Deepmind: AI delivers best weather forecast for more than two weeks in advance
Google's Deepmind has been researching AI-generated weather forecasts with minimal resource consumption for years. Now the next milestone has been reached.
(Image: eskystudio / Shutterstock.com)
Google's AI research lab Deepmind claims to have developed a weather forecast based on AI technology that beats even the leading technology currently in operational use. The high-resolution GenCast model can forecast the weather on a small scale over a period of 15 days better than the European Center for Medium-Range Weather Forecasts (ECMWF), the researchers explain in a blog post. At the same time, only one of Google's fifth-generation Tensor Processors (TPU) is required for such a forecast.
A processor instead of a supercomputer
GenCast was trained using four decades of weather data from the ECMWF archive, covering years up to 2018. According to Deepmind, the model learned "global weather patterns" from this data and then tested its performance on data from 2019. In more than 97 percent of cases, the forecasts generated were better than those of the best forecasting model currently used at ECMWF for forecasting. However, while a supercomputer took hours to calculate this, GenCast only needs around 8 minutes with a TPU.
Videos by heise
Deepmind has been researching AI-generated weather forecasts for some time now, and a year ago the laboratory was already reporting that established models were being beaten and that significantly less computing power was required. The greatest progress has now been made in terms of range; a year ago, the more precise forecasts only reached 10 days into the future. In the blog entry, the researchers now announce that they soon want to publish forecasts in real time. These could then be integrated into other forecasts. They present the latest progress in an article in the journal Nature.
(mho)
