Skip to main content

Google's WeatherNext 3: Sharper Forecasts, Hourly Updates

For three years, AI weather models have been nipping at the heels of traditional physics-based forecasting. But two stubborn problems remained: their forecasts were too blurry to capture local terrain, and they lagged behind by about six hours compared to conventional analysis. Now, Google DeepMind and Google Research have unveiled WeatherNext 3, a model that takes direct aim at both issues.

A Fresh Look at the Sky

Instead of relying solely on reanalysis data—which can smooth over the nuances of coastlines, valleys, and mountains—WeatherNext 3 trains directly on real-time satellite mosaics and ground station measurements. This means it can see the world as it is, not just as a smoothed grid. The model reinitializes every hour, producing forecasts at a resolution of 0.05 degrees, roughly 5 kilometers. According to Google, independent evaluations by Brightband have already rated it as the most accurate global weather model to date.

How It Works

Architecturally, WeatherNext 3 is a grid Transformer that belongs to the same probabilistic family as its predecessor, WeatherNext 2. But it extends to multi-resolution outputs. Its inputs include live geostationary satellite imagery and ECMWF HRES analysis, while training data draws from ERA5, NASA's IMERG, site observations, and satellite mosaics. A dedicated observation head calibrates temperature and dew point outputs against actual instrument readings, a departure from the common practice of learning only from reanalysis grids.

Sharper, Faster, Better

A single forward pass yields three levels of detail: 2-meter temperature and dew point at 5 km resolution, surface winds, pressure, sea surface temperature, cloud cover, solar radiation, and hourly precipitation at 10 km, and a broader 25 km atmospheric field spanning 13 pressure levels. Compare that to WeatherNext 2, which only managed a 25 km field every six hours—hence the claim of "five-fold clarity improvement."

Refresh frequency is another game-changer. The model runs 24 times a day, with four main cycles (at 00, 06, 12, and 18 UTC) extending forecasts up to 15 days with 64 ensemble members. The remaining hourly runs cover 48 hours. For fast-developing storms, this hourly refresh based on current satellite data makes a world of difference compared to the sluggish six-hour cycle.

Precipitation and Renewable Energy

Precipitation has always been a weak spot for global models, often producing blurry, indistinct storm boundaries. WeatherNext 3 tackles this by training on three data sources: ECMWF reanalysis, NASA's IMERG satellite data, and Google's own radar-based precipitation estimates. In short-term forecasts, Google reports improvements in the CRPS score of 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 up to 50% relative to traditional NWP baselines.

For renewable energy, the model outputs 100-meter wind speeds at turbine hub height, complete cloud cover at 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.

Availability and Limitations

Forecast data can be requested through Google Cloud's BigQuery, Earth Engine, and Cloud Storage. However, the model weights are not open-sourced, and custom inference still relies on WeatherNext 2 for now. So while the technology is impressive, it's not yet fully in the hands of the public.

Key Points

  • WeatherNext 3 offers 5 km resolution, a five-fold improvement over its predecessor.
  • The model updates hourly, using real-time satellite and ground data.
  • It shows significant gains in precipitation forecasting and renewable energy variables.
  • Data is accessible via Google Cloud, but weights remain proprietary.

This article is based on information from AIbase and Google's announcements.