Dynamically Adapting Mesh Refinement in an Unstructured Grid Global Model for Numerical Weather Prediction
Pak-Him Kam, Chi‐Yung Tam, Wai‐Pang Sze, Chi-Chiu Cheung, Ka Ki Ng, Sung-Ho Lee · Weather and Forecasting · 2025
Abstract Atmospheric circulation and weather systems often comprise phenomena of a variety of scales, in which there are mutual interactions. Adopting high spatial resolution can improve the representation of small-scale features and sometimes the overall performance of numerical weather prediction (NWP) systems; however, this requires substantial computational resources, posing a challenge in balancing resolution, efficiency, and accuracy. Various methods have been explored to address this issue, including static grid nesting and the use of unstructured grids in NWP. The use of adaptive mesh refinement (AMR) is another approach, but hitherto, this has mostly been applied on structured grid models. In fact, implementing AMR in the Model for Prediction Across Scales-Atmosphere (MPAS-A) can be challenging, in terms of mesh generation, domain regridding, and optimal time step management. In this study, AMR is adopted in a global unstructured grid NWP model, namely, ClusterTech Platform for Atmospheric Simulations (CPASs), which was developed based on MPAS-A. CPAS incorporates advanced techniques including customizable unstructured mesh generation (CUMG) and hierarchical time-stepping (HTS) to enable effective AMR, overcoming the limitations of MPAS-A. In a case study of tropical cyclone predictions, AMR is able to reduce both the intensity and track error. For predicting a cold surge event, AMR reduces the error of surface level variables and 850-hPa temperature. At the same time, the algorithm reduces the execution wall clock time by 12.3% and 14.0%, respectively. It is demonstrated that applying AMR on CPAS can further improve the latter’s forecast skill, in a computationally efficient way.