data into a brushed representation of the atmosphere.
- Processor: generates weather forecasts from the bred data.
- Decoder: Translates the predictions into specific local predictions.
To improve the performance and accuracy of earthvark, components are trained for the first time in advance on ERA5-Heranalysegegegeverens and historical dataset of high quality of ECMWF and then refined using Real-World weather observations.
“Data assimilation generally works as a car -grazing procedure. You start with the current atmospheric prediction, generated by large dynamic systems that estimate the current condition. At the time of zero you have this initial condition,” Requeima said. “But data assimilation must also take in real-time measurements of external sensors. So you collect actual observations in addition to the prediction of the model and adjust the estimate of the atmosphere accordingly.”
A fraction of the costs – and time
According to the report, Aardvark can generate a complete global prediction using four Nvidia A100 GPUs in just one second compared to the hours required by older models such as the European Center for the forecast of medium -sized weather forecasts.
This drastic reduction of computer requirements makes high-quality, adaptable predictions accessible to regions and agencies without the means to operate full NWP systems. It also makes much faster refinement of the model possible.
Aardvark agrees with a growing series of tools that are aimed at helping meteorologists to predict and respond extreme weather conditions. During recent storms, such as hurricanes Helene and Milton, who struck the American east coast in October 2024, predictors emphasized the importance of AI in improving the prediction of storm intensity.
Looking ahead, Requeima noted that the team is planning to open Source Aardvark to make technology more accessible.
“I think it is an important step in the direction of democratizing weather modeling – making it lighter and more accessible to the public,” he said. “That is our hope. It is also an important progress in end-to-end weather modeling, in particular by a data-driven, machine learning approach.”
Published by Sebastian Sinclair and Josh Quitittner
Generally intelligent Newsletter
A weekly AI trip told by Gen, a generative AI model.
