Google DeepMind has unveiled WeatherNext 2, a major upgrade to its AI-driven forecasting technology that the company says can generate accurate global weather predictions in a fraction of the time required by traditional numerical models. The system, announced on Monday, represents a significant step forward in how agencies and everyday users receive updates on severe conditions as climate-related disasters continue to intensify worldwide.
WeatherNext 2 can produce hundreds of global forecast scenarios from a single starting point, with each simulation completed in less than a minute using one Tensor Processing Unit, Google’s custom chip for AI workloads. This speed dramatically increases the frequency at which forecasters can refresh predictions, a limitation that has long hindered conventional modelling systems, which often take several hours per run.
Peter Battaglia, a research scientist at Google DeepMind, noted on X that accurate weather intelligence supports high-stakes decisions affecting food systems, supply chains, energy grids, and public safety. With AI accelerating the forecasting process, he said, the industry is undergoing a profound transformation.
Google has already integrated WeatherNext 2 into Search results, the Gemini assistant, Pixel Weather, and weather layers within Google Maps. Akib Uddin, who oversees the product, said a wider rollout is underway across the company’s ecosystem. He emphasised that improving forecast precision at scale has real-world benefits, from daily routines to disaster readiness.
DeepMind said WeatherNext 2 is substantially faster and more precise than last year’s WeatherNext Gen. The latest version operates at six times the resolution and updates in one-hour increments instead of six, producing more reliable estimates for temperature, wind, humidity, and pressure across nearly the entire globe through a 15-day window. Battaglia added that the system outperformed its predecessor on almost every variable measured.
The accuracy upgrade stems from a new modelling framework introduced in a research paper earlier this year called Functional Generative Networks, or FGN. Unlike other systems that require extensive multivariable datasets, FGN is trained on single-variable predictions at individual locations. Even so, it learns how different atmospheric factors interact and evolve together, enabling it to capture regional phenomena such as heat domes, wind extremes, and cyclone development.
According to Google’s evaluation, FGN matches earlier models on extreme temperature forecasting and significantly improves performance on extreme wind events. Using standard scoring metrics like the Continuous Ranked Probability Score, WeatherNext 2 achieved accuracy gains of up to 8.7% in large-area assessments.
One of the standout improvements lies in cyclone forecasting. When compared with historical storm tracks, the model demonstrated an average 24-hour advantage in accuracy for three- to five-day predictions. A coarser version running at 12-hour timesteps still exceeded the performance of the previous GenCast system, especially at longer lead times.
DeepMind has already shared experimental cyclone tools with global weather agencies. Uddin said the goal is simple: faster, more precise forecasts that give communities and responders more time to act as extreme weather continues to escalate.
