Developing a Weather Control Computation System for Estimating Optimal Control Inputs
To realize weather control, it is necessary to predict how the weather will respond when control inputs are given and to estimate when, where, and how we should add signals to the atmosphere to achieve desirable weather. However, there is currently no system capable of handling these tasks. In Item 4, we aim to develop a weather control computation system that can simulate the impact of control inputs on weather based on existing numerical weather prediction systems and estimate optimal control inputs.
In addition, Item 4 addresses the acceleration of prediction and control computations. To achieve weather control, as mentioned above, these computations must be completed before a disaster occurs or before the appropriate timing for control is missed. However, running numerical models used for weather prediction generally requires significant computational cost. With current computational resources and mainstream computational algorithms, it is anticipated that it would be difficult to complete these computations in a realistic timeframe. Therefore, in Item 4, we will introduce surrogate models and latent space representation techniques derived from mathematical research and deep leaning into the prediction and control computation system. This will enable the calculation of control inputs in a realistic timeframe while evaluating effective elemental technologies for weather control computation. We also aim to utilize quantum computers, which are cutting-edge computational technology, to optimize model predictive control and data assimilation calculations by mapping them to the Ising model and accelerating computations through quantum annealing.
Item 4-1Developing weather control computation system
Principal investigator: Atsushi Okazaki
Outline
This study develops a weather control computation system based on a regional weather prediction system (specifically SCALE-LETKF) that enables the estimation of control inputs and the execution of intervention experiments. We also develop data assimilation methods that effectively utilize ensemble intervention experiments required for a weather control computation system, improving the accuracy of initial value estimation and weather predictions.
Methods
We develop a weather control computation system based on a regional weather prediction system that can handle various control inputs, such as floating dome formation, cold pool formation, sea surface cooling, microwave heating, cloud seeding. We assume SCALE-LETKF as the regional weather prediction system. Estimating appropriate control inputs, such as when, where, and how to control, is expected to require ensemble prediction. We explore methods to generate high-quality ensembles (e.g., Multi-Physics Multi-Parameter Ensemble), and develop and upgrade data assimilation methods (e.g., local particle filters and four-dimensional ensemble variation methods) to effectively utilize these ensembles.
Importance
To realize weather control, a prediction system is needed that can estimate appropriate control inputs, determining when, where, and how to control, and conduct weather predictions based on these inputs within a single platform. The prediction system needs to be equipped with an actuator function that transmits control inputs to the weather model. However, existing weather prediction systems do not possess the functions. This study aims to develop a weather control computation system that enables the estimation of weather control inputs and their implementation, which is essential for the success of the project scenario. In addition, examining the effectiveness of weather control methods using on real atmospheric conditions with physical experiments is impractical due to immense effort and cost. However, this study enables a more efficient evaluation of these methods.
Expected problems and solutions
When developing complex and large systems such as weather prediction systems, diagnosing issues is expected to be challenging if problems arise. Therefore, we first develop ideal systems using simple models, subsequently address issues regarding real atmospheric conditions. Some interventions may be difficult to directly represent using existing weather prediction systems. We seek solutions by collaborating with other items and research teams to parameterize the effects of interventions and incorporate them into the prediction system.
Members
Item 4-2Reducing weather control computation demanding
Principal investigator: Shunji Kotsuki
Outline
This study introduces techniques such as surrogate models and latent space representations obtained from mathematical research and deep leaning into the weather prediction computation system. This approach enables the calculation of control inputs within a realistic computation time while evaluating effective technological elements for weather control computations. In addition, we develop techniques formulating the optimization calculations of model predictive control and data assimilation into Ising model, and enabling acceleration of computations through quantum annealing.
Methods
We select among multiple methods in parallel to hedge risks while deriving optimal solutions. This includes Model Predictive Control (MPC) using tangent linear and adjoint models, MPC reducing dimensionality of weather data to narrow the search space, computation acceleration or exploration of control inputs by using deep learning, surrogate models, and quantum annealing.
At the start of the research, no specific numerical targets will be established, because the scale of heavy rainfall over the sea and the reduction in rainfall over land for disaster mitigation are expected to vary depending on the specific disaster events. By the FY 2024, we select cases with high potential to generate heavy rainfall over the sea based on the results of Item 5-2. Flood inundation and economic damage estimates for the selected cases will be calculated by Item 8-1 and 8-2 to investigate the extent to which the heavy rainfall generation over the sea and rainfall reduction over land are required. Then, we set target values for the rainfall reduction to be achieved for each selected cases at an early stage, to serve as our goals.
Importance
The development of a computationally less demanding control input computation methods and the development of techniques for accelerating control input calculations are both essential to demonstrate the project goal on a computer, which is ‘reducing rainfall damage by heavy rainfall generation over the sea and rainfall reduction on land.’ nonlinear dynamic system control Nonlinear mechanics control, particularly Model Predictive Control is already established in engineering. However, due to the high dimensionality of the control search space, the enormous computation demands of repeated calculations can be a bottleneck. Simultaneous development of effective methods for reducing the dimensionality of the control search space and techniques for accelerating computation speed is essential.
Expected problems and solutions
The calculation of control inputs using model predictive control can become challenging if the numerical models or initial values are incomplete. In such cases, finding appropriate input values may be difficult. Therefore, we start the research from the experimental settings where the completeness of the numerical models and initial values are ensured. Then, we address more difficult problems. To hedge risks, we do not only focus on high-performance feedback control, but explore simpler and computationally less-demanding methods, such as feedforward control that explores the timing and location of one-time interventions, as well as ensemble sensitivity analysis.
Members
Item 4-3Developing observation simulator for monitoring heavy rainfall generation over the sea
Principal investigator: Kaya Kanemaru
Outline
This study considers a comprehensive observation system to monitor the locations of weather interventions aimed at generating heavy rainfall over the sea and to assess the effects of these intervention. Specifically, we improve existing weather radar simulator to develop a new observation simulator, where new observation methods, such as lidar for measuring water vapor distribution around clouds and radar for assessing water vapor distribution within clouds can be experimented numerically on a computer. By using the observation simulator, we explore the necessary observation methods for monitoring heavy rainfall generation over the sea. We combine existing and new observation systems to consider the feasibility including methods, costs, and required number of units.
Methods
We propose a monitoring method for heavy rainfall generation over the sea that combines both existing and new observation systems. To examine it, we enhance the function of the weather radar simulator (specifically, the POLARRIS module included in the Joint-Simulator package) being developed in Japan, creating an observation simulator that can demonstrate new observation methods, such as lidar for assessing water vapor distribution around clouds and radar for measuring water vapor distribution within clouds. In addition, since the location and effects of intervention methods are expected to change from time to time, we consider the specifications of an observation system that uses geostationary satellite observation and small satellite constellation data, enabling wide-area and high-frequency observation over the sea, while incorporating these new observation methods.
For efficient research development, the development of the observation simulator will proceed in coordination with atmospheric heating simulator under research and development of Item 6. Additionally, the consideration of a comprehensive observation system, which includes new observation methods, will be worked with Item 4-1 “Developing weather control computation system” and Item 5. The observation simulator being developed will be integrated as an observation operator into the weather control computation system.
Importance
To implement weather control, it is essential to identify intervention locations and conduct monitoring to verify the effects of those interventions. Monitoring of heavy rainfall over the sea, a target phenomenon of weather control, is challenging due to the difficulty of deploying observation over the sea. Therefore, it is necessary to develop a comprehensive observation system that combines various exiting and new observation methods to overcome these deployment issues.
In addition, the observation simulator developed for this purpose can also be used as an observable operator of the weather control computation system. The information obtained through monitoring includes data on changes in the weather conditions resulting from interventions. Since intervention strategies may be updated continuously, the implementation of observable operator will be one of the essential function necessary for weather control computation system.
Expected problems and solutions
There is a concern that accurately determining the variables to be monitored within the rain clouds that cause heavy rainfall over the sea will be challenging. In such cases, we solve the problems by combining information on wind and water vapor from the surroundings of the rain clouds, as well as partially available data on wind and water vapor within the rain clouds, with information from numerical weather models, under the collaboration with other member of Item 4.
Members
Item 4-4Development of rapid-update cloud & precipitation prediction systems
Principal investigator: Shigenori Otsuka
Outline
This project aims to conduct research and development to integrate three-dimensional precipitation nowcasting using high-frequency observations from phased-array weather radar and the Himawari meteorological satellite, as well as three-dimensional cloud and precipitation forecasting using SCALE-LETKF, into a framework for artificial generation of upstream maritime heavy rains. In addition, by combining conventional methods with machine learning approaches, we seek to improve forecast accuracy while maintaining interpretability. The project will provide forecast information necessary for decision-making during outdoor experiments on artificial generation of upstream maritime heavy rains, and will also advance SCALE-LETKF-MPC to contribute to the development of weather control systems.
Methods
- Three-dimensional nowcasting based on advection and machine learning
For very short-term weather prediction, it is well known that nowcasting, that is, spatiotemporal extrapolation based on observational data, outperforms numerical weather prediction (NWP) models in forecast accuracy. Furthermore, it has been demonstrated that three-dimensional prediction using phased-array weather radar is more effective for precipitation forecasting than conventional methods based on two-dimensional radar and satellite imagery. However, we cannot predict cloud regions with precipitation radars before precipitation begins. Therefore, this study aims to achieve integrated three-dimensional nowcasting of clouds and precipitation by combining satellite-based cloud imagery with three-dimensional precipitation observations from phased-array weather radar. Additionally, the knowledge gained from assimilating these observations will be fed back into the SCALE-LETKF-MPC system being developed in related tasks such as Task 4-1. - Rapid update numerical weather prediction system
While nowcasting is useful for real-time monitoring and decision-making, it does not explicitly solve physical equations and is therefore unsuitable for control applications requiring highly accurate physics-based predictions. Accordingly, in collaboration with other tasks, we will promote research on assimilating high-frequency observations from phased-array weather radar and the Himawari satellite into the numerical weather prediction and data assimilation system SCALE-LETKF. High-precision calculations in NWP systems are computationally expensive, and real-time forecasting currently requires large-scale computational resources. Therefore, we will advance research on replacing computationally intensive components with machine-learning-based surrogate models.
Importance
To analyze, predict, and control rapidly evolving rain clouds, a forecasting system that updates at high frequency is essential. The typical timescale of cumulus convection is around one hour; however, convective cells within systems that produce heavy rainfall can develop rapidly within 5 to 10 minutes. As a result, conventional observations and monitoring at 5-10 minute intervals often capture phenomena only after they have already reached their peak. Therefore, it is necessary to establish monitoring systems based on higher-frequency observations. Key observations include phased-array weather radar, which can obtain three-dimensional precipitation echoes at 30-second intervals, and the Himawari-8/9 meteorological satellites, which can capture images over Japan every 2.5 minutes at up to 500 m resolution. These observations are not currently utilized in the Japan Meteorological Agency’s operational precipitation nowcasting. At the research level, however, the principal investigator has implemented three-dimensional precipitation nowcasting using X-band phased-array weather radar observations, and the National Institute of Information and Communications Technology has developed a machine-learning-based precipitation nowcasting system; both have demonstrated their effectiveness.
Expected problems and solutions
- Adaptation to C-band phased-array weather radar
Considering seeding experiments in Toyama Bay, C-band phased-array weather radar is expected to be used as a source of high-frequency observations. Its characteristics are expected to differ from those of the X-band phased-array weather radar used in previous demonstrations. Rain attenuation is expected to be less significant in the C-band than in the X-band, but the characteristics of observation errors and other factors still need to be investigated. In addition to the difference in frequency band, the radar is developed by a different manufacturer, necessitating a review of quality control methods for observational data. - Integration of two-dimensional and three-dimensional observations
Observations from the Himawari meteorological satellite are two-dimensional images, whereas phased-array weather radar provides three-dimensional observations; however, the final output required is the three-dimensional distribution of clouds and precipitation. While this is not an issue in numerical weather prediction systems, differences in data structure must be addressed in the development of nowcasting methods. Approaches such as using machine learning techniques to handle heterogeneous data formats in a unified manner will be required.
Members
