First protein folding, now weather forecasting: London-based artificial intelligence (AI) company DeepMind continues to apply deep learning to tough scientific problems. In collaboration with the UK's national weather service, the Met Office, DeepMind has developed a deep learning tool called DGMR (deep generative model of rainfall) that can accurately predict the probability of rain in the coming 90 minutes, one of the most difficult challenges in weather forecasting.
In a blind comparison with existing tools, several dozen experts determined that DGMR's forecasts were the best on a number of factors (including their predictions of rainfall location, extent, movement, and intensity) on 89 % of the time. The results have been published in an article in Nature.
DeepMind's new tool is not AlphaFold, which solved a key problem in biology that scientists had struggled with for decades. But even a small improvement in forecasting is important.
Predicting rain, especially the heaviest, is crucial for many industries, from outdoor events to aviation and emergency services. But doing it right is hard. Knowing how much water is in the sky, and when and where it will fall, depends on a number of weather processes, including temperature changes, cloud formation, and wind. All of these factors are complex enough on their own, but even more so when combined.
The best existing forecasting techniques use massive computer simulations of atmospheric physics. They work well for long-term forecasting, but they're not as good at predicting what's going to happen in the next hour, known as nowcasting. Previous deep learning techniques have been developed, but they are often good at one thing, like location prediction, at the expense of others, like intensity prediction.
DGMR compared to actual radar data and two competing forecasting techniques for heavy rainfall in the eastern United States in April 2019. Photo: DeepMind
"Precipitation nowcasting remains a substantial challenge for meteorologists," says Greg Carbin, chief of forecasting operations at the US National Oceanic and Atmospheric Administration's (NOAA) Center for Weather Prediction. ., who did not participate in the work.
The DeepMind team trained their AI on radar data. Many countries publish throughout the day snapshots of radar measurements that track cloud formation and movement. In the UK, for example, a new reading is published every five minutes. Putting these snapshots together results in an up-to-date stop-motion video showing how rainfall patterns move across a country, similar to the forecast footage you see on TV.
The researchers fed this data with a deep generative network, similar to a GAN, a type of AI that is trained to generate new data samples that are very similar to the real data it was trained on. GANs have been used to generate fake faces, including fake Rembrandts. In this case, the DGMR learned to generate fake radar snapshots that continued the sequence of real measurements. It's the same idea as watching a few frames of a movie and guessing what's coming next, says Shakir Mohamed, who led the research at DeepMind.
To test the method, the team asked 56 Met Office meteorologists (who were not involved in the work) to rate the DGMR in a blind comparison with forecasts made by a state-of-the-art physical simulation and a rival deep learning tool; 89% said they preferred the results given by the DGMR.
"Machine learning algorithms often try to optimize a simple measure of how good their prediction is," says Niall Robinson, Met Office head of partnerships and product innovation, a co-author of the study. "However, weather forecasts can be good or bad in many different ways. One forecast might get precipitation in the right place but at the wrong intensity, or another might get the right mix of intensities but in the wrong places, etc. We have put a lot of effort into this research to evaluate our algorithm based on a broad set of parameters."
DeepMind's collaboration with the Met Office is a good example of AI development done in collaboration with the end user, something that seems like an obviously good idea but often doesn't happen. The team worked on the project for several years, with input from Met Office experts shaping the project. DeepMind research scientist Suman Ravuri says, "It drove our model development in a different way than we would have pursued on our own. Otherwise, we might have made a model that wasn't especially useful in the end."
DeepMind is also keen to show that its AI has practical applications. For Shakir, DGMR is part of the same story as AlphaFold: the company is cashing in on its years of solving tough game problems. Perhaps most importantly, DeepMind is beginning to cross off a list of real-world scientific problems.