AI tackles rainfall

London typical weather, Oxford St. © iStock.

After protein-folding predictions, DeepMind is moving on to the weather. The London-based artificial intelligence company applies deep learning to difficult scientific problems. Working with the Met Office, the UK’s national weather service, DeepMind has developed a deep-learning tool called DGMR (“Deep Generative Model of Rainfall”) that can accurately predict the likelihood of rain in the upcoming 90 minutes, which is one of the toughest weather forecasting challenges. In a blind comparison of existing tools, several dozen meteorologists judged the DGMR’s forecasts to be the best in a range of categories, including location, extent and intensity of the rain, 89% of the time. Forecasting rain, especially heavy rain, is crucial for many purposes, from outdoor events to aviation to emergency services. It’s also handy for planning a bike ride or a picnic. The DeepMind team has trained its deep generative network similar to a GAN (“Generative Antagonist Network”) with a storm of data from weather radars, and given the UK climate, they really had no shortage of numbers.

⇨ Nature, Suman Ravuri, Karel Lenc, Matthew Willson, et autres, “Skilful precipitation nowcasting using deep generative models of radar.”

⇨ MIT Technology Review, Will Douglas Heaven, “DeepMind’s AI predicts almost exactly when and where it’s going to rain.”

⇨ DeepMind Blog, the Nowcasting team, “Nowcasting the next hour of rain.”

2021-09-29