Skip to main content
AI Learning Resources

Quick Overview

This announcement video from Google DeepMind introduces WeatherNext 3, presented by Product Lead Nofar Peled Levi and Research Scientist Stephan Rasp. The video explains the technological advancements in their latest AI weather forecasting model following previous research milestones in atmospheric machine learning.

Key Points

  • 1.WeatherNext 3 is an operational global AI weather forecasting model that updates hourly and delivers forecasts down to five-kilometer spatial resolution.
  • 2.Unlike traditional physics-based simulations or earlier AI systems that relied on six-hour reanalysis cycles, the model directly processes raw satellite and weather station data.
  • 3.The model outputs three native spatial resolutions in a single pass, covering 25-kilometer atmospheric fields, 9-kilometer surface variables, and 5-kilometer temperature and humidity data.
  • 4.Specialized outputs are included for renewable energy operators, specifically measuring wind speed and direction at 100-meter turbine heights alongside solar radiation and cloud cover.
  • 5.Google DeepMind is integrating WeatherNext 3 across consumer platforms, including Google Search, Gemini, and Google Maps.

Summary

Predicting the weather is one of science's oldest and most computationally challenging tasks. Traditional numerical forecasting uses mathematical physics equations to calculate atmospheric changes step by step. Running these physics simulations globally at high fidelity is extremely expensive and time-consuming, historically forcing forecasters to compromise between localized high-resolution models and global low-resolution forecasts. Machine learning offers a solution by training on historical observations to predict atmospheric patterns much faster and with greater accuracy than traditional physics-based simulations.

While earlier machine learning weather models demonstrated significant speed and accuracy improvements, they were generally restricted to relying on reanalysis data that refreshed only every six hours. Atmospheric conditions such as storm clouds, sudden precipitation, and shifting winds evolve on the order of minutes, making a six-hour delay suboptimal. WeatherNext 3 solves this limitation by directly ingesting raw real-world observation streams, including inputs from orbiting satellites and ground-based weather stations. Processing these live observations allows the system to generate a brand-new global forecast every hour of the day.

Spatial resolution is critical for translating global weather patterns into meaningful local forecasts. Averaging weather variables over large grid squares, such as 25 kilometers by 25 kilometers, obscures local microclimates that people experience on the ground, especially in coastal or mountainous terrain. WeatherNext 3 addresses this by outputting three distinct spatial resolutions within a single pass. It provides a 25-kilometer resolution for broad atmospheric shifts, a 9-kilometer resolution for surface variables like wind and pressure, and a high-fidelity 5-kilometer resolution for local temperature and humidity predictions.

Beyond general weather tracking, the system is designed to support industries critically reliant on atmospheric precision, particularly renewable energy. WeatherNext 3 models wind speed and direction at a height of 100 meters, matching the typical hub height of modern wind turbines, while also calculating cloud cover and solar radiation to estimate solar panel output. DeepMind is deploying the model across Google products, including Google Search, Gemini, and Google Maps, to provide timely forecasts for farming, flood disaster response, and daily consumer planning.

The Challenge of Traditional Weather Forecasting

Traditional weather forecasting relies on complex physics equations that simulate atmospheric variables step by step across the globe. This approach requires vast computing resources and lengthy runtimes, forcing meteorologists to choose between regional high-resolution forecasts and low-resolution global models. In contrast, AI models learn directly from historical atmospheric observations, allowing them to deliver faster and more accurate global predictions without the computational bottleneck of standard physics simulations.

Hourly Updates from Direct Real-World Observations

Earlier AI weather models depended heavily on atmospheric reanalysis data, which limited forecast refresh intervals to roughly every six hours. Because atmospheric phenomena like precipitation and wind shifts can change rapidly within minutes, WeatherNext 3 directly assimilates raw real-world data from weather satellites and ground stations. This architectural change allows the system to produce an updated global forecast every single hour.

Multi-Scale Resolution and Renewable Energy Applications

To deliver localized detail where people live and operate, WeatherNext 3 provides three native spatial resolutions in a single pass: 25 kilometers for broad atmospheric patterns, 9 kilometers for surface wind and pressure, and 5 kilometers for temperature and humidity. The system also includes targeted variables for the renewable energy sector, such as solar radiation metrics for solar arrays and wind speed and direction calculated at a height of 100 meters for wind turbine management.

The Bottom Line

WeatherNext 3 establishes that operational AI weather forecasting can overcome the historical trade-off between global scope, high spatial resolution, and hourly update speed by ingesting raw observational data directly. It demonstrates practical advantages across sectors ranging from renewable energy management to flood evacuation planning. The presentation outlines the model's architecture and product rollout across Google surfaces while leaving the technical benchmarks against specific national meteorological operational systems unstated.

FAQ

What is WeatherNext 3 and what capabilities does it introduce for global weather forecasting?

WeatherNext 3 is an AI-based operational global weather forecasting model developed by Google DeepMind that produces hourly forecast updates at up to five-kilometer spatial resolution across major atmospheric variables.

How does the AI approach in WeatherNext 3 differ from traditional numerical weather prediction models?

Traditional models step through mathematical physics equations describing atmospheric variables, which is computationally expensive and slow. WeatherNext 3 instead trains on historical atmospheric observations to infer future conditions quickly and accurately.

Why can WeatherNext 3 update forecasts every hour while earlier AI models updated every six hours?

Earlier AI models relied on reanalysis data that only refreshed periodically, whereas WeatherNext 3 directly ingests raw observation data from satellites and ground stations, enabling continuous hourly forecast generation.

What are the three native spatial resolutions produced in a single pass by WeatherNext 3?

WeatherNext 3 provides 25-kilometer resolution for broad atmospheric changes, 9-kilometer resolution for surface wind and pressure, and 5-kilometer resolution for temperature and humidity.

What specific atmospheric variables does WeatherNext 3 predict for renewable energy management?

The model forecasts wind speed and wind direction at a height of 100 meters for wind turbines, as well as cloud cover and solar radiation to predict solar panel energy generation.

On which consumer platforms and Google surfaces will WeatherNext 3 weather forecasts be available?

WeatherNext 3 forecasts will be integrated across Google Search, Gemini, and Google Maps to reach users worldwide.

Worth watching for

Meteorologists, atmospheric researchers, renewable energy operators, and technologists interested in machine learning applications for operational weather forecasting.

  • deepmind
  • weather-forecasting
  • machine-learning
  • atmospheric-science
  • renewable-energy
Watch on YouTubeMore videos from Google DeepMind