DeepMind’s AI predicts almost exactly when and where it will rain
first protein folding, now weather forecast: The London-based AI firm continues to drive the application of in-depth learning to complex science problems. In partnership with the Met Office, the UK’s national weather service, DeepMind has developed an in-depth study called DGMR that can accurately predict the likelihood of rain over the next 90 minutes-one of the toughest challenges. to predict the weather.
In a separate comparison of existing equipment, dozens of experts judged DGMR announcements to be the best for a number of factors-including predictions of location, size, movement, and rainfall intensity-89% in time. The consequences published in a role in Nature now
New tool in DeepMind is not AlphaFold, which breaks down a fundamental problem in biology that scientists have struggled with for decades. Although even a slight improvement in forecasting factors.
Rainfall forecasting, especially heavy rainfall, is important for many industries, from outdoor activities to aviation to emergency services. But it’s hard to do well. Knowing how much water is in the sky, and when and where it will fall, depends on many weather processes, such as temperature changes, cloud formation, and wind. All of these factors are complex enough on their own, but they are even more complicated when combined.
The best available forecasting methods use multiple computer simulations of atmospheric physics. This works well for longer betting but is not very good at predicting what will happen next hour or so, known as nowcasting. The earlier methods of deep learning have been developed, but they are often well done with one thing, such as predicting location, at the expense of others, such as predicting speed.
“Rainfall today remains a significant challenge for meteorologists,” said Greg Carbin, head of forecast operations at the NOAA Weather Prediction Center in the U.S., who was not involved in the work.
The DeepMind team trained their AI on radar data. Many countries release regular snapshots of all -day radar resistances that track the formation and movement of clouds. In the UK, for example, a new reading is released every five minutes. Combining these snapshots provides a state-of-the-art stop-motion video that shows how rain patterns move in a country, similar to the visual effects you see on TV.