DeepMind has developed an algorithm to predict rainfall several hours ahead

The UK-based DeepMind has developed a deep learning tool DGMR that accurately predicts the probability of rain in the next 90 minutes.

From packing an umbrella to preparing for extreme conditions, predicting short term weather patterns is crucial for daily life.

New research with the @metoffice and SOTA model advances the science of Precipitation Nowcasting — the prediction of rain: https://t.co/gKGx76zwxS 1/4 pic.twitter.com/eHYVeNLWP9

— DeepMind (@DeepMind) September 29, 2021

Researchers trained the AI to work with radar data that meteorological agencies in many countries regularly publish. For example, in the United Kingdom, new readings are issued every five minutes.

Combining these scans formed a frame-by-frame video that shows the movement of clouds, similar to the televised visuals used in weather forecast programs.

The researchers fed the data into a deep generative network, resembling a GAN, where the algorithm learned to generate synthetic radar images that continued the sequence of real measurements.

“This is the same as watching a few frames from a movie and guessing what happens next,” said the study’s lead researcher Shakir Mohamed.

To test the algorithm, the team asked 56 forecasters from the UK Met Office to evaluate DGMR in a blind comparison with predictions from competing models. 89% of those surveyed said they preferred DeepMind’s results.

According to the scientists, short-term precipitation forecasting remains a challenging problem in meteorology. Existing methods rely on large-scale computer modelling of atmospheric physics. They are well-suited for long-range forecasting, but less effective for predicting weather over the next several hours.

In August, DeepMind developed universal architecture for creating artificial intelligence.

In late July, the AI lab researchers presented an extensive gaming environment XLand for training universal artificial intelligence agents.

In July, DeepMind assembled and published the most complete database of human protein structures, created by the AlphaFold neural network.

Subscribe to ForkLog news on Telegram: ForkLog AI — all news from the world of AI!