Skip to main content

Google's WeatherNext 3: Sharper Forecasts, Hourly Updates

For years, AI weather models have been nipping at the heels of traditional physics-based forecasting. Yet two stubborn problems remained: their views were too coarse to capture local quirks like mountains and coastlines, and they updated too slowly—often lagging about six hours behind the latest observations. Now, Google DeepMind and Google Research have unveiled WeatherNext 3, a model designed to tackle both issues head-on.

A Fresh Perspective: Learning from the Ground Up

Most AI forecasters learn from reanalysis data—a blend of past observations and model output that can smooth over the very local variations that matter for accurate predictions. WeatherNext 3 breaks that mold. Instead of relying solely on reanalysis grids, it trains directly on raw meteorological station measurements, using a dedicated observation head to calibrate its temperature and dew point outputs to actual instrument readings.

What truly sets it apart, though, is its input. The model ingests real-time mosaics from global geostationary satellites, refreshing its forecasts every single hour. This constant stream of current data allows it to capture rapidly developing weather systems that older models might miss. According to Google, independent evaluations by Brightband have already rated it the most accurate global weather model to date.

Sharper, Faster, More Detailed

A single forward pass of WeatherNext 3 produces forecasts at three resolutions. The finest, at 0.05 degrees (about 5 kilometers), covers 2-meter temperature and dew point. A mid-tier 0.1-degree grid (roughly 10 kilometers) handles surface winds, pressure, sea surface temperature, cloud cover, solar radiation, and hourly precipitation. The coarsest, 0.25 degrees (about 25 kilometers), provides atmospheric fields across 13 pressure levels.

Compare that to its predecessor, WeatherNext 2, which only managed a 0.25-degree output every six hours. The leap in clarity is staggering—Google calls it a five-fold improvement.

But resolution isn't the only upgrade. WeatherNext 3 initializes 24 times a day. Four of those runs—at 00, 06, 12, and 18 UTC—extend forecasts out to 15 days with 64 ensemble members, while the remaining hourly runs cover the next 48 hours. For fast-moving storms, that hourly refresh based on live satellite imagery can make all the difference.

Precipitation: The Achilles' Heel, Addressed

Global models have long struggled with precipitation, often producing blurry, indistinct rain fields that miss storm boundaries. WeatherNext 3 tackles this by training on three data sources: ECMWF reanalysis, NASA's IMERG satellite estimates, and Google's own satellite-radar precipitation reanalysis. The results are impressive. In short-term forecasts, the model's CRPS—a measure of forecast accuracy—improves by up to 60% compared to IMERG, 30% against MRMS, and 10% versus rain gauges. When evaluated against IMERG, the Brier score and CRPS drop by as much as 50% relative to traditional NWP baselines.

Powering the Renewable Revolution

For the energy sector, WeatherNext 3 is a game-changer. It outputs 100-meter wind speeds at typical turbine hub heights, complete cloud cover data across low, mid, and high levels, and two components of solar irradiance—SSRD and FDIR. That's precisely the combination grid operators need to predict wind and solar power generation, helping them balance supply and demand more efficiently.

Access and Limitations

Forecast data from WeatherNext 3 is available through Google Cloud's BigQuery, Earth Engine, and Cloud Storage. However, the model's weights are not open-sourced, and custom inference still runs on WeatherNext 2. That means while the technology is impressive, its full potential remains locked behind Google's proprietary walls.

Key Points:

  • WeatherNext 3 delivers forecasts at up to 5-kilometer resolution, refreshed hourly.
  • It trains on real-time satellite data and direct station observations, not just reanalysis.
  • Precipitation forecasts show dramatic improvements, with CRPS gains up to 60%.
  • The model supports renewable energy planning with detailed wind, cloud, and solar data.
  • Data is accessible via Google Cloud, but model weights are not open-sourced.