Progress Report
Last updated:
Artificial generation of upstream maritime heavy rains to govern intense-rain-induced disasters over land (AMAGOI)[3] Dimensionality Reduction
Progress until FY2025
1. Outline of the project
This research and development project aims to prevent damage from torrential rainfall over land by generating heavy rainfall over the ocean. To achieve this goal, it is necessary to guide the trajectory of weather away from a scenario that leads to heavy rainfall over land and toward a scenario in which heavy rainfall over the ocean reduces the amount of atmospheric water vapor, thereby preventing rainfall-related disasters. However, if meteorological data such as temperature, humidity, and wind direction are handled exhaustively, the resulting data become extremely high-dimensional, making prediction and control difficult. On the other hand, if similar meteorological fields evolve into several typical patterns, then efficient control may be possible by appropriately describing the “branching points of fate” among these scenarios (Fig.1). Accordingly, Research and Development Item 3 aims to acquire the essential low-dimensional degrees of freedom that are important for weather control, namely latent-space representations. To this end, we are pursuing three approaches: reservoir computing, Koopman mode decomposition, and landscape analysis.
2. Outcome so far
In the reservoir computing approach, we developed methods to extract and control low-dimensional structures related to heavy rainfall. Using the temporary low-dimensional structure of ensemble forecasts in chaotic systems, we derived control inputs to avoid extreme events. We also developed a bottom-up control method that dynamically selects optimal interventions from limited candidates (Fig. 2). SCALE-RM experiments on the 2015 Kanto-Tohoku heavy rainfall event suggested fractal-like sensitivity to small intervention differences.
In the Koopman mode decomposition approach, we developed a low-dimensional modeling method that represents nonlinear meteorological dynamics through the temporal evolution of observables (Fig.3). Warm-bubble simulations with SCALE-RM showed that the method can extract compact models while balancing accuracy and simplicity.
In the landscape analysis approach, we built a workflow combining PCA, manifold embedding, and geometric indicators to analyze state transitions and scenario branching in ensemble weather predictions. We also developed a mathematical theory to evaluate controllability under realistic control inputs (Fig.4).
3. Future plan
From the next fiscal year onward, we plan to use the insights obtained through the research conducted so far to formulate more concrete methods for acquiring latent-space representations of meteorological phenomena. We will also apply these methods to more complex data, including real meteorological data, verify their effectiveness, and advance research toward the realization of practical weather control.