The goal of this research project is to prevent heavy rainfall damage on land by generating torrential rain over the sea. To achieve this, we aim to redirect weather scenarios from those leading to intense rainfall on land to ones where offshore rainstorms reduce atmospheric water vapor and mitigate disaster risks. However, comprehensive meteorological data—such as temperature, humidity, and wind direction at each location—are extremely high-dimensional, making prediction and control challenging. If future developments of similar meteorological fields can be grouped into a few typical patterns—one of which results in undesirable outcomes—then describing these abstracted scenarios could improve control efficiency. In Research Item 3, we aim to extract a small number of essential degrees of freedom that govern such branching (Fig.1). We refer to this as a latent space representation. To obtain latent space representations potentially useful for weather control, we are exploring three approaches: reservoir computing, Koopman mode decomposition, and landscape analysis.

In the reservoir computing approach, we developed a method to reduce meteorological data dimensionality using deep learning and predict its evolution via reservoir computing. For example, we compressed time-series data (like video frames) to extract key features, then learned and predicted their dynamics (Fig.2). We also devised a control strategy focusing on the direction of greatest variation in meteorological states for more efficient guidance.