Google presented a new model Gencast artificial intelligence, which sets a new standard in weather forecasting. Developed by the DeepMind team, this technology exceeds existing systems, including the European system of medium-term forecasts (ENS), which is considered one of the most accurate in the world. A study published in Nature magazine showed that Gencast provides more accurate forecasts in 97.2% of cases when comparing ENS on data for 2019.
Unlike the previous version, which proposed one single event scenario, Gencast uses the ensemble approach and generates more than 50 different options. Experts can make a clear idea of how weather conditions can develop in the future, as well as evaluate possible risks and uncertainty.
Gencast is built on a diffusion model – technology that is already used to create images, video, and music. However, its key feature is adaptation to the spherical form of the Earth. The model studied at the 40-year-old data archive ERA5. It includes parameters such as temperature, wind speed, and pressure at different heights. Due to the highly resolution of forecasts (0.25 °), Gencast allows you to analyze in detail changes in the atmosphere.
To verify the accuracy of Gencast, they trained on data until 2018, and then tested on data for 2019. During the experiments, the model demonstrated superiority over the ENS in 1320 of various combinations of forecasts, which included different variables and temporary intervals. A particularly noticeable advantage was observed with forecasts for a period of more than 36 hours: there the accuracy of Gencast reached 99.8%.
One of the main strengths of the model is its ability to predict extreme weather phenomena. For example, abnormal heat, hurricanes, or strong winds. In the case of typhoon Hagibis, Gencast created an exact scheme of its trajectory for 7, 5, 3, and 1 day before going to land. The closer the event was, the more accurate the predictions became.
Gencast is a 15-day forecast in just 8 minutes on one TPU V5 TPU from the Google Cloud. Traditional physical models like ENS require several hours of work of supercomputer with thousands of processors.
Accuracy plays a decisive role in preventing the consequences of climatic disasters. From timely evacuation to energy planning: