Extreme weather events such as heavy rain are increasing worldwide in the course of climate change. So far, however, it was not possible to reliably predict the exact place where such masses of water will go down. This makes it difficult to prepare and the protection of people in the affected areas. Now researchers have developed a method that converts roughly resolved global weather data into high -precision regional precipitation cards using artificial intelligence. This enables a quick and reliable heavy rain forecast – and all over the world.
Due to climate change, weather extremes occur more and more and increase in intensity. “Heavy rain and floods are much more common in many regions of the world today than a few decades ago,” says senior author Christian Chwala from the Karlsruhe Institute of Technology (KIT). Where the water masses appear unexpectedly, houses and streets are washed away, fields are flooded and people killed. “But so far, data has been missing in many places in order to reliably assess such extreme events,” said Chwala.
From global made regional
Researchers around Chwala and his colleague Luca Glawion have now developed a new tool that could solve this problem. To do this, they trained a generative AI model (Spategan-A-A-A-A-A-A-A-A-A-A-A-A-A-AEST) with historical data from global weather models that describe the precipitation to 24 kilometers and one hour-a rather rough information situation. This data (ERA5) comes from the European Climate Service Copernicus and currently comprised every year since 1940, but the team only chose the years 2009 to 2020. In addition, Glawion and her colleagues trained the AI with significantly higher measurement data from weather radars of the German Weather Service, also from 2009 to 2020. close.
“Our AI model does not simply create a subsequent version of the input data, but creates several realizations of physically plausible, finely resolved precipitation cards,” explains Glawion. “Details up to two kilometers are visible every ten minutes.” This prediction is spatial and time -more precise than previous weather forecasts. At the same time, the model provides information on the probability with which the calculated regional extreme weather will really occur and how much water will fall on earth.
Weather forecast even without local data
The researchers then tested their trained AI model with weather radard data from 2021 from Germany, the USA and Australia-three countries with completely different climatic conditions. It was shown that the method can be applied to all these regions of world regions. The AI supported forecasts matched the rainfall actually occurred. The tests also showed that artificial intelligence can also predict heavy rain where there are no local weather data, solely based on the information from other parts of the other, such as Germany, the USA and Australia.
The AI-based forecast could therefore be used worldwide in the future to predict heavy rain events precisely and in good time, the team concludes. “Especially in the particularly vulnerable areas, the resources for close weather observation are often missing,” says co-author Julius Polz from KIT with a view of regions such as the global south. “With our approach, we can also estimate much more reliably in such data low areas, where heavy rain and floods threaten.” This means that regional climate risks can be better evaluated and residents in an emergency used to warn of floods and floods. In addition to acute disaster protection, the AI could also help in the long term to design better prevention measures, for example in flood protection.
Source: Luca Glawion (Karlsruhe Institute of Technology) et al.; NPJ Climate and Atmospheric Science, DOI: 10.1038/S41612-025-01103-Y
