Google's WeatherNext 3 ditches physics simulations and learns weather directly from live satellite data
Google and DeepMind released WeatherNext 3, an AI weather model that abandons traditional physics-based numerical weather prediction (NWP) simulations in favor of learning directly from real-time geostationary satellite data The model delivers hourly forecasts at up to five-kilometer resolution—five times sharper than WeatherNext 2's 25-kilometer grid—enabling it to capture local terrain features like mountain ranges, coastlines, and valleys Precipitation forecasts show up to 50% improvement for
Analysis
TL;DR
- Google and DeepMind released WeatherNext 3, an AI weather model that abandons traditional physics-based numerical weather prediction (NWP) simulations in favor of learning directly from real-time geostationary satellite data
- The model delivers hourly forecasts at up to five-kilometer resolution—five times sharper than WeatherNext 2's 25-kilometer grid—enabling it to capture local terrain features like mountain ranges, coastlines, and valleys
- Precipitation forecasts show up to 50% improvement for day-ahead planning, with CRPS gains of up to 60% over NASA's IMERG, 30% over MRMS, and 10% over rain gauges at short lead times
- WeatherNext 3 generates renewable energy-specific outputs including 100-meter wind speed predictions for wind farms and cloud cover/solar irradiance estimates for solar installations
- The model is now live across Google Search, Maps, Gemini, the Weather API, and Google Earth Engine, with data accessible via BigQuery and Google Cloud Storage
Why It Matters
WeatherNext 3 represents a significant paradigm shift in operational meteorology by replacing delayed supercomputer-based NWP simulations with live satellite ingestion, directly addressing the six-hour latency that has historically degraded accuracy for fast-changing variables like rainfall and temperature. For AI practitioners and climate tech professionals, it demonstrates how end-to-end learning from raw observational data can outperform traditional physics-informed approaches at scale, while also opening new commercial applications in renewable energy grid management.
Technical Details
- Architecture: Built on the Functional Generative Network (FGN), the same foundation as WeatherNext 2 and GenCast, representing DeepMind's ongoing investment in generative approaches to weather prediction
- Data pipeline: Ingests live geostationary satellite data rather than pre-computed NWP outputs, combining this with individual weather station observations and NASA's IMERG satellite-based precipitation dataset alongside Google's own global precipitation analysis from satellite radar
- Multi-resolution output: Temperature and humidity at 5 km, other surface variables at 10 km, and atmospheric values like wind speed at 25 km—contrasting with WeatherNext 2's uniform 25 km grid at six-hour intervals
- Benchmarking: Evaluated using Continuous Ranked Probability Score (CRPS), showing 60% improvement over IMERG, 30% over MRMS, and 10% over rain gauges at short lead times for medium-range global forecasts
- Infrastructure: Runs on Google TPUs, with hourly forecast generation; data distributed via BigQuery, Google Cloud Storage, and Earth Engine APIs
Industry Insight
- The shift from physics-simulation-dependent training to live satellite data ingestion sets a new benchmark for AI-native weather modeling, suggesting that future operational models will prioritize direct observational learning over numerical approximation—this has implications for how meteorological institutions invest in data infrastructure versus supercomputing
- Renewable energy integration is a strategic differentiator: by predicting wind speeds at turbine height and solar irradiance with high spatial resolution, WeatherNext 3 positions Google to capture value in the growing clean energy grid management market, potentially influencing how utilities and energy traders source forecast data
- Regions historically underserved by expensive traditional regional models—particularly Latin America, Africa, and the Asia-Pacific—stand to benefit most from this open, high-resolution global approach, which could accelerate AI-driven meteorological equity and reduce dependency on legacy forecasting systems in developing economies
Disclaimer: The above content is generated by AI and is for reference only.