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 just gotten a whole lot clearer. Google DeepMind, in collaboration with Google Research, has unveiled WeatherNext 3, a new AI model that promises to shake up how we predict the weather on a global scale.

What sets WeatherNext 3 apart? For starters, it doesn't rely on the same old data sources that have long been the backbone of meteorological AI. Traditional models often depend on reanalysis grids from numerical weather prediction, which can smooth over the subtle local quirks caused by coastlines, valleys, and mountains. WeatherNext 3, however, goes straight to the source: it pulls in real-time geostationary satellite mosaics and raw weather station observations. That means its forecasts are grounded in what's actually happening on the ground, not just what a model thinks should happen.

The result? A model that can generate forecasts with a resolution of about 5 kilometers—roughly five times finer than the previous 25-kilometer standard. And it doesn't stop there. WeatherNext 3 updates its forecasts every hour, a far cry from the six-hour cycles we're used to. This kind of rapid refresh is a game-changer for tracking fast-moving storms and other rapidly developing weather events.

But resolution and speed are only part of the story. WeatherNext 3 is also a multitasker. With a single pass, it can output three levels of forecast data: fine-grained temperature and dew point at 5 kilometers, surface details like wind speed and precipitation at 10 kilometers, and broader atmospheric fields at 25 kilometers. This multi-resolution approach means it can serve a wide range of users, from local farmers checking frost risk to energy companies planning wind and solar output.

Speaking of energy, WeatherNext 3 has some tricks up its sleeve for the renewable sector. It can accurately predict wind speeds at turbine hub heights, cloud cover at various altitudes, and all components of solar irradiance. For grid operators, that's like having a crystal ball for wind and solar power generation—crucial for balancing supply and demand as we lean more on clean energy.

Precipitation forecasting has always been a thorn in the side of global models, often producing blurry, unreliable fields. WeatherNext 3 tackles this by training on a rich mix of data, including ECMWF reanalysis, NASA's IMERG satellite data, and Google's own radar-based precipitation estimates. Early results are promising: its probability scores improve by up to 60% compared to the IMERG benchmark, and it also holds its own against rain gauge measurements.

So, when can you get your hands on this? The forecast data from WeatherNext 3 is already available through Google's BigQuery, Earth Engine, and Cloud Storage. However, the model's inner workings—the weights—are not open-sourced, and for custom inference, you'll still need to use the older WeatherNext 2. Still, as WeatherNext 3 rolls out fully, it's clear that the era of high-precision, high-timeliness weather forecasting is just around the corner.

Key Points

  • Breakthrough Resolution: WeatherNext 3 delivers forecasts at 5km resolution, five times sharper than previous models.
  • Hourly Updates: The model refreshes every hour, enabling better tracking of fast-changing weather.
  • Multi-Resolution Output: Generates three levels of forecast data in one pass, catering to diverse needs.
  • Renewable Energy Support: Accurately predicts wind, cloud, and solar variables to aid clean energy management.
  • Improved Precipitation Forecasts: Significant gains in accuracy over existing benchmarks.
  • Data Availability: Forecasts are accessible via BigQuery, Earth Engine, and Cloud Storage, though model weights remain proprietary.