Google is turning the idea of weather forecasts upside down

Google unveiled GenCast, a new AI weather forecasting model that can predict climate risks with high accuracy. Unlike traditional methods, GenCast uses ensemble forecasting, which allows for multiple possible scenarios to be considered at once. The model exceeds current accuracy standards by 97.2% for forecasts up to 15 days in advance.

GenCast is based on diffusion model technology widely used for imaging and video, but adapted to Earth's spherical geometry. Trained on 40 years of data, the model analyses key indicators such as temperature, wind and pressure and generates forecasts at 0.25° resolution - much faster and more efficient than traditional models that require powerful supercomputers.

Using Google Cloud TPU v5, GenCast generates a full ensemble forecast in 8 minutes, whereas traditional methods can take hours.

GenCast is particularly effective at predicting extreme weather events, including hurricanes, high winds and extreme heat. For example, its ability to accurately predict the path of Typhoon Hagibis days before it hit Japan highlights the practical relevance of the model.

Google is making GenCast open source: code, model weights and predictions have been published. This opens up opportunities to integrate AI predictions into research and practice, including climate science, resource management and disaster prevention.