WeatherNext 2 release notes / what's new

Google DeepMind and Google Research have introduced WeatherNext 2, a weather forecasting model that generates predictions 8x faster than previous iterations and provides resolution down to 1-hour. This model enables the generation of hundreds of possible weather scenarios from a single input, significantly reducing the time required to produce forecasts compared to traditional physics-based supercomputer models.

Functional Generative Network (FGN) Architecture

WeatherNext 2 utilizes a new AI modeling approach called a Functional Generative Network (FGN). This architecture injects noise directly into the model's function space to create coherent variability in predictions while ensuring the forecasts remain physically realistic and interconnected.

Marginal and Joint Forecasting

The FGN approach allows the model to be trained on "marginals"—individual, standalone weather elements such as humidity, wind speed at a specific altitude, or temperature at a specific location. Despite being trained only on these marginals, the model learns to forecast "joints," which are large, complex, interconnected systems. This capability is critical for predicting regional effects, such as identifying areas affected by high heat or estimating power output across an entire wind farm.

Performance and Accuracy Benchmarks

WeatherNext 2 demonstrates significant improvements over the previous state-of-the-art WeatherNext model:

  • Variable Accuracy: The model surpasses its predecessor on 99.9% of variables, including wind, temperature, and humidity.
  • Lead Times: Improvements are observed across lead times ranging from 0 to 15 days.
  • Inference Speed: A single prediction takes less than one minute on a single TPU, whereas physics-based models on supercomputers typically require hours.

Integration and Availability

Google is integrating WeatherNext 2 technology into several consumer and developer products to move the research from the lab to real-world application:

Consumer Product Integration

Weather forecasts in the following services have been upgraded:

  • Google Search
  • Gemini
  • Pixel Weather
  • Google Maps Platform Weather API
  • Google Maps (integration coming in the following weeks)

Developer and Research Access

Forecast data and model access are available through the following channels:

  • Earth Engine: Data is available in the Earth Engine Data Catalog.
  • BigQuery: Forecast data can be queried via BigQuery.
  • Vertex AI: An early access program is available on Google Cloud's Vertex AI platform for custom model inference.

Sources