How machine learning will help us prepare for extreme weather events
·
Published in
·
3 min read
·
23 hours ago
IMAGE: DeepMind
Here’s a topic I’ve been interested in for a while: using machine learning for weather modeling or forecasting. I came across this model published by DeepMind, the company founded by Demis Hassabis, Shane Legg and Mustafa Suleyman in 2010, and acquired by Alphabet for $650 million in 2014. It’s something that one of the co-founders of BigML, Tom Dietterich, has also been working on for some time, and that I’ve also had the opportunity to talk to him about on occasion.
The model, called GraphCast, outperforms the most accurate systems created by European and US government agencies, improving on 90% of the 1,380 verification targets, and very accurately forecasting severe events such as tropical cyclones, atmospheric rivers, heat waves, and other extreme temperatures.
On a planet suffering increasing destabilization due to a man-made climate emergency, it is vital to have systems capable of reliably forecasting the evolution of meteorological phenomena. Which is why DeepMind has created a machine learning model based on the re-analysis of more than forty years of historical data annotated with the climatological evolutions that they generated at the time, and using the computing capacity necessary to manage a model with several hundred variables. The result is a model capable of predicting hundreds of meteorological variables over a period of 10 days and with an overall resolution of 0.25°C, in less than a minute.
Last September, artificial intelligence models developed by Google, Nvidia and Huawei forecast Hurricane Lee’s track a week in advance: Lee rapidly intensified into a Category 5 hurricane in the Atlantic Ocean east of the Caribbean, then weakened before making landfall in Nova Scotia with a strength equivalent to a tropical storm. Despite being a relatively unusual event historically speaking, derived from the abnormally high sea surface temperatures that “feed” this type of hurricane, machine learning models are apparently able to predict them with great accuracy.
While traditional weather models make forecasts based on complex mathematical calculations and equations, and require enormous amounts of computing power, AI-based models like GraphCast take a different approach, recognizing patterns in large amounts of historical…
