Progress Report
Last updated:
Artificial generation of upstream maritime heavy rains to govern intense-rain-induced disasters over land (AMAGOI)[10] Environmental Monitoring
Progress until FY2025
1. Outline of the project
This project promotes public communication necessary for the realization of weather control, leveraging information science. By utilizing deep learning and visual information processing technologies, we aim to accelerate research on weather intervention through four-dimensional (4D) visualization of the processes behind local heavy rainfall and the effects of weather intervention. This visualization system will also contribute to the project's outreach promotions.
By FY2025, we constructed a four-dimensional visualization system integrating ray marching and neural fields for qualitative analysis of weather intervention effects. The introduction of ray marching and optimization of VRAM transfer processing achieved over 30 times higher data processing efficiency than conventional approaches. Moreover, temporal interpolation techniques based on advection modeling with neural fields enable smooth time-series visualization.
2. Outcome so far
(1) High-Precision Advection Modeling Using Neural Fields
In FY2025, we developed an advection modeling technique based on an interpolation method using neural fields with geometric transformations. The proposed method models spatiotemporal advection from grid-point data generated by numerical weather models and performs spatiotemporal interpolation, achieving higher accuracy (RMSE:0.70km/10min) than conventional linear interpolation methods (RMSE:0.92 km/10min). In particular, we successfully reduced the model size from 100.7MB for a linear interpolation model constructed from down sampled original data to 1.0MB as presented in Fig.1.
(2) Advanced Visualization of Weather Intervention Effects Using Ray Marching (Task 10-2)
To eliminate the discontinuities observed in conventional point-cloud-based visualization, we introduced a ray marching method implemented with custom shaders in Unity. As shown in Fig.2, atmospheric components could be represented as spatially continuous and smooth volumetric data, facilitating qualitative comparison before and after weather intervention. Furthermore, through runtime resampling and optimization of VRAM transfer processing, we achieved more than 30 times improvement in data processing efficiency compared with conventional methods. This demonstrated the feasibility of real-time processing for large-scale four-dimensional datasets on the order of 64GB.
(The heavy rainfall in Hiroshima in August 2014)
3. Future plans
In FY2026, we will advance the integration of the developed technologies and apply them to diverse case studies. First, we will further enhance the 4D visualization system integrating neural-field-based advection modeling and ray marching. In collaboration with meteorological researchers, we will obtain high-resolution numerical simulation data and improve temporal interpolation quality to realize more natural weather intervention simulations. Using this system, we will visualize multiple intervention scenarios, including cloud seeding and ground-wind control, and conduct multifaceted evaluations of weather control effects.
In addition, we will expand outreach activities through the application of the system to observational radar data, the publication of demonstration videos.