Google DeepMind has unveiled a new artificial intelligence model that it claims can forecast the path and intensity of tropical cyclones with an extra day of lead time compared to the best existing systems, a leap the researchers equate to a decade of progress in meteorology.
The model, named WeatherNext Cyclones, was detailed in a study published in the journal Nature on Thursday. It was developed in collaboration with Google Research, the U.S. National Hurricane Center (NHC), the Cooperative Institute for Research in the Atmosphere (CIRA), the UK Met Office, and other weather agencies worldwide. Google is releasing both the code and model weights for WeatherNext Cyclones and the broader WeatherNext 2 system on GitHub, opening the technology to the global research community.
The core breakthrough is a shift in the forecast timeline. According to the research team, a three-day prediction from WeatherNext Cyclones is now as accurate as a two-day forecast from prior state-of-the-art models. On the ground, that additional 24-hour window is critical for emergency managers staging supplies, organizing evacuations, and positioning response teams before a storm makes landfall.
“Time is really golden when it comes to those types of decisions, so the ability to push forecast accuracy out as much as a day beyond what we’ve previously been able to do is really valuable,” said Mike Brennan, director of the NHC.
Tropical cyclones—known as hurricanes in the Atlantic and typhoons in the western Pacific—have long presented a unique forecasting challenge because they operate across vastly different spatial scales. Predicting a storm’s track requires analyzing global atmospheric currents, a task traditionally handled by coarser global models. Predicting intensity, however, depends on fine-scale thermodynamics at the storm’s core, which has typically required specialized high-resolution local models.
WeatherNext Cyclones bridges this divide by training a single model end-to-end on nearly 20 terabytes of global atmospheric data combined with historical information from the International Best Track Archive for Climate Stewardship (IBTrACS), which contains records of nearly 5,000 historical storms. The result is an AI system that generates a complete 15-day forecast—covering track, intensity, and wind structure—in under a minute on a single Google TPU chip.
“We don’t have that much cyclone data, but we have a lot of weather data,” said Ferran Alet, a research scientist at Google DeepMind and a lead author on the Nature paper. “So what we did was train a model to be both good at weather as well as cyclones.”
Perhaps the most surprising aspect of the model is its resolution. WeatherNext Cyclones operates on atmospheric data at a 28-by-28-kilometer grid, roughly 100 times coarser than the high-resolution inputs traditionally required for intensity forecasting. A smaller variant, WeatherNext 2-mini, runs at 111-by-111 kilometers and still delivers strong results. The research community, Alet noted, was “shocked” that such coarse inputs could capture signals about impending intensity changes.
“It’s a black box at the end of the day,” Alet said, “but that gives physicists a signal that something is happening that was not previously understood.”
The model’s real-world credibility was bolstered during the 2025 Atlantic hurricane season. In October of that year, WeatherNext flagged an 80 percent probability five days in advance that a developing storm system would intensify into a Category 5 hurricane and strike Jamaica. That storm became Hurricane Melissa, which caused catastrophic flooding and landslides across the island nation. It marked the first time forecasters at the NHC were able to predict a Category 5 hurricane when the storm was still at Category 1 strength.
Kate Musgrave, tropical cyclone group lead at CIRA and a co-author on the paper, said researchers were initially skeptical that the strong retrospective test results would hold up in live operations. “I think everybody was surprised at just how well it did,” she said.
This year, the collaboration is scaling up. While the 2025 season saw the model generating 50 potential scenarios per cyclone to capture uncertainty, the 2026 operational version produces 1,000 ensemble members. That expanded output is designed to better surface rare but devastating tail-risk events, such as rapid intensification overnight.
“That’s something that, with our computing power, we simply can’t do with our existing numerical models,” Musgrave said.
Google is making the technology broadly accessible. Alongside the Nature publication, the company is open-sourcing WeatherNext Cyclones, the WeatherNext 2 operational model, and a Colab notebook that allows anyone to run WeatherNext 2-mini on a single TPU for free. A visualization tool called Weather Lab, part of Google Earth AI, lets users explore cyclone and global weather views.
Google DeepMind said the open-source release is intended to accelerate research, help specialized local models, and equip regional forecasters and nonprofits with modern tools. The company emphasized, however, that official watches and warnings remain the sole responsibility of national and local weather agencies.
Brennan echoed that point, cautioning that no single model is a silver bullet. “There’s no guarantee that one model, because it did well last year or really did well for this particular storm, is necessarily going to be the best model for the next season or the next storm,” he said. “A hurricane is not just a track or an intensity forecast. It requires experts to translate that into what the impacts are going to be—and it’s the impacts that kill people.”
Tropical cyclones have caused more than $1.4 trillion in economic losses and over 700,000 deaths globally over the past 50 years, underscoring the stakes of even marginal improvements in forecast lead time. For Alet and the research team, the open-source release is also a bet on scientific discovery. “I think AI is giving us new tools to poke into the laws of the universe,” he said.