Google wants you to know that artificial intelligence is much more than a conversational chatbot. The subsidiary DeepMind presented in September a model capable of identifying genetic mutations that cause diseases. Now it has created a tool with the ability to predict where a hurricane will make landfall, much further in advance than current systems.
Scientists involved in the project say it is a revolutionary invention. And according to the research, published this Tuesday in the magazine Science, no wonder. GraphCast, the new model developed by Google DeepMind, was able in its tests predict weather conditions up to 10 days in advancewith greater precision and much faster than the current gold standard.
GraphCast, for example, outperformed the European Center for Medium-Range Weather Forecasts (ECMWF) model by more than 90% of 1380 test areas. In predictions for Earth’s troposphere — the lowest part of the atmosphere where most weather events occur — GraphCast outperformed the ECMWF model in more than 99% of the climatic variables. Among them, rain and air temperature.
The ECMWF has one of the largest supercomputer facilities and meteorological data archives in the world. It supports the operations of programs such as Copernicus, financed by the European Union and one of the key sources to keep track of climate change. Therefore, for a tool to be presented as superior is a lot.
How do current systems work?
The Google DeepMind tool managed to predict where Hurricane Lee would make landfall in Canada, a powerful event recorded in September, three days earlier than existing methods. Being able to warn earlier offers key time for authorities and populations to better prepare. Critical time to save lives.
A study published in October explains that Atlantic hurricanes are now more than twice as likely to intensify rapidly. Climate change is to blame. Hurricane Lee, for example, went in less than a day from being a category 1 event – with winds of 129 kilometers per hour – to a category 5 – 249 kilometers per hour. Therefore, it is crucial to save time.
Traditional weather forecasting is based on measurements of what is happening in real time in the atmosphere. In the best cases, like the ECMWF team, these measurements come from different parts of the world and from different instruments: satellites, buoys in the ocean, sensors on commercial airplanes.
Matthew Chantry of ECMRWF told BBC that a single one of its predictions can take into account around 10 million measurements. All this large amount of data is processed in some of its supercomputers to solve complex equations, based on physics and different climatic variables. Those in this European center can make up to a thousand billion calculations per second. And thus, knowing what the probability is of a certain event occurring in the future.
Models like these, however, require large computing resources. And, even with all their power, they can sometimes take hours to dictate their forecasts. Artificial intelligence comes to inject speed into analysis, even with less energy expenditure.
The contribution of Google’s tool to predict hurricanes
GraphCast uses machine learning to perform these calculations in less than a minute. Instead of physics-based equations, it leverages four decades of historical weather data to predict hurricanes and other events much faster.
Google’s DeepMind tool uses graph neural networks, which They map the Earth’s surface in more than a million grid points. At each point, the model predicts the temperature, wind speed and direction. Also the average pressure at sea level, humidity and other variables. With this, the neural network identifies patterns and predicts what will happen for each of these data points.
“GraphCast draws on decades of historical weather data to learn a model of the cause-and-effect relationships that govern how Earth’s climate evolves, from the present into the future,” the company explained on its blog. This Google DeepMind tool predicted the path of the hurricane Lee nine days early, while the ECMRWF achieved it six days early.
But this tool does not replace the measurements of centers such as the ECMWF, but rather complements them. “They go hand in hand,” said Google DeepMind. In fact, one of the great inputs that GraphCast uses are the measurements of this center. And the ECMWF team is already taking advantage of the new artificial intelligence system.
Peter Dueben, head of Earth system modeling at ECMWF, said when they introduced GraphCast to his team it felt like Christmas. “This showed that these models are so good that we can no longer avoid them”told MIT Technology Review. Google DeepMind says it doesn’t just want to predict the climate: “By developing new tools and accelerating research, we hope artificial intelligence can empower the global community to address our biggest environmental challenges.”