The journal Nature has today published two artificial intelligence models that predict the weather with an accuracy that, for the first time, is at a level similar to that of the methods used today by meteorologists. The publications show the potential of AI in this field, drastically reducing the time needed to make forecasts and describe hitherto intractable phenomena.
The first of the models, developed by scientists from the multinational Huawei, in China, and named Pangu-Weather, has been trained with 39 years of global meteorological data to predict temperature, wind, pressure and humidity at different heights. The program has managed to make forecasts with an accuracy comparable to that of numerical forecasting methods, the most accurate tool available to meteorologists today. It is the first time that an artificial intelligence model has achieved such a feat.
The program is 10,000 times faster than current methods —it performs tasks in seconds in which hours are invested today—, and it can be used from any computer. This represents economic savings for meteorological agencies, which can allocate the budget to other fields of research and to improve the accuracy of forecasts.
Instead of solving very complex physical equations that take hours of calculations on supercomputers, as the currently used numerical prediction does, the model presented by Huawei compares the current situation with those that occurred in the past and deduces what is most likely to happen in the future. future.
In the second case, scientists from the School of Software at Tsinghua University, also in China, and from the University of California, in the United States, have used data from Chinese and North American radars to develop NowcastNet. The model predicts the precipitation that will take place in the next few minutes in a specific area, which is known as nowcasting, and which is one of the great current challenges in meteorology.
In this case, scientists have trained the model with physical equations, in addition to observations. The goal was to provide predictions of extreme events—such as torrential rains or hurricanes—that were physically plausible. Despite achieving this, the authors warn that their model is still incomplete and that it can be improved by integrating “more physical principles” and “more meteorological data” into the training.
“The works confirm the opening of new horizons for the AI ??techniques used in operational forecasting”, endorses Robert Monjo, director of Research and Innovation of the Foundation for Climate Research (FIClima), in statements to the Science Media Center Spain, who did not has participated in the investigation. For the expert, these models already “compete directly with traditional methods” of prediction.
Complement, not replace
However, despite the fact that the publication of these two investigations entails an important advance in the capacity of artificial intelligence to improve weather predictions, the independent experts consulted by this means rule out the idea that automatic learning (or machine learning) is going to to completely replace physical equations as a forecasting element, at least in the short term.
“Currently, AI elements are applied in weather forecasting,” says Jordi Moré, a member of the Applied Research and Modeling Area of ??the Meteorological Service of Catalonia (Meteocat), in an email. These elements are used simultaneously with traditional forecasting methods “although it is expected that in the coming years there will be more and more AI-based components within the weather forecasting system.”
The Pangu-Weather developers agree with Moré: “We believe that the future will be a combination of AI-based methods and physics-based methods,” says Lingxi Xie, one of the authors of the publication, in an email. “AI can capture some laws and relationships from training data that humans are not aware of, and physics is still important by offering principles that forecasting has to satisfy,” he concludes.
One of the cases in which the effectiveness of artificial intelligence models falters is precisely in that of the most extreme weather events, whose frequency and intensity will increase due to climate change. Being rare, they are underrepresented in model training.
That is why “an extreme weather event can trigger erratic predictions,” they point out in an analysis of the models also published today in Nature by Imme Ebert-Uphoff and Kyle Hilburn, from the Cooperative Institute for Atmospheric Research at Colorado State University, in the United States. Joined.
For David Quintero, Aemet’s Senior Meteorology Technician, this is especially important, since it is fair “in cases of extremes in which weather forecasting is more necessary than ever.” More than 34% of all disasters and 22% of deaths related to them were the result of extreme precipitation events, according to the World Meteorological Organization (WMO).
