How Google’s DeepMind tool is ‘more quickly’ forecasting hurricane behavior | Google

As Tropical Storm Melissa churned south of Haiti, National Hurricane Center (NHC) meteorologist Philippe Papin was confident it was about to become a monster hurricane.

As the chief forecaster on duty, he predicted that the storm would strengthen into a category 4 hurricane in just 24 hours and begin turning toward the coast of Jamaica. No NHC forecaster has ever published this. a very bold prediction For fast strengthening.

But Papin had an ace up his sleeve: artificial intelligence in the form of Google’s new DeepMind hurricane model, which was first released in June. And as predicted, Melissa became a storm of surprising power that wreaked havoc on Jamaica.

Forecasters at NHC are increasingly relying on Google DeepMind. On the morning of October 25, Papin declared in his public discussion: and on social media HE The main reason why he was so confident was Google’s model: “About 40/50 Google DeepMind community members indicate Melissa falls into Category 5. While I’m not yet ready to predict that intensity, given the uncertainty in the track, it remains a possibility.”

“A period of rapid intensification appears likely to occur as the storm slowly moves over very warm ocean waters, the highest ocean heat content in the entire Atlantic basin.”

Google DeepMind first artificial intelligence model dedicated to hurricanesand has now become the first to beat traditional weather forecasters at their own game. All 13 Atlantic storms so far this yearGoogle’s model is the best; it even outperforms human forecasters in accurate predictions.

Melissa eventually made landfall in Jamaica at category 5 strength; This is one of the most powerful landfalls ever documented in nearly two centuries of record-keeping in the Atlantic basin. Papin’s bold prediction likely gave people in Jamaica extra time to prepare for the disaster and likely saved lives and property.

Google DeepMind makes weather forecasts for several yearsand the main forecasting system from which the new hurricane model is derived also performed exceptionally well at diagnosing large-scale weather patterns last year.

Google’s model works by detecting patterns that traditional time-intensive physics-based weather models might miss.

“They do this much faster than their physics-based cousins, and the computing power is less expensive and time consuming,” said Michael Lowry, a former NHC forecaster.

“What this hurricane season has quickly proven is that incoming AI weather models can compete with, and in some cases are more accurate than, the slower physics-based weather models we have traditionally relied on,” Lowry said.

Of course, Google DeepMind is an example of machine learning, a technique that has been used for years in data-intensive sciences such as meteorology, and is not a generative AI like ChatGPT.

Machine learning takes masses of data and extracts patterns from it; so it only takes a few minutes for his model to come up with an answer, and he can do it on a desktop computer; This is a strong contrast to the flagship models that governments have used for decades, which can take hours and hours to run. some of the world’s largest supercomputers.

Still, the fact that Google’s model was able to so quickly outperform previous gold-standard legacy models is nothing short of surprising to meteorologists who have spent their careers predicting the world’s strongest storms.

“I’m impressed,” said James Franklin, a retired NHC forecaster. “The sample is large enough now that it’s pretty clear that this is not a beginner’s chance.”

Although Google DeepMind outperformed all other models this year in predicting the future path of hurricanes around the world, Franklin said, As with many AI models, high-level density predictions can sometimes be inaccurate. Hurricane Erin, which rapidly intensified to category 5 in the north of the Caribbean, also struggled with difficulties at the beginning of this year. He also battled Typhoon Kalmaegi – Landed in the Philippines on Monday.

Next offseason, Franklin said he plans to talk to Google about how to make DeepMind output more useful to forecasters by providing additional confidential data they can use to evaluate exactly why it found the answers.

“What bothers me is that even though these predictions look really good, the output of the model is kind of a black box,” Franklin said.

Unlike almost all other models that are made available in their entirety free to the public by the governments that design and maintain them, there has never been a private for-profit company that produces a high-end weather model that allows researchers to peek into its methods. Google takes DeepMind’s high-level output publicly available in real time on a dedicated websitetheir methods are still largely hidden.

Google isn’t the only one starting to use AI to solve tough weather forecasting problems. The US and European governments also have their own AI weather models in the works. improved skill over previous non-AI versions.

The next steps in AI weather forecasting look like startups are turning to sub-seasonal outlooks and problems that have previously been difficult to solve. better advance warnings of tornado outbreaks and flash flooding – and they they receive funding from the US government to do this. Even a company called WindBorne Systems launches his own weather balloon To fill gaps in the US weather observation network, which was recently scaled back by the Trump administration.