WeatherNext 2 - Google DeepMind's AI weather forecasting model

What is WeatherNext 2?

WeatherNext 2 is a next-generation AI weather forecasting model jointly developed by DeepMind and Google Research. The model employs a Functional Generative Network (FGN) architecture, offering approximately eight times the computational speed of its predecessor and generating detailed hourly forecasts. It can extrapolate hundreds of possible weather scenarios from a single initial state, with each extrapolation taking less than a minute on a single TPU. In forecast periods of 0 to 15 days, WeatherNext 2 outperforms its predecessor in 99.9% of meteorological variables (such as temperature, wind speed, and humidity).

Main features of WeatherNext 2

• High-precision weather forecastIt can generate detailed hourly forecasts, covering a forecast period from 0 to 15 days, and comprehensively surpasses previous models in various meteorological variables such as temperature, wind speed, and humidity.
• fast calculationIt adopts an innovative "function generation network" architecture, which improves the computing speed by about 8 times compared to the previous generation, and a single prediction takes less than one minute on a single TPU.
• Multi-scenario simulationIt can deduce hundreds of possible weather development scenarios from a single initial state, providing a more comprehensive reference for meteorological research and disaster early warning.
• Low probability event predictionAccurately predict low-probability catastrophic weather events, such as typhoons and heavy rainfall, to improve early warning capabilities for extreme weather.
• Data platform integrationThe prediction data has been uploaded to the Earth Engine and BigQuery platforms for researchers and developers to access and use.
• Cloud service supportGoogle Cloud’s Vertex AI platform has launched an early access program for custom model inference, allowing users to make customized predictions based on their own needs.
• Multi-scenario applicationsIt has been integrated into the Weather API of Google Search, Gemini, Pixel Weather, and Google Maps Platform, and will be further applied to weather information in Google Maps in the coming weeks to provide users with a more convenient weather query service.

The technical principles of WeatherNext 2

• Noise injection into the function spaceFGN generates globally consistent variability by injecting low-dimensional noise vectors (such as 32-dimensional ones) into the function space and applying the noise to the entire network through a conditional normalization layer. This mechanism enables the model to generate hundreds of possible weather scenarios from a single input while maintaining physical plausibility.
• Graph Neural Network ArchitectureThe FGN employs a graph neural network (GNN) encoder/decoder structure, mapping latitude and longitude grids to a latent space on a spherical icosahedral grid, and using a Graph-Transformer processor to operate on this grid. This architecture helps capture the spatial dependencies between complex meteorological variables.
• Multi-stage training strategyThe training of FGN is divided into multiple stages, including pre-training using ERA5 data and fine-tuning using HRES-fc0 data. The final stage includes autoregressive training, where the model performs multi-step rolling predictions and calculates the loss to optimize prediction performance.
• Marginal and Joint Distribution PredictionAlthough FGN is trained only on marginal distributions (such as single variables like temperature and wind speed), it can learn to capture joint distributions (i.e., complex interactions between multiple variables). This enables the model to generate more comprehensive and accurate weather forecasts.
• High resolution and fast predictionThe FGN model can generate detailed hourly forecasts, and a single prediction takes less than a minute on a single TPU. This high-speed and high-resolution forecasting capability significantly improves the efficiency of weather forecasting.

WeatherNext 2 project address

• Project official websitehttps://deepmind.google/science/weathernext/

Application scenarios of WeatherNext 2

• Consumer-side weather servicesWeatherNext 2 is integrated into services such as Google Search, Gemini, and Pixel Weather, providing users with more accurate and high-resolution weather forecasts.
• Developers and Enterprise ApplicationsIts forecast data is available on Earth Engine and BigQuery platforms, and is offered through an early access program on Google Cloud's Vertex AI platform for developers and enterprises to perform customized model inference. This enables enterprises and developers to develop customized applications based on high-precision weather data, such as optimizing logistics routes and managing supply chains.
• Meteorological agencies and disaster early warningBy providing hundreds of possible weather scenarios, WeatherNext 2 helps meteorological agencies conduct multi-scenario analysis, identify extreme weather events in advance, and improve emergency response efficiency.
• Energy and AgricultureThe model can be used to optimize the use of renewable energy, such as wind and solar power, and improve grid stability. Meanwhile, its medium-term weather forecasting capabilities are also applicable to agricultural planning, helping farmers develop planting and harvesting plans.
• Scientific research and data sharingWeatherNext 2’s openness enables researchers to access high-precision weather data, further driving innovation in meteorological research and related fields.