Automated MPAS Mesh Generation: Herding Cats with the Push of a Button

The air-quality model development group at the U. S. Environmental Protection Agency (US EPA) is developing a new generation of global air quality models using the Model for Prediction Across Scales (MPAS) atmospheric component as the meteorological driver. MPAS is a collaborative model development project led primarily by the National Center for Atmospheric Research (NCAR) and the Los Alamos National Laboratory to provide atmosphere, ocean, and other earth-system components within a unified modeling framework. MPAS offers gradual refinement of horizontal resolution thus avoiding discontinuities that have presented difficulties for nested-domain modeling applications.

The computational framework of MPAS is based on Spherical Centroidal Voronoi Tessellations (SCVTs) and is composed of an irregular mesh of polygons (mostly hexagons) with specific properties that allow MPAS simulations to be mass conservative, an important property for air quality modeling. These polygons can vary in size to provide mesh refinement where needed and gradually transition to a coarser mesh covering the rest of the globe. NCAR is heading the development of the atmospheric component of MPAS which is named MPAS-Atmosphere (MPAS-A). The variable-resolution meshes currently available from NCAR have refinement areas of simple shape (nominally circular) and sized to cover continental areas. These refined areas are generally appropriate for meteorological modeling but do not satisfy the needs of air quality modelers who desire mesh refinement over areas with important emissions of pollutants or their precursors, or sensitive receptors of those pollutants. Thus, the US EPA has endeavored to generate its own MPAS meshes for specific air-quality applications.

MPAS developers at NCAR provided a method description and the computing codes they use for mesh generation. The process for creating variable-resolution SCVTs involves a number of actions in a specific sequence as this presentation will demonstrate. The most critical step is an iterative relaxation based on Lloyd’s method that adjusts the SVT generating points based on a density function specified for the desired refinement pattern. Occasionally, an iterative relaxation will fail to converge sufficiently to the desired solution and the entire process must be restarted. The iterative relaxation is computationally expensive and only partially applicable to parallel computing. Computing a solution for a large and complex mesh can take months on a modern compute server. Experiments have been conducted to identify the conditions leading to convergence failures. As will be shown, the results provide some clues. However, specific conditions for success or failure have yet to been found. Guidance from NCAR indicates that density functions with continuous first derivatives “seem to work well”, but examples will be shown where this is not a sufficient condition for convergence. There are also questions about the convergence criteria used to determine when the SCVT is sufficiently centroidal. Because of the complex nature of the meshes desired and the critical need for mass conservation in air quality modeling, work is continuing to find answers to these questions.