Google has introduced GenCast, an advanced AI-powered weather forecasting model. In a paper published in Nature, Google detailed how GenCast uses a diffusion-based approach, specifically tuned for Earth’s geometry, to deliver precise weather predictions. By training the model on four decades of historical data from ECMWF archives, Google has enabled GenCast to outperform traditional systems like ECMWF’s ENS, especially for forecasts extending up to 25 days.
To evaluate its performance, Google conducted rigorous tests by running 1,320 forecasts from 2019 using GenCast, which had been trained on data up to 2018. The results showed that GenCast outdid ENS in 97.2% of cases and achieved up to 99.8% accuracy for predictions exceeding 36 hours. For example, when tasked with forecasting the path of Typhoon Hagibis in 2019, GenCast effectively narrowed its predictions as the storm approached landfall. This capability highlights its potential for improving disaster preparedness and response efforts.
Moreover, Google GenCast’s applications extend far beyond disaster management. The model effectively predicts wind speeds near wind farms and weather conditions over solar farms, offering valuable insights for renewable energy planning. GenCast generates over 50 predictions with varying probabilities as an ensemble system, providing comprehensive forecasts. Impressively, the model creates a 15-day forecast in just eight minutes on Google Cloud TPU v5, significantly reducing the time required compared to traditional supercomputing methods.
In addition, Google has embraced an open-access approach by releasing GenCast’s code and weights to the public. This move encourages collaboration with weather agencies and researchers worldwide. By working together, Google aims to refine GenCast further and push the boundaries of meteorological science. With its high level of precision, efficiency, and accessibility, GenCast represents a pivotal step forward in the evolution of weather forecasting.