"Scientific ML"
Global atmospheric prediction has long been dominated by physics-based Numerical Weather Prediction (NWP) models executing complex partial differential equations on massive supercomputing clusters. While early AI weather models (including GraphCast and WeatherNext 2) demonstrated that deep neural networks could match or exceed NWP accuracy at orders-of-magnitude lower inference compute, they inherited a severe structural flaw: they trained on NWP reanalysis data, carrying a mandatory six-hour data assimilation lag and coarse 25-kilometer spatial grids that missed localized convective storms, coastal microclimates, and mountain terrain variations.
On September 3, 2026, Google DeepMind and Google Research announced WeatherNext 3, establishing a new state of the art on the independent Operational WeatherBench (OWB) leaderboard run by Brightband. Bypassing the six-hour NWP data latency, WeatherNext 3 learns directly from a continuous mosaic of live geostationary satellite feeds, NASA IMERG precipitation radar, and sparse ground station networks. Powered by an end-to-end Functional Generative Network (FGN) mesh transformer, the model outputs high-fidelity global forecasts every hour at native 5-kilometer resolution—delivering a global picture five times sharper than its predecessor while improving precipitation accuracy by up to 60%.
Key Breakthroughs.
1. Functional
Generative Network (FGN) Mesh Transformer Architecture Previous machine learning weather systems typically decoupled coarse atmospheric modeling from localized spatial downscaling using multi-stage neural pipelines.
Unified Spherical Mesh Representation: WeatherNext 3 formulates atmospheric dynamics over an unstructured multi-scale geometric mesh mapped directly to the Earth's sphere.
- End-to-End Multitask Conditioning: A single, flexible FGN mesh transformer simultaneously processes 1-hour live geostationary satellite mosaics, historical reanalysis, and sparse station observations. Multi-Format Native Outputs: The model jointly predicts dense gridded fields across 3D atmospheric pressure levels, discrete tropical cyclone trajectory tracks, and point-level sparse coordinates for individual ground weather stations in a single inference pass.
2. Five-Times
Sharper: Native 5km Multi-Scale Spatial Resolution A forecast's practical utility depends on its ability to resolve fine spatial gradients near coastlines, urban heat islands, and mountain ranges:
Multi-Scale Field Hierarchy: WeatherNext 3 predicts key surface variables (such as 2-meter air temperature and surface moisture) at an unprecedented 5-kilometer (0.05°) grid, other surface variables at 10 kilometers, and upper-atmospheric wind vectors at 25 kilometers. Topographic Fidelity: By learning directly from ground-level observation stations rather than smoothed NWP grids, the model resolves intricate valley cold-air pools and coastal marine layers, eliminating the pixelated, over-smoothed thermal artifacts characteristic of 25km models.
3. Eliminating the 6-Hour
Lag via Real-Time Satellite Ingestion Traditional operational NWP cycles require hours to collect, quality-check, and assimilate global sensor data before running multi-hour simulations on supercomputers:
Hourly Forecast Cadence: By ingesting real-time geostationary satellite radiances continuously, WeatherNext 3 produces a fresh global forecast every hour. Rapid Storm Front Tracking: When convective storm systems, squall lines, or flash-flood atmospheric rivers materialize suddenly, the 1-hour refresh cadence provides early detection and warning margins that traditional 6-hour simulation cycles completely miss. Global South Inclusivity: Provides localized, high-resolution forecast intelligence across Latin America, Africa, and Southeast Asia, bridging regions historically underserved by regional high-resolution weather models due to local supercomputing costs. 4. 60% Improvement on Convective Precipitation Forecasting Precipitation forecasting has historically been the most stubborn bottleneck in AI atmospheric modeling due to the chaotic, small-scale cloud physics that drive rain and snow:
Dual Radar-Satellite Training: WeatherNext 3 is trained directly on NASA's Integrated Multi-satellite Retrievals for GPM (IMERG) and Google's high-resolution global radar reanalysis. CRPS Benchmark Gains: Evaluated using the Continuous Ranked Probability Score (CRPS), WeatherNext 3 achieves an accuracy improvement of up to 60% against IMERG, 30% against Multi-Radar Multi-Sensor (MRMS), and 10% against rain gauge measurements during early lead times, accurately tracing sharp convective precipitation bands rather than blurry probabilistic blobs.
5. Specialized
Variables for Clean Energy Grids To accelerate the transition to renewable power, WeatherNext 3 introduces purpose-built variables for energy operators:
100-Meter Turbine Wind Vectors: Predicts wind velocity and shear at the exact hub height of modern wind turbines, providing utility operators with actionable day-ahead and hour-ahead generation estimates. Surface Solar Irradiance & Cloud Transmissivity: Generates high-resolution cloud cover and downward shortwave solar radiation forecasts to model photovoltaic farm output with precision.
Technical Specifications & Benchmark Overview
& Benchmark Overview Metric / Dimension WeatherNext 3 WeatherNext 2 Operational NWP (ECMWF HRES) Developing Organization Google DeepMind & Google Research Google DeepMind European Centre (ECMWF) Release Date September 3, 2026 November 2025 Operational Baseline Model Architecture FGN Spherical Mesh Transformer Graph Neural Network (GNN) Hydrostatic Physics Equations Forecast Frequency Every 1 hour (Real-Time) Every 6 hours Every 6 hours Surface Spatial Resolution 5 km (0.05° grid) 25 km (0.25° grid) 9 km (0.1° grid) Data Ingestion Live Geostationary Satellites & Stations NWP Reanalysis (ERA5) Global Data Assimilation Precipitation CRPS Gain Up to +60% vs. IMERG Baseline Traditional Simulation Operational WeatherBench Ranked #1 (Brightband OWB) Prior Generation Leader Benchmark Standard Inference Runtime Minutes on Google Cloud TPUs ~1 minute on TPU v4 Hours on Supercomputers
Verified Integration & API Usage
WeatherNext 3 predictions are globally accessible via Google Earth Engine and BigQuery for enterprise climate analytics:
pythonimport ee # Initialize Google Earth Engine with your Google Cloud Project ee.Initialize(project="your-cloud-project-id") # Load WeatherNext 3 global hourly surface temperature and precipitation asset weathernext_collection = ee.ImageCollection("GOOGLE/DEEPMIND/WEATHERNEXT/V3/HOURLY") # Filter for the latest hourly forecast over a specific renewable energy site point_of_interest = ee.Geometry.Point([-122.084, 37.422]) latest_forecast = ( weathernext_collection .filterDate("2026-09-04T00:00:00", "2026-09-04T23:59:59") .filterBounds(point_of_interest) .sort("system:time_start", False) .first() ) # Extract 5km surface temperature and 100m wind vectors variables = latest_forecast.select(["temperature_2m", "wind_speed_100m", "precipitation_probability"]) sampled_data = variables.reduceRegion( reducer=ee.Reducer.first(), geometry=point_of_interest, scale=5000 ) print("WeatherNext 3 Local Forecast:", sampled_data.getInfo())