Google unveils yet another AI model for weather forecasting: GenCast.

The refinement of models using generative AI to predict weather and extreme events shows no signs of slowing down.

In December, Google introduced yet another advancement in this field with GenCast, the successor to GraphCast. According to Google, the new model represents a significant step forward because GenCast changes the approach to data usage and output generation.

While GraphCast was a deterministic model designed to provide a single, highly accurate forecast, GenCast takes a completely different approach: starting from the initial data, the model generates 50 or more predictions. This change reflects Google’s recognition that, since achieving a perfect long-term forecast is nearly impossible, decision-makers need to rely on probabilistic ensemble predictions.

GenCast was trained using a dataset comprising four decades of historical data from the ERA5 archive of the ECMWF. These data include variables such as temperature, wind speed, and atmospheric pressure at various altitudes. To evaluate the model’s effectiveness, the dataset was used up to 2018, enabling testing of GenCast with 2019 data and comparing its forecasts to those of the ECMWF’s ENS model. The results demonstrated GenCast’s overall superiority over the traditional model.

The new model can process an input to generate a series of predictions that outline a probabilistic picture of the area affected by a meteorological phenomenon. According to Google, GenCast is designed to provide a sufficiently reliable trend regarding the affected area.

Another noteworthy aspect is GenCast’s computational speed: it takes only eight minutes to produce a 15-day forecast using a single Google Cloud TPU v5. In contrast, traditional physics-based ensemble forecasts require hours and the use of supercomputers with tens of thousands of processors.