Skip to main content

Google DeepMind's WeatherNext 3: Hourly, 5km Global Forecasts

When it comes to weather forecasting, the devil is in the details—and the details have always been a challenge for AI models. They often miss the nuances of local terrain or lag behind fast-moving storms. But Google DeepMind's latest creation, WeatherNext 3, aims to change that.

Developed in collaboration with Google Research, WeatherNext 3 is a new global weather AI model that promises hourly updates and high-resolution forecasts down to 5 kilometers. That's a big deal for anyone who's ever been caught in an unexpected downpour or wondered if the wind will be strong enough for their weekend kite-flying plans.

A Fresh Approach to Data

Traditional weather AI models rely on reanalysis grids from numerical weather prediction (NWP) systems. These grids smooth out local variations caused by coastlines, valleys, and mountains—exactly the kind of details that matter for accurate forecasts. WeatherNext 3 sidesteps this limitation by directly ingesting real-time geostationary satellite mosaics and raw meteorological station observations during training. This means its temperature and dew point forecasts at 0.05-degree resolution (about 5 kilometers) are more aligned with actual instrument readings, not just the model's internal physics.

The model builds on the same probabilistic functional generation network (FGN) architecture as its predecessor, WeatherNext 2, but expands to multi-resolution outputs. It also incorporates data from the European Centre for Medium-Range Weather Forecasts (ECMWF), NASA's IMERG satellite precipitation data, and Google's own satellite radar precipitation reanalysis.

Three Levels of Detail

One of the standout features of WeatherNext 3 is its ability to output three levels of forecast results in a single pass:

  • Level 1: 2-meter air temperature and dew point at 0.05 degrees (about 5 km), trained on station data.
  • Level 2: Grid surface data at 0.1 degrees (about 10 km), including wind speeds at 10 and 100 meters, pressure, sea surface temperature, cloud cover, solar radiation, and one-hour precipitation.
  • Level 3: Atmospheric fields across 13 pressure levels at 0.25 degrees (about 25 km).

Compared to the previous model's 0.25-degree output every six hours, this represents a fivefold improvement in spatial resolution. And the update frequency is equally impressive: the model initializes 24 times a day. The main cycles at UTC 00, 06, 12, and 18 can extend forecasts up to 15 days (360 hours) with 64 ensemble members, while hourly runs cover a 48-hour window. This rolling refresh based on current satellite observations breaks free from the old six-hour lag, making it particularly valuable for tracking rapidly developing convective storms.

Better Precipitation Predictions

Precipitation has always been a weak spot for global models—they often produce blurry, vague fields that miss storm boundaries. WeatherNext 3 tackles this by training on a rich mix of ECMWF reanalysis, NASA's IMERG satellite inversion, and Google's own satellite radar precipitation data. In early forecast stages, its Continuous Ranked Probability Score (CRPS) improves by up to 60% compared to the IMERG benchmark, and it also shows significant gains over rain gauge measurements.

Empowering Clean Energy

For the renewable energy sector, WeatherNext 3 is a game-changer. It can accurately output 100-meter wind speeds at turbine hub height, detailed cloud distributions (low, medium, and high), and all components of solar irradiance. These data points are crucial for grid operators to predict wind and solar power generation with greater precision, helping to balance supply and demand more efficiently.

Open Data, Closed Weights

Currently, WeatherNext 3's forecast data is available through Google's BigQuery, Earth Engine, and Cloud Storage. However, the model weights are not open-sourced, and custom inference still runs on the older WeatherNext 2. As this new model rolls out fully, it's clear that global weather forecasting is moving toward a future of higher precision and timeliness—one where AI doesn't just mimic the weather, but truly understands it.

Key Points

  • Hourly updates: WeatherNext 3 refreshes forecasts every hour, a major upgrade from six-hour cycles.
  • High resolution: Offers 5-kilometer detail for temperature and dew point, with multi-level outputs up to 25 kilometers.
  • Data-rich training: Uses real-time satellite mosaics and raw station observations to capture local variations.
  • Precipitation improvements: Up to 60% better CRPS scores compared to IMERG benchmarks.
  • Clean energy support: Provides wind, cloud, and solar irradiance data to aid renewable energy forecasting.
  • Access: Forecast data is open via BigQuery, Earth Engine, and Cloud Storage, but model weights remain proprietary.