How DeepMind's artificial intelligence almost accurately predicts when and where it will rain
Posted: Sat Feb 01, 2025 4:52 am
DeepMind, a UK company, in collaboration with the Met Office, has developed DGMR (Deep Generative Model of Rainfall), a deep learning tool that solves the complex problem of weather forecasting. DGMR predicts with near-perfect accuracy the probability of rain occurring in the next 90 minutes. The results of the study are published in the journal Nature .
Forecasting rain, especially heavy rain, is a challenging task. However, it is of critical importance to many industries, from outdoor events to aviation and emergency services.
The amount of precipitation in the sky, as well as when and colombia number data where it falls, depends on a multitude of complex factors, such as temperature changes, cloud formation, and wind. Previous deep learning methods have typically been good at predicting just one of these, such as the location of precipitation.
computer simulations of atmospheric physics. They can tell what will happen in the long term, but they are less effective at making operational predictions for the next few hours.
How the new model works
The DeepMind team trained an AI model on radar data. Throughout the day, many countries release radar images that track cloud formation and movement. For example, in the UK, this happens every five minutes.
Together, these images provide a time-lapse video showing how precipitation moves across the country. Similar visual forecasts are commonly shown on television.
The researchers fed this data into a deep generative network, similar to a GAN—an artificial intelligence model that generates new data samples similar to the ones it was trained on. Such networks are commonly used to create fake faces. DGMR learned to create fake radar images that continued the sequence of real measurements.
Forecasting rain, especially heavy rain, is a challenging task. However, it is of critical importance to many industries, from outdoor events to aviation and emergency services.
The amount of precipitation in the sky, as well as when and colombia number data where it falls, depends on a multitude of complex factors, such as temperature changes, cloud formation, and wind. Previous deep learning methods have typically been good at predicting just one of these, such as the location of precipitation.
computer simulations of atmospheric physics. They can tell what will happen in the long term, but they are less effective at making operational predictions for the next few hours.
How the new model works
The DeepMind team trained an AI model on radar data. Throughout the day, many countries release radar images that track cloud formation and movement. For example, in the UK, this happens every five minutes.
Together, these images provide a time-lapse video showing how precipitation moves across the country. Similar visual forecasts are commonly shown on television.
The researchers fed this data into a deep generative network, similar to a GAN—an artificial intelligence model that generates new data samples similar to the ones it was trained on. Such networks are commonly used to create fake faces. DGMR learned to create fake radar images that continued the sequence of real measurements.