How artificial intelligence is speeding up predictions of hurricanes and global weather patterns

As Hurricane Beryl tore through the Caribbean in early July, Europe's top weather agency expected A series of final landfalls, warning that landfall in Mexico is most likely. The alerts are based on global observations from aircraft, buoys and spacecraft, which are then turned into predictions by room-sized supercomputers.

That same day, experts running artificial intelligence software on a much smaller computer predicted that the storm would make landfall in Texas. The prediction is based solely on what the machine previously knew about Earth's atmosphere.

Four days later, on July 8, Hurricane Beryl hit Texas with deadly force, flooding roads, killing at least 36 people and leaving millions of residents without power. In Houston, at least two victims were killed when high winds blew trees into homes.

The Texas forecast offers a glimpse into the emerging world of artificial intelligence weather forecasting, where increasingly intelligent machines are predicting future global weather patterns with new speed and accuracy. In this case, the experimental program is GraphCast, created by DeepMind, a Google company in London. It takes just minutes and seconds to do what used to take hours.

“This is a really exciting step,” said Matthew Chantry, an artificial intelligence expert at the European Center for Medium-Range Weather Forecasts, whose Beryl forecast stole the show. He added that, on average, GraphCast and its smarter cousins ​​outperformed his agency at predicting hurricane paths.

Christopher S. Bretherton, professor emeritus of atmospheric science at the University of Washington, said that in general, ultra-fast artificial intelligence can play a role in detecting upcoming dangers. He said the usual warnings for dangerous heat, high winds and heavy downpours would be “much more timely than they are now”, saving countless lives.

Rapid AI weather forecasting will also aid scientific discovery, said Amy McGovern, a professor of meteorology and computer science at the University of Oklahoma and director of the Artificial Intelligence Weather Institute. Weather detectives now use artificial intelligence to create thousands of subtle changes in forecasts, allowing them to find unexpected factors that could trigger extreme events like tornadoes, she said.

“It lets us look for fundamental processes,” Dr. McGovern said. “It's a valuable tool for discovering new things.”

Importantly, AI models can be run on desktop computers, making the technology easier to adopt than the room-sized supercomputers that now dominate the world of global predictions.

“This is a turning point,” said Maria Molina, a meteorologist at the University of Maryland who studies artificial intelligence programs for predicting extreme events. “You don’t need a supercomputer to generate predictions. You can do it on a laptop, which makes science easier to understand.

People rely on accurate weather forecasts to make decisions about things like how to dress, where to travel, and whether to flee severe storms.

Even so, reliable weather forecasts remain elusive. The problem is complexity. Astronomers can predict the paths of the solar system's planets for centuries to come because one factor dominates their motion – the sun and its immense gravitational pull.

In contrast, weather patterns on Earth are caused by a variety of factors. The Earth's tilt, rotation, wobble, and day-night cycle turn the atmosphere into a turbulent maelstrom of wind, rain, clouds, temperature, and pressure. To make matters worse, the atmosphere was already chaotic. For their part, specific areas can quickly go from stable to erratic in the absence of external stimulation.

Therefore, weather forecasts may become invalid days or even hours later. Errors increase as forecast times get longer – today's forecast times can be as long as ten days, up from three a few decades ago. Slow improvements come from improvements in global observations and the supercomputers that make predictions.

This is not to say that supercomputing work has become easier. Preparation requires skill and labor. Modelers constructed a virtual planet crisscrossed by millions of data gaps and filled the empty spaces with current weather observations.

The University of Washington's Dr. Bretherton said the input was critical and somewhat improvised. “You have to fuse data from multiple sources to make a guess about the current atmospheric conditions,” he said.

Complex fluid dynamics equations then transform the mixed observations into predictions. Despite the vast power of supercomputers, number crunching can take an hour or more. Of course, as the weather changes, the weather forecast must be updated.

Artificial intelligence’s approach is completely different. Instead of relying on current readings and millions of calculations, artificial intelligence agents use what they know about the cause-and-effect relationships that control Earth's weather.

Overall, this advancement stems from an ongoing revolution in machine learning—a branch of artificial intelligence that mimics the way humans learn. This approach has been a huge success because artificial intelligence is good at pattern recognition. It can quickly sort through mountains of information and uncover complex content that humans cannot discern. Doing so has led to breakthroughs in speech recognition, drug discovery, computer vision and cancer detection.

In weather forecasting, artificial intelligence scans a database of real-world observations to understand atmospheric forces. It then identifies subtle patterns and uses this knowledge to predict weather with remarkable speed and accuracy.

Recently, the DeepMind team that built GraphCast won the UK's highest engineering award from the Royal Academy of Engineering. Jury chairman Sir Richard Friend, a physicist at the University of Cambridge, praised the team's “revolutionary progress”.

Rémi Lam, GraphCast's chief scientist, said in an interview that his team has trained the artificial intelligence program based on four decades of global weather observations compiled by the European Forecast Center. “It learns directly from historical data,” he said. GraphCast can produce a 10-day forecast in seconds, which would take a supercomputer more than an hour to complete, he added.

Dr. Lam said GraphCast runs best and fastest on computers designed for artificial intelligence, but can also run on desktops and even laptops, albeit at a slower speed.

Dr. Lam reported that in a series of tests, GraphCast outperformed the European Center for Medium-Range Weather Forecasts' best forecast model more than 90 percent of the time. “If you know where the cyclone is going, that's very important,” he added. “This is important to saving lives.”

In response to questions, Dr. Lam said that he and his team were computer scientists, not cyclone experts, and had not evaluated how accurate GraphCast's predictions for Hurricane Beryl were compared to other predictions.

But he added that DeepMind did conduct research on Hurricane Lee, an Atlantic storm thought to potentially threaten New England or further east into Canada in September. Dr. Lam said the study found GraphCast targeted the Nova Scotia landfall three days before the supercomputer came to the same conclusion.

Impressed by these achievements, the European center recently adopted GraphCast and an AI prediction project developed by Nvidia, Huawei and Fudan University in China. Now, the company is showing off a global map of its artificial intelligence tests on its website, including the range of predicted paths the intelligent machines made for Hurricane Beryl on July 4.

The path predicted by DeepMind's GraphCast (labeled DMGC on the July 4 map) shows Beryl making landfall in the Corpus Christi, Texas area, not far from where the hurricane actually hit.

Dr Chantry of the European Center said the agency believed the experimental technique was becoming a routine part of global weather forecasting, including cyclones. He added that a new team is now building on the “great work” of experimentalists to create an operational artificial intelligence system for the agency.

Its adoption could happen quickly, Dr. Chantry said. However, he added that AI technology as a routine tool could coexist with the center's legacy forecasting system.

Dr. Bretherton, now a team leader at the Allen Institute for Artificial Intelligence (founded by Microsoft co-founder Paul G. Allen), said the European center is considered the world's top weather agency because comparative tests often show its Predictions were shown to outperform all others in accuracy. As a result, he added, its interest in artificial intelligence allowed meteorologists to “see this and say, 'Hey, we have to match this.'”

Weather experts say artificial intelligence systems may complement supercomputer methods because each method has its own unique advantages.

“All models are wrong to some degree,” said Dr. Molina of the University of Maryland. She added that AI machines “may be able to predict hurricane paths accurately, but what about rainfall, maximum winds and storm surge? There are so many different impacts” that need to be predicted reliably and carefully evaluated.

Even so, Dr. Molina noted that AI scientists are still rushing to publish papers demonstrating new predictive skills. “The revolution continues,” she said. “It's wild.”

Jamie Rhome, deputy director of the National Hurricane Center in Miami, agreed that a variety of tools are needed. He called artificial intelligence “an evolution rather than a revolution” and predicted that humans and supercomputers will continue to play important roles.

“Having a human on the table applying situational awareness is one of the reasons we have such high accuracy,” he said.

Mr. Roem added that the hurricane center has been using aspects of artificial intelligence in forecasting for more than a decade, and the agency will evaluate and potentially exploit these clever new initiatives.

“With the rapid development of artificial intelligence, many people believe that humans’ role is diminishing,” Mr. Roem added. “But our forecasters are making a huge contribution. The human role is still very important.