Decoding Official Information and Mechanisms of Google DeepMind’s Weather Prediction AI “WeatherNext 3”

This article is a technical explanation and implementation example created using AI. The published code and procedures are based on primary sources, but the author has not verified their operation on actual devices. Operation may vary depending on the environment and version.

Announced by Google DeepMind and Google Research, “WeatherNext 3” is a state-of-the-art global weather AI model that performs high-precision and high-resolution forecasting using real-time observational data, without relying on traditional large-scale physical simulations. Based on published primary sources, this article organizes the model’s architecture, data processing characteristics, and expansion into the ecosystem.

Overview and Background of WeatherNext 3

Weather information influences numerous decision-making processes, ranging from daily umbrella preparations to agriculture, supply chains, clean energy production, and national economies. In recent years, AI has achieved faster forecasting than traditional methods by utilizing past weather records, but predicting localized and rapidly changing weather has presented challenges.

Conventional models suffered from issues such as insufficient spatial resolution and difficulty in incorporating real-time data like satellites. WeatherNext 3 adopts an approach that learns directly from real-time observational data to address these challenges, providing timely and more localized forecasts. High accuracy has also been demonstrated in independent live evaluations.

Architecture and System Configuration

Primary sources describe the system architecture and data flow of WeatherNext 3. The following Mermaid diagram shows the components and information flow described in the primary sources.

flowchart TD
    A[Live 1-hour Geostationary Satellite Mosaics] --> C[Functional Generative Network FGN Mesh Transformer]
    B[Traditional Historical Analysis] --> C
    C --> D[Dense Gridded Fields]
    C --> E[Discrete Cyclone Tracks]
    C --> F[Station-level Sparse Coordinates]

In this configuration, hourly real-time geostationary satellite mosaics and traditional historical analysis data are combined and fed into a single, flexible Functional Generative Network (FGN) mesh transformer. This natively predicts dense gridded fields, discrete cyclone tracks, and station-level sparse coordinates.

Improved Resolution and Update Frequency

The usefulness of a forecast largely depends on its temporal and spatial granularity (resolution). WeatherNext 3 generates hourly forecasts at multiple spatial resolutions, maintaining consistency from broad, macro-scale atmospheric wind patterns down to regional terrain.

• Surface variables (temperature, humidity, etc.): Visualized at 5-kilometer resolution.

• Other surface variables: 10-kilometer resolution.

• Atmospheric variables (wind speed, etc.): 25-kilometer resolution.

While its predecessor, WeatherNext 2, made forecasts every 6 hours on a 25-kilometer grid, WeatherNext 3 delivers global weather imagery that is overall about 5 times sharper.

Utilization of Real-Time Satellite Data and Sparse Observations

Many AI weather models are trained on data from Numerical Weather Prediction (NWP) models. NWP models are complex physical simulations driven by supercomputers, but they introduce a data lag of about 6 hours, which causes biases in rapidly changing variables like rain and surface temperature.

WeatherNext 3 addresses this challenge through the following mechanisms:

1. Ingestion of real-time satellite mosaics: Acquires continuously updated views of the atmosphere, enabling the generation of hourly forecasts based on the latest satellite observations.

2. Use of sparse weather observation data: Because temperature and humidity fluctuate dramatically over ranges of just a few kilometers in areas such as coastlines, valleys, and mountain ranges, WeatherNext 3 trains directly using sparse data from weather stations. This enables global forecasting on a 5-kilometer grid that accounts for topographical details, bringing benefits to regions such as Central and South America, Africa, and the Asia-Pacific that traditional high-resolution models struggled to maintain.

Variables for Clean Energy and Improved Precipitation Forecasting Accuracy

Support for Clean Energy Production

In addition to improved resolution and update frequency, forecasting capabilities dedicated to clean energy production have been introduced. In addition to predicting wind speeds at 100 meters—equivalent to turbine height—to accurately estimate wind power generation output, it predicts high-resolution cloud cover and solar radiation to help estimate the amount of light solar power plants receive on the ground.

Improved Accuracy in Precipitation Forecasting

Rain and snow systems originate from small-scale, rapidly moving cloud processes, an area where traditional physical simulations were prone to blurred estimations or missing severe storm boundaries. WeatherNext 3 is trained using two high-quality sources: NASA’s IMERG (Integrated Multi-satellite Retrievals for GPM) and global precipitation reanalysis data based on satellite radar. This demonstrates significant improvements in initial lead-time rainfall measurements in evaluations such as the Continuous Ranked Probability Score (CRPS).

Integration into the Google Ecosystem and Usage

WeatherNext 3 goes beyond laboratory achievements and is integrated into Google’s core ecosystem and related services.

• High-resolution forecast data: Provides global weather forecasts that can be integrated into workflows without model configuration. Researchers, developers, and enterprises can query data in BigQuery and Earth Engine, and download it in bulk from Google Cloud Storage.

• Global deployment: Enhances weather experiences across Google Search, the Gemini app, Google Maps, the Google Maps Platform Weather API, and Google Earth Engine. Long-range forecast accuracy is improved by up to 50%, bringing major enhancements to regions where forecast reliability was previously low.

Usage Precautions

As noted in the primary sources, for official weather forecasts, weather warnings, and public safety advisories, it is recommended to refer to information from local meteorological agencies and national weather services.

Reference Information

• source_title: Introducing WeatherNext 3, our most advanced and accurate global weather AI model

• source_url: https://deepmind.google/blog/introducing-weathernext-3-our-most-advanced-and-accurate-global-weather-ai-model/