DeepMind Releases Typhoon Prediction AI as Open Source

Google DeepMind's weather prediction AI model 'WeatherNext' has been revealed to have the capability to simultaneously predict the trajectory and intensity of tropical cyclones (typhoons, hurricanes, and similar storms). The model can make accurate predictions approximately one day further into the future compared to current major weather prediction systems, and both the code and model weights (trained parameters) have been made publicly available as open source on GitHub.

The improvement of weather prediction accuracy is a field that has advanced gradually through the accumulation of dedicated research over many years. Traditional weather models based on numerical calculations apply vast amounts of atmospheric data to physical laws, requiring massive supercomputers and several hours of computation time to achieve high accuracy. According to DeepMind, the 'approximately one day improvement in prediction accuracy' achieved by WeatherNext is equivalent to approximately ten years of progress accumulated through conventional methods.

A major characteristic of WeatherNext is its ability to simultaneously predict both the 'trajectory' and 'intensity (strength)' of tropical cyclones. In conventional models, trajectory prediction and intensity prediction have typically been handled by separate systems, and simultaneous high-precision prediction of both has been considered a difficult challenge. WeatherNext addresses both of these aspects in an integrated manner, enabling more comprehensive typhoon forecasting.

A notable technical point is that the model's trained weights and source code are made available on GitHub. This enables research institutions, weather agencies, universities, and similar organizations to operate the model independently or make their own improvements. The publication of technology as open source, without being premised on commercial use, is expected to lead to enhanced capabilities across the entire research community for AI weather prediction.

Weather prediction accuracy, particularly typhoon prediction accuracy, is a social issue directly relevant to disaster prevention, evacuation planning, agriculture, logistics, and many other fields. Cases exist where minor deviations in trajectory or errors in intensity estimation lead to delayed evacuation orders or insufficient countermeasures. In this context, the significance of obtaining accurate predictions 'approximately one day earlier' compared to conventional methods is substantial. In the AI weather field, major weather agencies in Western countries have also been advancing the adoption and evaluation of AI models, and DeepMind's efforts can be positioned as aligned with this trend.

A key point of future interest is how WeatherNext will be utilized in actual meteorological operations and disaster prevention activities. By making it available as open source, an environment has been created in which researchers and weather agencies from various countries can independently evaluate and improve it. It can be said that AI-based weather prediction is entering a phase of transition from 'research-level achievements' to 'tools used in the field'.