Exploring the Effects of Multiple Refinement Regions on Computational Efficiency and Forecast Accuracy in MPAS-Atmosphere

Variable-resolution (VR) meshes allow high resolution in regions of interest without requiring the immense computational power that is needed for uniformly high-resolution meshes. In regions with refined resolution, more accurate representations of fine-scale physical processes can be captured, resulting in improved forecast accuracy. In this study, the effects of multiple refinement regions on forecast accuracy and computational efficiency in MPAS-Atmosphere is explored. We generate global variable-resolution meshes with different configurations of refined regions, where the refined regions exhibit a Gaussian-bell shaped profile with radii ranging from 300 km to 700 km. Multiple refinement regions with a minimum grid spacing of 4 km or 1 km are located over the U.S. continent, whereas outside of the refinement regions have a maximum grid spacing of 20 km or 40 km. The computational efficiency is evaluated with respect to the number of the multi-refinement regions and stable time step size. The increment in the number of global cells when adding a single 4 km refinement region is found to be less than 3% compared to the quasi-uniform (QU) 20 km mesh, whereas the maximum allowable time step size for the VR mesh must be reduced by up to 75% for stable time integration. Therefore, the computational efficiency of MPAS-Atmosphere is primarily influenced by the model's time step size rather than the number of global cells. Since the time step size remains unchanged when adding multiple refinement regions with the same minimum grid spacing of 4 km, it is observed that adding 5 (10) refinement regions to the QU 20 km mesh increases the runtime by only 11% (21%) compared to a single refinement region. A heavy rainfall event over the U.S. continent demonstrates the impact of refinement regions on forecast accuracy in MPAS-Atmosphere. The daily accumulated precipitation shows significant improvement in the refinement regions compared to the QU 20 km mesh. This study will contribute to identifying the optimal balance between increasing resolution across multiple regions and minimizing runtime penalties through the incorporation of VR meshes.