This project will build the WeatherGenerator – the world’s best Generative Foundation MachineLearning Model ofthe Earth system – that will serve as Digital Twin for Destination Earth. Similar to the leap in quality that was achieved when foundation models were replacing conventional machinelearningmodels in language modelling, the WeatherGenerator will provide machinelearning applications in Earth system science with improved skill at lower computational costs compared to the state-of-the-art. The WeatherGenerator will be based on representation learning to create a task-independent tool that modelsthe dynamics ofthe Earth system much more faithfully and robustly compared to conventional machinelearning tools that are trained for a specific task. The WeatherGenerator is generic and task-independent as the large pre-trained model can be adjusted to specific needs for a wide range of application areas in Earth sciences. Thereby, it will likely be much more resilient for climate when the underlying data distributions are changing and to niche applications where training data is limited. Through its generality and robustness, the WeatherGenerator will form a machine-learned Digital Twin ofthe Earth that will be incorporated into the Digital Twin family of Destination Earth. The WeatherGenerator will be tested in a wide range of application spanning weather prediction, renewable energy generation, flooding, food security, and health.