Google DeepMind、最先端の気象予測モデルWeatherNext 2を発表 WeatherNext 2: Our most advanced weather forecasting model

Google DeepMind has unveiled WeatherNext 2, the latest iteration of its AI-driven weather forecasting system. The new model produces hourly forecasts up to eight times faster than its predecessor and reportedly outperforms it on the majority of key variables, including temperature, wind, humidity and geopotential.

At the heart of WeatherNext 2 is a new technique the team calls a Functional Generative Network (FGN). Rather than perturbing inputs as traditional ensemble forecasting does, FGN injects noise directly into the model's weight space, enabling a single model to generate hundreds of plausible weather scenarios efficiently. DeepMind says this design improves the calibration of probabilistic forecasts for extreme events such as heatwaves, cold snaps and high winds — the kind of tail risks that matter most to emergency planners and energy markets.

The model's outputs will be rolled into a range of Google products, including Search, Maps, Gemini and Pixel Weather, exposing AI-generated forecasts to a mass consumer audience. For researchers and enterprise users, real-time predictions will be made available through Google Cloud's Earth Engine and BigQuery, as well as via Vertex AI, continuing Google's strategy of dual-tracking consumer integration and developer access.

WeatherNext 2 lands in a field that has been transformed by machine learning over the past two years. Numerical weather prediction, long the domain of physics-based simulations running on national supercomputers, is increasingly being challenged — and in some metrics surpassed — by neural approaches. ECMWF has released its AIFS system, NVIDIA has promoted FourCastNet, and Huawei's Pangu-Weather demonstrated that transformer-based models could match operational forecasts at a fraction of the compute cost. DeepMind itself broke ground with GraphCast in 2023 and the diffusion-based GenCast last year; WeatherNext 2 can be read as the consolidation of that lineage into a productized offering.

The shift toward generative ensembles is particularly notable. Traditional ensemble forecasts require running expensive physical simulations dozens of times with slightly different initial conditions. A generative model that can sample many futures from learned weather dynamics offers similar uncertainty quantification at a fraction of the runtime, which could allow forecast centers to issue more frequent updates or run higher-resolution scenarios.

Some caveats remain. AI weather models are typically trained on reanalysis datasets such as ERA5, meaning their behaviour on truly unprecedented events — the kind climate change is making more common — may be harder to validate. Researchers have also raised questions about whether learned models faithfully capture rare dynamical regimes or simply interpolate plausible-looking fields. DeepMind's emphasis on probabilistic skill and extreme-event calibration suggests it is aware of these concerns, though independent evaluation by national meteorological agencies will likely determine how quickly such systems are adopted operationally.