Kingy AI
Subscribe, 2.1M subscribersWork with Kingy
AI News

GenCast’s 97.2% Weather Forecast Result: What It Means

December 11, 2024 AI News
Illustration of a weather-forecasting operations room

Correction: GenCast did not achieve a universal 97.2% weather-forecast accuracy rate. In Google DeepMind’s evaluation, it showed greater forecasting skill than ECMWF’s ENS system on 97.2% of 1,320 tested combinations of variables and lead times.

Google DeepMind introduced GenCast in December 2024 as a probabilistic machine-learning model for global medium-range weather forecasting. Instead of producing one definitive prediction, GenCast generates an ensemble of possible weather trajectories. That approach is designed to represent uncertainty, which is especially important when forecasting extreme weather.

Illustration accompanying Google DeepMind GenCast weather forecasting research

What the 97.2% result means

DeepMind and the authors of the peer-reviewed Nature paper compared GenCast with ENS, the European Centre for Medium-Range Weather Forecasts’ operational ensemble system. The evaluation covered 1,320 targets made up of different weather variables and forecast lead times.

GenCast performed better than ENS on 97.2% of those targets. For lead times longer than 36 hours, it performed better on 99.8% of the evaluated targets. These percentages describe how often GenCast outperformed the comparison system across the benchmark. They do not mean that 97.2% or 99.8% of individual forecasts were correct.

How GenCast works

GenCast is a diffusion-based probabilistic model adapted to the geometry of Earth. It produces global forecasts up to 15 days ahead at 0.25-degree resolution and advances in 12-hour steps. A GenCast ensemble contains 50 or more forecast trajectories, each representing a plausible future weather scenario.

The model was trained on ERA5 reanalysis data from 1979 through 2018 and evaluated using 2019 conditions. ERA5 is produced by ECMWF and reconstructs historical weather using observations and forecasting systems. The training data includes variables such as temperature, pressure, humidity and wind at the surface and at multiple atmospheric levels.

Speed and practical applications

According to the research paper, GenCast can produce one 15-day forecast in about eight minutes on a Google Cloud TPU v5 device. Multiple ensemble members can be generated in parallel. That makes the model substantially faster to run than large physics-based ensemble systems, although traditional forecasting systems remain essential for observations, initial conditions, training data and operational expertise.

The researchers also reported stronger results for several decision-relevant tasks, including extreme heat and wind, tropical-cyclone tracks and wind-power forecasting. An ensemble is useful in these settings because it shows a range of possible outcomes instead of hiding uncertainty behind a single forecast.

What the study does not establish

  • It does not show that every GenCast forecast is 97.2% accurate.
  • It does not establish one universal accuracy percentage for every location, variable or forecast horizon.
  • It does not make conventional meteorology obsolete. GenCast depends on weather observations, reanalysis data and initial conditions produced through the broader forecasting ecosystem.
  • The reported benchmark used historical 2019 conditions, so the result should be described as a research evaluation rather than a guarantee about a particular consumer forecast.

Bottom line

GenCast’s result is significant because the model outperformed a leading operational ensemble system across most of a large benchmark while generating forecasts quickly. The accurate way to report the headline number is that GenCast beat ENS on 97.2% of 1,320 evaluated targets—not that it predicted the weather correctly 97.2% of the time.

Primary sources