Using DeepMind’s AI, DeepMind predicts when, where, and when it’s going to rain!

The firm developed a model that was better for short-term forecasts than existing systems, after working with UK weather forecasters.

Initially, DeepMind applied deep learning to a protein folding problem, and now it’s tackling weather forecasting. DeepMind has developed a deep-learning tool called DGMR in collaboration with the Met Office, the UK’s national weather service, that can accurately predict how likely it is to rain in the next 90 minutes.

Experts evaluated DGMR’s forecasts, which ranked first overall on several factors, including the direction, extent, and intensity of rain, in a blind comparison to existing tools 89% of the time. A Nature paper published today reports the results.

The DeepMind tool is not AlphaFold, which cracked open a biology problem that scientists had been battling for decades. However, every step forward counts.

Several industries rely on the forecasting of rain, including outdoor events, aviation and emergency services. However, this can be a challenging task. In order to figure out how much and where water is in the sky, a variety of weather processes have to be taken into account, such as temperature changes, cloud formations, and wind. Even though each one of these factors is complex on its own, when combined, they become even more complicated.

Simulations of atmospheric physics using massive computer programs are the best existing forecasting techniques. In the short term, these methods work well. However, they are not as effective at predicting what will happen in the next hour or so, known as nowcasting. Previous deep-learning techniques have been developed, but they usually excel at predicting location but not intensity, for example.

Heavy rainfall in eastern USA in April 2019 compared to DGMR and two rival forecasting techniques

“The nowcasting of precipitation remains a substantial challenge for meteorologists,” says Greg Carbin, chief of forecast operations at the NOAA Weather Prediction Center in the US, who was not involved in the work.

Radar data was used to train DeepMind’s AI. Various countries release frequently updated radar measurements throughout the day that trace cloud formation and movement. An updated reading is issued every five minutes, for example, in the UK. In a similar way to forecast visuals on TV, the stop-motion video built from these snapshots shows how rain patterns move across a country as they occur.

Data from these experiments was fed into a deep generative network similar to a GAN, which ensures that new samples of data will look like the original data it was trained on. Fake Rembrandts have been created using GANs. For this one, the DGMR used a fake radar snapshot that followed the actual measurement sequence. Shakir Mohamed, the DeepMind researcher who led the research, says the technique is the same idea as seeing a few frames from a movie and guessing what will happen next.

To prove the effectiveness of this approach, they asked 56 weather forecasters at the Met Office (who weren’t otherwise involved with the project) to compare DGMR’s results to a physics simulation and a deep-learning competitor tool; 89% reported preferring DGMR.

“Machine-learning algorithms generally try and optimize for one simple measure of how good its prediction is,” says Niall Robinson, head of partnerships and product innovation at the Met Office, who coauthored the study. “However, weather forecasts can be good or bad in lots of different ways. Perhaps one forecast gets precipitation in the right location but at the wrong intensity, or another gets the right mix of intensities but in the wrong places, and so on. We went to a lot of effort in this research to assess our algorithm against a wide suite of metrics.”

This collaboration between DeepMind and the Met Office demonstrates AI development that is done in collaboration with the end user, an apparent good idea that is rarely implemented. A Met Office expert’s input shaped the project as it was developed over several years. “It pushed our model development in a different way than we would have gone down on our own,” says Suman Ravuri, a research scientist at DeepMind. “Otherwise we might have made a model that was ultimately not particularly useful.”

In addition, DeepMind is eager to show that it can develop practical applications for its artificial intelligence. According to Shakir, DGMR is part of the same trend as AlphaFold: it is capitalizing on years of experience solving hard problems in games. DeepMind is finally checking off their bucket list of problems related to real-world science, which is perhaps the most important takeaway here.

Written by IOI

Get the latest stories from Tech & Innovation from around the globe. Subscribe Now!

Leave a Reply

Your email address will not be published. Required fields are marked *


GameStop soaring their way into NFT space with a non fungible asset management platform! Check it out!


Sanjay Bhargava hired as Country Director of SpaceX, India!