Graphcast
Better than any weatherman: Google AI is supposed to predict weather more accurately than was previously possible
Especially in extreme weather like Hurricane Lee, it is important to be able to predict the course as accurately as possible.
© NOAA via NASA / DPA
With Graphcast, Google has developed an AI that is supposed to outshine any previous weather model. Extreme weather situations should thus be able to be recognized and calculated much more precisely.
The weather frog could become an AI in the future: According to a report in the Financial Times, a new weather AI called Graphcast beat the world’s leading system for calculating forecasts in 90 percent of the metrics used – and only needed a fraction of the time. Google is behind the development of Graphcast.
The AI was 40 years old Weather data from the European Center for Medium-Range Weather Forecasts (ECMWF), it is said. Based on their analysis, Graphcast should already be able to make very precise predictions for the next ten days. According to the report, experts are surprised by the speed at which Graphcast learns.
13 images
Weather AI is already cheaper and faster today
With the weather data in memory, Graphcast analyzes the current state of the global atmosphere. This means that the system is able to produce a very precise forecast within a few minutes. According to the “FT”, this is not only more accurate than previous models, but also saves a lot of time.
The so-called numerical weather forecast is currently common. Here, supercomputers calculate equations based on scientific knowledge of atmospheric physics. This is not only very energy-intensive, but sometimes takes several hours.
Matthew Chantry machine learning coordinator at ECMWF, told the newspaper: “Once trained, Graphcast is incredibly cost-effective to run. In terms of energy consumption, we can talk about a 1,000-fold saving. This is a wonderful improvement.”
A concrete example shows how an AI like Graphcast could soon impact forecasts: In September, the system was able to calculate the path of Hurricane Lee nine days in advance. Conventional models only put it at six days before the hurricane arrived in Nova Scotia. In extreme weather conditions, three days is an enormous amount of time to take measures.
Combination of tradition and modern Google technology
The system probably still has difficulties with unexpectedly rapid developments, such as those that recently occurred with Hurricane “Otis”. To further improve the AI, a combination of the Graphcast approach and traditional calculations will now be used. “There is an opportunity to bring our understanding of physics into these machine learning systems,” Chantry told the Financial Times.
Experts agree that weather models will become significantly more reliable and detailed in the future with the help of AI. The ECMWF therefore wants to research its own system in the next step.