In early July, as Hurricane Beryl battered the Caribbean, a top European weather agency predicted a range of possible destinations and warned that landfall was most likely in Mexico. This forecast was based on global observations from aircraft, buoys and satellites, transformed into forecasts by large supercomputers.
That same day, experts using artificial intelligence software on a much smaller computer predicted that the hurricane would make landfall in Texas. This forecast was based solely on the prior knowledge the machine had acquired about the planet’s atmosphere.
Four days later, on July 8, Hurricane Beryl struck Texas with deadly force, flooding roads, killing at least 36 people and leaving millions without power. In Houston, high winds toppled trees onto homes, killing at least two people.
A satellite view of Hurricane Beryl approaching the Texas coast on July 8. NOAA, via European Press Agency, via Shutterstock
The Texas prediction illustrates the new landscape of AI-powered weather forecasting, in which a growing number of smart machines are anticipating future weather patterns with impressive speed and accuracy. In this case, the experimental program GraphCast, created in London by DeepMind, a Google company, performed calculations in seconds that previously took hours.
“It’s a really exciting step,” said Matthew Chantry, an artificial intelligence specialist at the European Centre for Medium-Range Weather Forecasts, the agency that was eclipsed by the Beryl forecast. Chantry said GraphCast and similar models could outperform his agency in predicting hurricane paths.
Overall, ultrafast AI excels at detecting future hazards, according to Christopher S. Bretherton, professor emeritus of atmospheric sciences at the University of Washington. For heat, wind and torrential rain, warnings will be “more up-to-date than they are now,” potentially saving countless lives.
In addition, AI-powered rapid weather forecasts also drive scientific discovery, according to Amy McGovern, a professor of meteorology and computer science at the University of Oklahoma who runs an institute for AI meteorology. McGovern said meteorologists now use AI to create thousands of subtle variations in forecasts, allowing them to identify unexpected factors that can trigger extreme events like tornadoes.
“It allows us to look for fundamental processes,” said Dr. McGovern. “It’s a valuable tool for discovering new things.”
A significant advantage of AI models is that they can run on desktop computers, making them easier to adopt compared to the room-sized supercomputers that currently dominate the realm of global predictions.
“It’s a game-changer,” said Maria Molina, a research meteorologist at the University of Maryland who studies AI programs for predicting extreme events. “You don’t need a supercomputer to generate a forecast. You can do it on a laptop, which makes the science more accessible.”
People rely on accurate weather forecasts to make decisions about what to wear, where to travel and whether to evacuate in the event of a violent storm.
Still, obtaining reliable weather forecasts remains extremely difficult due to the complexity of weather patterns. Unlike astronomers, who can predict the trajectories of the planets due to the dominance of the sun, Earth’s weather patterns arise from a variety of factors and are inherently chaotic. This causes forecasts to become less reliable over time, although they now extend out to ten days, compared with three days a few decades ago.
Slow progress is due to improved global observations and supercomputer capabilities, which require considerable skill and effort to build accurate models of the planet, filling data gaps with current observations.
Dr Bretherton, from the University of Washington, stressed the importance of combining data from multiple sources to estimate the current state of the atmosphere. “You have to combine data from many sources to make an estimate of what the atmosphere is like right now,” he said.
Complicated equations in fluid mechanics turn the combined observations into predictions. Despite the power of supercomputers, the analysis can take more than an hour, and forecasts must be continually updated as the climate changes.
The AI approach is radically different. Instead of relying on current readings and extensive calculations, AI learns from the causal relationships that govern the planet’s climate. The revolution in machine learning, a branch of AI that mimics human learning, allows AI to recognize patterns with great success, applying this ability to weather forecasting.
Recently, the DeepMind team that developed GraphCast won Britain’s prestigious engineering prize awarded by the Royal Academy of Engineering. Sir Richard Friend, a physicist at the University of Cambridge and chair of the judging panel, praised the team for what he called “a revolutionary breakthrough.”
In an interview, Rémi Lam, GraphCast’s chief scientist, explained that his team trained the program on four decades of global weather observations from the European forecasting center. “It learns directly from historical data,” he said. In seconds, he added, GraphCast can generate a 10-day forecast that would take a supercomputer more than an hour.
Dr. Lam said GraphCast works best on computers specifically designed for AI, but can also run on desktops and even laptops, albeit at a slower speed.
In tests, Dr Lam reported that GraphCast outperformed the best model from the European Centre for Medium-Range Weather Forecasts in more than 90% of cases. “Knowing where a cyclone is heading is crucial,” he said. “It is important to save lives.”
A view of a severely damaged home with debris strewn across the yard after the hurricane. Brandon Bell/Getty Images
In response to a question, Dr Lam said his team, although comprised of computer scientists rather than cyclone experts, had not assessed how GraphCast’s predictions for Hurricane Beryl compared to other forecasts in terms of accuracy.
However, DeepMind did conduct a study on Hurricane Lee, an Atlantic storm that was considered to be threatening New England or Canada in September. Dr Lam said the study found that GraphCast had predicted the landfall point in Nova Scotia three days before supercomputers came to the same conclusion.
Impressed by these achievements, the European center recently adopted GraphCast, as well as AI programs developed by Nvidia, Huawei and Fudan University in China. On its website, it now displays global maps of its AI tests, including trajectory forecasts for Hurricane Beryl on July 4.
DeepMind’s GraphCast projected path, labeled DMGC on the July 4 map, showed Beryl making landfall near Corpus Christi, Texas, not far from where the hurricane actually hit.
Dr Chantry, from the European centre, said the institution believes the experimental technology will become a regular part of global weather forecasting, including for cyclones. A new team is working to create an operational AI system for the agency.
Dr. Chantry predicts adoption could happen soon, but noted that AI technology could coexist with the center’s traditional forecasting system.
#Artificial #intelligence #offers #advantage #meteorologists
