Google DeepMind and Google Research introduced WeatherNext 2, their most advanced weather forecasting AI model, on a specified date. The model delivers 8x faster global forecasts at up to 1-hour resolution by generating hundreds of scenarios from a single input through noise injection in function space. This supports weather agencies with experimental cyclone predictions and integrates into Google products.
Weather influences daily decisions across global supply chains, flight paths, and personal commutes. Artificial intelligence has expanded capabilities in weather forecasting over recent years. WeatherNext 2 represents the latest advancement from the WeatherNext team, focusing on efficiency, accuracy, and resolution in predictions worldwide.
The model generates forecasts eight times faster than predecessors while achieving hourly resolution. This speed stems from processing on a single Tensor Processing Unit (TPU), where each scenario prediction completes in less than one minute. In contrast, physics-based models on supercomputers require hours for equivalent computations. WeatherNext 2 has already aided weather agencies by providing experimental cyclone predictions based on scenario ranges, enabling decisions informed by multiple outcomes.
Forecast data from WeatherNext 2 now resides in Earth Engine and BigQuery, allowing users direct access to these datasets. Google launched an early access program on Google Cloud’s Vertex AI platform, permitting custom model inference for participants. These steps move the technology from research labs into practical applications for broader utilization.
WeatherNext technology now enhances forecasts within Google Search, Gemini, Pixel Weather, and Google Maps Platform’s Weather API. In the coming weeks, WeatherNext 2 will power weather information displayed in Google Maps, extending its reach to mobile navigation and planning tools used by millions daily.
From one initial input, WeatherNext 2 employs independently trained neural networks and injects noise directly into function space. This method produces coherent variability across hundreds of possible weather outcomes, capturing the full spectrum of possibilities. Such coverage proves essential for planning around worst-case scenarios, which demand precise preparation in meteorology and beyond.
WeatherNext 2 outperforms the prior WeatherNext model across 99.9% of variables, including temperature, wind, and humidity, and all lead times from zero to 15 days. These metrics reflect higher skill levels and finer hourly resolution. The Continuous Ranked Probability Score (CRPS) comparisons confirm these gains against WeatherNext Gen, quantifying superior probabilistic accuracy in ensemble predictions.
Central to these improvements is the Functional Generative Network (FGN), a new AI modeling approach. The FGN injects noise into the model architecture itself, ensuring generated forecasts maintain physical realism and interconnections between variables. This architectural innovation preserves spatial and temporal consistency, distinguishing it from traditional noise addition techniques.
Video: Google
The model excels in forecasting both marginals and joints. Marginals encompass individual weather elements, such as precise temperature at a specific location, wind speed at a designated altitude, or humidity levels at a given point. Training occurs exclusively on these marginals. Despite this limited scope, the model acquires the ability to predict joints accurately—complex, interconnected systems reliant on interactions among multiple elements.
Joints include predictions for entire regions impacted by high heat or aggregated power output from wind farms spanning large areas. These forecasts depend on how individual marginals combine dynamically. WeatherNext 2’s capacity to derive joint distributions from marginal training data marks a key technical achievement, enabling applications in energy production, agriculture, and disaster management that require holistic system views.
Development of WeatherNext 2 translates research into operational tools. Google DeepMind and Google Research commit to advancing model capabilities through integration of new data sources. Plans encompass further expansion of access to these tools. Provision of open data and platforms aims to support researchers, developers, and businesses. Users can explore related geospatial and AI initiatives via Google Earth, Earth Engine, AlphaEarth Foundations, and Earth AI for deeper context on these efforts.
