A group of researchers from the University of Cambridge has developed a New weather model capable of predicting time more accurately and up to 10 times faster than conventional models. The model, created jointly with the Alan Turing Institute, Microsoft Research and the European Center for Medium -range weather predictions, is capable of doing so consuming less resources than conventional models. Is called Aardvark Weather and has potential to revolutionize the field of weather.
The model, whose details and results have been published in an article in the journal Nature, can make weather predictions up to 10 times more precise than current systems. In addition, to function it needs a lot of less computing power than other systems.
The research that has resulted in model development, led by the Professor of the Cambridge Richard Turner Engineering Departmentalso points out that Aardvark is several thousand times faster than other time prediction methods.
Researchers who have collaborated in their creation indicate that thanks to the AI that integrates the system, they have been able to replace the entire prediction of conventional time with a single Machine Learning model. This model can take as sources satellite observations data, weather stations and various types of sensors. It is able to generate local and global time forecasts.
According to Turner, one of the secrets of the improvements that Aardvark Weather brings to the weather prediction is its focus difference. Until now, the projects in charge of improving time forecast were limited to their prediction, and were not fixed in all tasks prior to them, which are carried out at a stage known as initialization. As they point out in New Scientist, with Aardvark they have also been able to restructure this phase, which now uses 90% less data in this phase than traditional systems.
It includes data from satellites, weather balloons and weather stations from around the world, and then purify, manipulate and merge them into an organized network that is taken as the basis for prediction. Turner points out that all these tasks consume half of the computing resources used in the process.
So far, in addition, a weather forecast needed the observation of several models. Each one of them needs a superordinate to generate, in addition to a support team to make it work. You also need a lot of time to generate these models. In general, it takes several hours, and even days in its creation. But with Aardvark Weather, the same job can be done in minutes, and even in seconds, and using a conventional desktop computer.
Of course, it will still be necessary to adjust the model to improve its results in complex and unexpected time patterns. This is because Aardvark uses a grid model of the land surface that has 1.5 square degrees.
Meanwhile, ECMWF’s ERA5 model uses a grid with only 0.3 degrees grids. That is why it is not yet able to predict with precision and detail extreme time events. Even so, Turner says that Aardvark already improves some models in the forecast of unusual events, such as cyclones.
Another weak point that the model still presents is that it depends completely on physics -based models for training, and that does not work if training data is deleted and is only used with data from observation. They have already tried to do it, but for now it has not worked.
Therefore, he believes that at least in the medium term the Future of the weather forecast It happens because lscientists work on models based on even more precise physicsand then use them to train AI models that replicate their exit more quickly and a lower hardware use. Some are even more optimistic about the future of weather forecast if AI intervenes.
