When then Tropical Storm Melissa was churning south of Haiti, Philippe Papin, a National Hurricane Center (NHC) meteorologist, had confidence it was about to grow into a monster hurricane.
As the lead forecaster on duty, he predicted that in just 24 hours the storm would become a category 4 hurricane and begin a turn towards the coast of Jamaica. No NHC forecaster had ever issued such a bold forecast for rapid strengthening.
But Papin had an ace up his sleeve: artificial intelligence in the form of Google’s new DeepMind hurricane model – released for the first time in June. And, as predicted, Melissa did become a storm of astonishing strength that tore through Jamaica.
Forecasters at the NHC are increasingly leaning hard on Google DeepMind. On the morning of 25 October, Papin explained in his public discussion and on social media that Google’s model was a primary reason he was so confident: “Roughly 40/50 Google DeepMind ensemble members show Melissa becoming a Category 5. While I am not ready to forecast that intensity yet given the track uncertainty, that remains a possibility.
“It appears likely that a period of rapid intensification will occur as the storm moves slowly over very warm ocean waters which is the highest oceanic heat content in the entire Atlantic basin.”
Google DeepMind is the first AI model dedicated to hurricanes, and now the first to beat traditional weather forecasters at their own game. Through all 13 Atlantic storms so far this year, Google’s model is the best – even beating human forecasters on track predictions.
Melissa eventually made landfall in Jamaica at category 5 strength, one of the strongest landfalls ever documented in nearly two centuries of record-keeping across the Atlantic basin. Papin’s bold forecast likely gave people in Jamaica extra time to prepare for the disaster, possibly saving lives and property.
Google DeepMind has been making weather forecasts for a few years now, and the parent forecast system from which the new hurricane model is derived also performed spectacularly well in diagnosing large-scale weather patterns last year.
Google’s model works by spotting patterns that traditional time-intensive physics-based weather models may miss.
“They do it much more quickly than their physics-based cousins, and the computing power is less expensive and time consuming,” Michael Lowry, a former NHC forecaster, said.
“What this hurricane season has proven in short order is that the newcomer AI weather models are competitive with and, in some cases, more accurate than the slower physics-based weather models we’ve traditionally leaned on,” Lowry said.
To be sure, Google DeepMind is an example of machine learning – a technique that has been used in data-heavy sciences like meteorology for years – and is not generative AI like ChatGPT.
Machine learning takes mounds of data and pulls out patterns from them in a such a way that its model only takes a few minutes to come up with an answer, and can do so on a desktop computer – in strong contrast to the flagship models that governments have used for decades that can take hours to run and require some of the biggest supercomputers in the world.
Still, the fact that Google’s model could outperform previous gold-standard legacy models so quickly is nothing short of amazing to meteorologists who have spent their careers trying to forecast the world’s strongest storms.
“I’m impressed,” said James Franklin, a retired NHC forecaster. “The sample is now large enough that it’s pretty clear this is not a case of beginner’s luck.”
Franklin said that although Google DeepMind is beating all other models on forecasting the future path of hurricanes worldwide this year, like many AI models it occasionally gets high-end intensity forecasts wrong. It struggled with Hurricane Erin earlier this year, as it was also undergoing rapid intensification to category 5 north of the Caribbean. It also struggled with Typhoon Kalmaegi – which made landfall in the Philippines on Monday.
In the coming offseason, Franklin said he plans to talk with Google about how it can make the DeepMind output even more helpful for forecasters by providing additional under-the-hood data they can use to assess exactly why it is coming up with the its answers.
“The one thing that nags at me is that while these forecasts seem to be really, really good, the output of the model is kind of a black box,” said Franklin.
There has never been a private, for-profit company that has produced a top-level weather model which allows researchers a peek into its methods – unlike nearly all other models which are provided free to the public in their entirety by the governments that designed and maintain them. While Google has made top-level output of DeepMind publicly available in real time on a dedicated website, its methods have still largely been hidden.
Google is not alone in starting to use AI to solve difficult weather forecasting problems. The US and European governments also have their own AI weather models in the works – which have also shown improved skill over previous non-AI versions.
The next steps in AI weather forecasts seem to be startup companies taking swings at previously tough-to-solve problems such as sub-seasonal outlooks and better advance warnings of tornado outbreaks and flash flooding – and they are receiving US government funding to do so. One company, WindBorne Systems, is even launching its own weather balloons to fill the gaps in the US weather-observing network, which has recently been downsized by the Trump administration.
