TL;DR: Google DeepMind and Google Research have released WeatherNext 2, an AI forecasting model that generates predictions 8x faster than predecessors with resolution down to 1 hour. The model produces hundreds of possible scenarios from a single input in under a minute on a single TPU, surpassing the previous WeatherNext model on 99.9% of variables and lead times.

Functional Generative Networks Enable Scenario Modeling

WeatherNext 2 employs a new AI modelling approach called Functional Generative Networks (FGN), which injects noise directly into model architecture to maintain physically realistic and interconnected forecasts. The model generates hundreds of possible weather outcomes from a single starting point, with each prediction completing in less than a minute on a single TPU—a process requiring hours on supercomputers using physics-based models.

The FGN approach proves particularly effective for predicting “marginals” (individual weather elements like temperature at specific locations) and “joints” (large, interconnected systems depending on how individual pieces interact). Despite training only on marginals, the model learns to skilfully forecast joints—essential for identifying regions affected by extreme heat or expected power output across wind farms.

Performance Improvements and Resolution

WeatherNext 2 surpasses the previous WeatherNext model on 99.9% of variables including temperature, wind and humidity, across lead times spanning 0-15 days. The model achieves higher-resolution predictions down to hourly intervals, enabling more useful and accurate forecasts for time-sensitive decisions.

This performance improvement addresses weather forecasting’s critical requirement to capture the full range of possibilities, including worst-case scenarios requiring advance planning. The model’s ability to generate hundreds of coherent scenarios from independently trained neural networks with function space noise injection provides meteorologists and decision-makers with comprehensive probability ranges.

Integration Across Google Products and Cloud Services

Google has integrated WeatherNext 2 technology across multiple products, upgrading weather forecasts in Search, Gemini, Pixel Weather and Google Maps Platform’s Weather API. The model will power weather information in Google Maps in coming weeks.

For developers and researchers, WeatherNext 2 forecast data is now available in Earth Engine and BigQuery. Google Cloud has launched an early access programme on Vertex AI platform for custom model inference, expanding accessibility beyond Google’s product ecosystem.

The company has also supported weather agencies through experimental cyclone predictions using the technology’s scenario-generation capabilities, enabling decisions based on ranges of possible outcomes rather than single-point forecasts.

Future Research Directions

Google DeepMind indicates active research into capabilities including integrating new data sources and expanding access further. The team aims to provide powerful tools and open data to accelerate scientific discovery and empower researchers, developers and businesses to address complex problems with geospatial and weather forecasting applications.

The release positions AI-powered weather forecasting as a practical alternative to traditional physics-based supercomputer models, offering speed and efficiency advantages whilst maintaining prediction accuracy across variables and timeframes. By making forecast data accessible through multiple channels—from consumer products to enterprise cloud platforms—Google extends advanced weather prediction capabilities to diverse use cases spanning supply chain management, aviation planning and daily commute decisions.

Source: Google Blog