Meteorological forecasts are a private preserve. A new AI model wants to democratize the prediction of time

A group of researchers has just raised a most striking option: democratize meteorology. Your new weather prediction system challenges traditional systems, very expensive in computational resourcesand make use of AI so that (almost) any of us can become a meteorologist who performs his own personalized predictions.

Aardvark Weather. This is the name of a new system of weather predictions that according to those responsible will make any researcher with a desktop PC in a full -fledged meteorologist. The system makes use of AI algorithms and raises an alternative to the conventional systems they use thousands of times more computing capacity.

Homemade predictions. The normal thing is that a weather forecast platform takes several hours to process a prognosis. For this, it also needs supercomputers and a team of experts who develop, maintain and display those forecast systems. Aardvark Weather allows you to train an AI model with data from Meteorological stationssatellites, ships or airplanes worldwide and then make predictions based on that data.

The investigation. The study Published in Nature this week comes from a group of researchers from the University of Cambridge, the Alan Turing Institute, Microsoft Research and the European Center for Medium-Russian Weather Forecasts (ECMWF). In it they explain how the numerical weather prediction (NWP) is being replaced For automatic learning and neuronal networks that allow “improving the speed and precision” of the prediction.

Hyperlocalized forecasts. Among other things the system would allow to offer hyperlocalized forecasts and adapted to specific industries. Richard Turner, automatic learning professor at the University of Cambridge, explained In The Guardian how this model could be used to predict temperatures for agricultural crops in areas of Africa or The wind speeds For a renewable energy company in Europe.

The time in the next eight days. Turner adds that the model could be able to generate precise forecasts for a range of up to eight days in the future, when it is normal for precise forecasts to only be guaranteed to five days.

Rapid. This system is capable of generating a complete forecast from observational data in a second when processing it in four NVIDIA A100 GPUS, when 1,000 hours-nodo is normally taken in the HRES model of the ECMWF.

Ideal for developing countries. There are regions in which these types of forecasts are especially important, and having a “personalized” system would be very useful. Aardvark Weather offers that option according to its creators, because both its implementation and its use is much more accessible.

Previous attempts. At the end of 2023 Deepmind precisely Graphcast announceda meteorological prediction system based on AI that had an operation up to 1,000 times cheaper in energy consumption. Its precision was in fact greater than the best of current systems, but it does not seem that development has been implemented in practice. A few months ago Deepmind researchers presented their evolution, called Gencastanother prediction based on automatic learning that improved its predecessor and that of course competes with Aardvark Weather. Everything therefore points to this type of systems are gaining ground and interest, but remains to be seen if they apply massively.

Image | Brian McGowan

In Xataka | Meteorological prediction will improve a lot in Spain: Aemet has invested 25 million euros in it