Google WeatherNext 3 Deep Dive: The AI Weather Forecasting Revolution Driven by Satellite Data

Weather forecasting, a seemingly everyday topic, actually influences billions of decisions daily — from the simple matter of whether to bring an umbrella, to major planning in agriculture, supply chains, clean energy, and national economies. Recently, Google DeepMind and Google Research jointly released WeatherNext 3, which according to real-time evaluations by independent assessment organization Brightband, is the most advanced and accurate global AI weather forecasting model to date.

Unlike traditional Numerical Weather Prediction (NWP) models that rely on physics simulations, WeatherNext 3 learns directly from real-time satellite observation data, generating localized forecasts at higher resolution and faster frequency. This shift not only dramatically improves forecast accuracy but also brings high-quality weather forecasting to more corners of the globe.

Unprecedented Spatiotemporal Resolution: Five Times Sharper Than the Previous Generation

The practical value of a weather forecast often depends on its ability to capture fine details across time and space. WeatherNext 3 generates hourly forecasts and supports multiple spatial resolutions, maintaining physical consistency from large-scale global wind patterns down to local terrain details.

Specifically, the model presents different meteorological variables at the following resolution levels:

• 5 km resolution: Key surface variables such as temperature and humidity
• 10 km resolution: Other surface variables
• 25 km resolution: Atmospheric variables such as wind speed

Overall, this provides approximately five times sharper global weather imagery compared to the previous generation WeatherNext 2 — which could only output forecasts on a 25 km grid at 6-hour increments.

Take the 2-meter temperature forecast over the UK as an example: WeatherNext 2 at 25 km resolution showed pixelated, overly smoothed thermal distributions, while WeatherNext 3 at its native 5 km resolution can clearly resolve complex local terrain, avoiding the blurriness of the older model. This level of detail is especially critical for communities near coastlines, valleys, or mountain ranges, where temperature and humidity can change dramatically within just a few kilometers. To understand the significance of this improvement, think of it like map navigation — a 25 km resolution forecast can only tell you "London will have rain today," while 5 km resolution can distinguish that the southeast of the city is raining while the northwest is sunny, making a fundamental difference for travel decisions and local disaster preparedness.

Real-Time Satellite Data: WeatherNext 3's Biggest Technical Leap

The most core breakthrough of WeatherNext 3 lies in a fundamental change in training data sources. Most AI weather models, including WeatherNext 2, are trained on output data from Numerical Weather Prediction (NWP) models. While NWP is useful, it is essentially a complex physics simulation driven by supercomputers, with approximately 6 hours of data latency that introduces noticeable bias for rapidly changing variables like precipitation and surface temperature.

Numerical Weather Prediction (NWP) is the cornerstone of modern meteorology. Its core idea is to treat the atmosphere as a physical system governed by fluid dynamics equations, simulating atmospheric evolution by solving systems of partial differential equations including the Navier-Stokes equations, thermodynamic equations, and moisture continuity equations. This approach was first conceptualized by Norwegian meteorologist Vilhelm Bjerknes in 1904, but didn't become practical until the advent of computers in the 1950s. Currently, the world's mainstream NWP systems include the European Centre for Medium-Range Weather Forecasts (ECMWF) IFS model, the U.S. National Weather Service GFS model, and others, all of which require hours of computation across thousands of CPU cores to complete a single global forecast. A fundamental challenge of NWP lies in the "initial condition problem" — simulations must start from an atmospheric initial state that is as accurate as possible, yet constructing this initial state itself requires a complex data assimilation process that inherently introduces delays and errors. It is precisely this bottleneck that motivated WeatherNext 3 to bypass NWP outputs and learn directly from raw observational data.

A Paradigm Shift: From Physics Simulation to Real-World Observations

By ingesting real-time composite data from global geostationary satellites, WeatherNext 3 gains a continuously updated, rich perspective of the atmospheric state. This enables the model to generate a new forecast every hour, each based on the latest satellite observations at up to 5 km resolution.

Geostationary satellites are critical infrastructure for meteorological observation. These satellites orbit approximately 35,786 km above the equator, with an orbital period that exactly matches Earth's rotation, allowing them to continuously monitor weather changes over the same area. The world's major geostationary weather satellites currently include the U.S. GOES series, Europe's Meteosat series, China's Fengyun-4, Japan's Himawari series, and South Korea's GEO-KOMPSAT satellites, which together form a global observation network. These satellites carry advanced imagers that acquire imagery at minute-level frequency across multiple spectral channels including visible light, infrared, and water vapor, providing rich information on cloud-top temperature, atmospheric humidity profiles, sea surface temperature, and more. WeatherNext 3 composites the real-time data streams from these satellites to construct a continuously updated global atmospheric state map, fundamentally bypassing the time bottleneck of the data assimilation step in traditional NWP systems.

This capability is crucial because critical weather events often develop rapidly. When storms, fronts, or precipitation systems form suddenly, rapid update cycles and higher spatial resolution can provide earlier, more detailed early warning insights, supporting more effective emergency response.

Additionally, the model is trained directly on sparse weather station observation data, allowing the 5 km global grid forecasts to fully account for regional details such as terrain. This breakthrough is particularly significant for Latin America, Africa, and the Asia-Pacific region — areas that have long lacked high-resolution forecast services due to the prohibitive supercomputing costs required for traditional regional models.

Running a regional high-resolution NWP system is extremely expensive. Take the U.S. National Weather Service as an example: operating NOAA's weather forecasting supercomputer systems requires hundreds of millions of dollars in annual budget, and ECMWF's supercomputing center in Europe is similarly costly. Establishing a 3-5 km resolution regional model covering a nation's territory typically requires thousands of computing cores running continuously, plus teams of professional meteorologists for model tuning and post-processing. World Meteorological Organization (WMO) reports indicate that weather station density across the African continent is only one-eighth of the WMO recommended standard, and many Pacific island nations and Central Asian regions have virtually no localized high-resolution forecasting capability. This "forecast gap" means that the world's most vulnerable communities — those most susceptible to extreme weather impacts — receive the least and coarsest warning information. By replacing local supercomputer deployments with cloud-based AI inference, WeatherNext 3 reduces the marginal cost of high-resolution forecasting to near zero, giving billions of people and local businesses access to localized, high-fidelity weather forecasts for the first time. This has profound implications for closing the global meteorological service gap.

Specialized Forecasting Capabilities for Clean Energy

Notably, WeatherNext 3 also introduces forecasting features specifically designed for renewable energy production:

• 100-meter altitude wind speed forecasts: Corresponding to wind turbine hub height, for precise wind power output estimation
• High-resolution cloud cover forecasts: Helping solar power plants assess shading impacts
• Solar radiation level forecasts: Estimating actual ground-level light intensity received

Modern large wind turbines have their hubs (the point where the blades' central rotation axis is located) typically positioned 80 to 150 meters above ground level, with mainstream commercial models around 100 meters. Wind speed at this height differs significantly from ground-level measurements at 10 meters (the traditional weather station measurement height) — due to atmospheric boundary layer wind shear, wind speeds at 100 meters are typically 40%-80% higher than at the surface. More critically, wind power output is proportional to the cube of wind speed, meaning that every 1% improvement in wind speed forecast accuracy can reduce power generation estimation error by up to 3%. For wind farms with installed capacities of hundreds of megawatts, this directly affects daily dispatch decisions worth millions of dollars. WeatherNext 3's native 100-meter altitude wind speed forecasts enable wind power operators to directly obtain the meteorological parameters most relevant to power generation, without relying on imprecise height extrapolation empirical formulas.

This data is vital for global clean energy planning, enabling grid operators and renewable energy developers to accurately predict power generation and intelligently match it with consumer demand.

Precipitation Forecast Accuracy Breakthrough: CRPS Improvement of Up to 60%

Accurately predicting precipitation has always been a well-recognized challenge for global weather models. Rain and snow systems are driven by micro-scale, fast-moving cloud physics processes that are difficult for traditional physics simulations to model accurately, leading AI forecasts to often produce blurry estimates or completely miss the boundaries of strong storms.

To tackle this challenge, the Google team used two high-quality precipitation data sources for training:

• NASA IMERG dataset: Satellite-based global precipitation estimates
• Google's proprietary precipitation reanalysis data: A satellite radar-based global precipitation product

NASA IMERG (Integrated Multi-satellitE Retrievals for GPM) is the core data product of the Global Precipitation Measurement (GPM) mission. The GPM mission is jointly led by NASA and the Japan Aerospace Exploration Agency (JAXA), carried out by a constellation consisting of one core observation satellite and over a dozen international partner satellites. IMERG fuses microwave and infrared remote sensing data to provide precipitation estimates covering the globe between 60°N and 60°S latitude, with 30-minute temporal resolution and 0.1-degree (approximately 11 km) spatial resolution. It fills the gaps where ground rain gauges and radar lack observation capability — over oceans, deserts, mountains, and other uninhabited areas — and is one of the most authoritative global precipitation reference datasets, widely used in climate research, hydrological modeling, and disaster monitoring.

The result is a significant leap in precipitation forecast accuracy. In medium-range global forecast evaluations, the Continuous Ranked Probability Score (CRPS) achieved impressive results compared to baselines:

• Up to 60% improvement relative to the IMERG dataset
• 30% improvement relative to MRMS
• 10% improvement relative to rain gauge measurements at early lead times

The Continuous Ranked Probability Score (CRPS) is a widely used probabilistic evaluation metric in weather forecast verification. Unlike simply comparing forecast values to observations using mean squared error, CRPS comprehensively evaluates the full distribution characteristics of probabilistic forecasts — it measures the "distance" between the forecast's cumulative distribution function and the actual observation. Lower CRPS values indicate that the forecast probability distribution more closely matches reality. The advantage of this metric is that it simultaneously penalizes both forecast bias (insufficient accuracy) and excessive uncertainty (overly wide forecast ranges), making it considered a more comprehensive evaluation standard than single deterministic scores. A 60% CRPS improvement means WeatherNext 3's precipitation probability forecasts have achieved a qualitative leap in agreement with observations — an extraordinarily rare magnitude of single-generation improvement in the meteorological community.

In visual comparisons of precipitation probability forecasts, WeatherNext 3 at 11 km resolution precisely captures the sharp boundaries of convective bands, nearly matching real satellite observations, while WeatherNext 2 could only produce highly diffuse and pixelated precipitation footprints.

Deployment Across the Google Ecosystem

Google's core objective is to make weather intelligence universally accessible — whether for emergency responders tracking sudden wind shifts, air traffic controllers planning flight paths, or farmers managing crops. To bring these technical breakthroughs from the lab into real-world applications, Google is integrating WeatherNext 3 across its complete product ecosystem:

Data Access for Developers and Researchers

Global weather forecast data is updated hourly, accessible for integration into workflows without building your own model:

• Query and analyze data in BigQuery and Earth Engine
• Bulk download raw forecast data from Google Cloud Storage

Product Integration for Everyday Users

Starting today, WeatherNext 3 powers forecasts in the following products:

• Google Search
• Gemini app
• Google Maps
• Google Maps Platform Weather API
• Google Earth Engine

When users plan trips of a day or longer, they will see precipitation forecast accuracy improvements of up to 50%, with the most noticeable improvements in regions where forecast reliability has historically been lower.

Conclusion: AI Weather Forecasting Moves from "Second-Hand Data" to "First-Hand Observations"

The atmosphere will always retain a certain degree of unpredictability. But by training on real observational data and bypassing traditional modeling constraints, WeatherNext 3 brings us closer to a future where forecasts truly match what's actually happening on the ground.

From a technical roadmap perspective, this upgrade represents a paradigm shift in AI weather forecasting — from relying on "second-hand data" from physics simulations to directly learning from "first-hand observations" from satellites and weather stations. This not only improves accuracy and speed, but more importantly dramatically lowers the barrier to high-quality forecasting, enabling developing regions long excluded by supercomputing costs to benefit as well.

However, Google also reminds users that for official forecasts, severe weather warnings, and public safety advisories, information published by local meteorological agencies should remain the authoritative source.