5:15 PM - 7:15 PM
[AAS01-P08] Evaluation of the Effectiveness of Black-Box Optimization Methods in Weather Intervention Design
Keywords:Weather Control, Black-Box Optimization Method, Warm Bubble Experiment, Real Atmosphere Experiment
Weather control involves complex and large-scale atmospheric phenomena, presenting three major challenges: (1) identifying effective interventions, (2) selecting feasible interventions, and (3) reducing the computational time required for intervention calculations. Addressing these challenges necessitates the application of control theory; however, conventional control theories struggle to directly handle large-scale phenomena such as weather.
To tackle these challenges, this study focuses on identifying effective interventions by optimizing parameters such as intervention location and intensity. In weather intervention, the large-scale and complex nature of atmospheric phenomena makes it difficult to obtain accurate gradient information for the objective function. Moreover, numerical weather prediction (NWP) model-based simulations require enormous computational resources, necessitating the identification of appropriate parameters while minimizing the number of function evaluations.
Under these constraints, black-box optimization methods offer a promising approach. These methods search for optimal input values that maximize or minimize an objective function using only input-output relationships. Since they do not require gradient information and enable efficient exploration by utilizing evaluation results, they are well-suited for weather intervention optimization. However, to the best of our knowledge, no prior studies have applied black-box optimization methods to weather intervention optimization, and it remains unclear which algorithms are most effective.
This study applies black-box optimization methods to weather intervention optimization and evaluates their effectiveness through two experiments using the SCALE-RM NWP model. The experiments include (1) the Warm Bubble Experiment, an idealized two-dimensional model with artificial initial conditions, and (2) the Real Atmosphere Experiment, a three-dimensional model reproducing realistic atmospheric behavior with observational data. As an evaluation metric, we adopt the reduction rate of cumulative precipitation caused by intervention and examine its variation across multiple function evaluations. Additionally, we compare the effectiveness of representative black-box optimization methods: Bayesian optimization, random search, particle swarm optimization, and genetic algorithms. The results show that Bayesian optimization and particle swarm optimization outperform other methods, highlighting their effectiveness in optimizing intervention locations and intensities.
This study contributes to the development of efficient weather intervention optimization methods under limited computational resources and broadens the applicability of black-box optimization methods to weather control.
To tackle these challenges, this study focuses on identifying effective interventions by optimizing parameters such as intervention location and intensity. In weather intervention, the large-scale and complex nature of atmospheric phenomena makes it difficult to obtain accurate gradient information for the objective function. Moreover, numerical weather prediction (NWP) model-based simulations require enormous computational resources, necessitating the identification of appropriate parameters while minimizing the number of function evaluations.
Under these constraints, black-box optimization methods offer a promising approach. These methods search for optimal input values that maximize or minimize an objective function using only input-output relationships. Since they do not require gradient information and enable efficient exploration by utilizing evaluation results, they are well-suited for weather intervention optimization. However, to the best of our knowledge, no prior studies have applied black-box optimization methods to weather intervention optimization, and it remains unclear which algorithms are most effective.
This study applies black-box optimization methods to weather intervention optimization and evaluates their effectiveness through two experiments using the SCALE-RM NWP model. The experiments include (1) the Warm Bubble Experiment, an idealized two-dimensional model with artificial initial conditions, and (2) the Real Atmosphere Experiment, a three-dimensional model reproducing realistic atmospheric behavior with observational data. As an evaluation metric, we adopt the reduction rate of cumulative precipitation caused by intervention and examine its variation across multiple function evaluations. Additionally, we compare the effectiveness of representative black-box optimization methods: Bayesian optimization, random search, particle swarm optimization, and genetic algorithms. The results show that Bayesian optimization and particle swarm optimization outperform other methods, highlighting their effectiveness in optimizing intervention locations and intensities.
This study contributes to the development of efficient weather intervention optimization methods under limited computational resources and broadens the applicability of black-box optimization methods to weather control.