The Improvement of Computational Efficiency in KIM with a Variable-Resolution Grid

The Korean Integrated Model (KIM) is a global numerical weather prediction system that considers a cubed-sphere grid with uniform-resolution to ensure numerical stability and simplicity. However, high-resolution simulations significantly increase the number of grid points and reduce the time-step, which leads to high computational costs. To overcome this issue, we set the Korean Peninsula as an area of interest and implement a variable-resolution system in KIM to obtain high-resolution predictive performance in that area. The variableresolution system in KIM generates a stretched grid based on the Schmidt transformation. Although the total number of grids does not change compared to the reference resolution, the grid size in the high-resolution region become smaller due to the relaxation/contraction ratio, which affects the time-step. That is, the time-step becomes smaller than time-step of reference resolution, which increases the overall computational cost. To mitigate this limitation, we apply an adaptive time-step algorithm to KIM’s time integration scheme. This method dynamically adjusts the time-step based on the Courant-Friedrichs-Lewy (CFL) condition for each integration step. By allowing a larger time-step where possible, this approach reduces the total number of integration steps while preserving forecast accuracy.
 In this study, we will evaluate the computational efficiency of the variable-resolution system using KIM’s adaptive time-step algorithm through numerical simulations. Numerical results will show that this approach achieves similar prediction accuracy to that of the uniform high-resolution system in the area around the Korean Peninsula while significantly reducing the computational cost compared to the variable-resolution system using KIM’s static timestep. Therefore, it suggests that the variable-resolution system of KIM with adaptive timestep algorithm is an effective method for high-resolution prediction in the domain of interest if the overall computational resources are reduced while maintaining the prediction performance similar to that of a high-resolution uniform grid.