Google DeepMindの「WeatherNext 2」AIモデルがサイクロン予測で大幅な精度向上を達成Our WeatherNext 2 AI model demonstrated a massive leap forward in predicting cyclones.

Google DeepMind has unveiled WeatherNext 2, the newest generation of its artificial-intelligence weather forecasting system, and reports that the model delivers state-of-the-art accuracy in cyclone prediction. The development is significant because tropical cyclones—called hurricanes, typhoons or simply cyclones depending on the region—rank among the deadliest and costliest natural disasters, and improvements in forecasting their movement and strength can give communities and emergency planners more time to respond.

The company frames the advance primarily around two elements that determine how useful a cyclone forecast is: track and intensity. Track forecasting predicts the path a storm will follow, while intensity forecasting estimates how powerful it will become, usually expressed through maximum sustained winds and central pressure. Intensity has long been considered one of the harder problems in meteorology, since rapid intensification can occur over just hours and has historically challenged both physics-based models and earlier AI approaches. WeatherNext 2 appears to target these weaknesses, positioning itself as an improvement over its predecessor and over conventional methods.

Modern AI weather models differ fundamentally from traditional numerical weather prediction, or NWP, which powers most operational forecasts today. NWP works by solving the physical equations that govern the atmosphere on supercomputers, a process that is accurate but computationally intensive and time-consuming. Machine-learning models such as WeatherNext instead learn statistical patterns from decades of historical weather records—commonly drawing on reanalysis datasets like ECMWF's ERA5—and can then produce forecasts in a fraction of the time and cost once trained. That efficiency allows forecasters to generate more scenarios, run larger ensembles, or update predictions more frequently.

WeatherNext builds on a lineage of Google research in this field. Its earlier work includes GraphCast, a deterministic model published in the journal Science, and GenCast, a probabilistic ensemble model published in Nature that produces a range of possible outcomes rather than a single forecast. Ensemble and probabilistic techniques are particularly valuable for cyclones, where communicating uncertainty—such as the "cone" of a storm's possible track—is essential to sound decision-making. Google has been packaging these capabilities under the WeatherNext brand and making them accessible through its cloud and developer tools, including Google Cloud and platforms such as Vertex AI and Earth Engine, so that researchers and organizations can integrate the forecasts into their own systems.

The broader context is a rapid shift across the weather industry toward AI-driven forecasting. In recent years, technology companies and research institutions have released a series of competing models, including Nvidia's FourCastNet, Huawei's Pangu-Weather and Microsoft's Aurora, while established forecasting bodies have begun adopting the approach themselves. The European Centre for Medium-Range Weather Forecasts, long regarded as a global leader in NWP, has developed its