...
Here is a low-resolution CAM-SE grid. This is a "ne4np4" grid, meaning that it has 4x4 spectral elements in each cube face (the "ne" value), and each spectral element has a grid of 4x4 (the "np" value) Gauss-Lobatto points inside it. The blue lines show the edges of the spectral elements. The green dots show the Gauss-Lobatto points-Legendre (GLL) nodes. CAM-SE is "collocated", meaning that all variables (U,V,T, etc...) are carried around on the green dotsGLL nodes. CAM physics is computed in columns located at all the green pointsat GLL nodes. Each column has a Gauss-Lobatto weight associated with it, which is also that columns area. The sum of these weights is 4pi (the area of the unit sphere). Note that the Gauss-Lobatto points One can see from the figure that GLL nodes are not quite equally spaced - they cluster at the edges of the spectral elements.
...
Most analysis tools and other tools can not handle higher order elements such as used in CAM-SE. For these codes, we produce metadata that will divide each spectral element into a subcellsubcells. This subcell grid is shown below. Note that the Gauss-Lobatto points (where CAM physics is computed) The GLL nodes are the vertices of this the subcell grid. This meta data can be used by Paraview and Visit to plot native grid CAM-SE output. The metadata is stored in a grid template file that typically has a file name of the form "<gridname>_latlon_<date>.nc" in the grids directory of the ACME inputdata server.
...
For conservative remapping, the areas of these polygons must match the Gauss-Lobatto weight of each Gauss-Lobatto GLL node. If we construct this dual grid in the usual way, by connecting the centers of all the subcells shown in the subcell grid, the areas of the cells will in general not match the Gauss-Lobatto GLL weights. We thus have to perform an iteration, similar to spring dynamics, tweaking the polygons until the areas are correct. We have a couple of algorithms to do this, but the resulting polygons can be a little odd. The result of one of these algorithms is shown below. For For this figure, we have allowed pentagons and hexagons so that the algorithm will converge faster and to more uniform cells, but this means that some of them will be slightly non-convex. ESMF can handle non-convex cells, but other utilities may require they be convex.
...
Also, the "physics grid" work, when completed, will allow us to run the CAM physics at cell centers of the subcell grid (instead of vertices). It will also allow us to run the physics on a subcell grid with equally spaced subcell centers. Either of these options will mean we would no longer need the SE dual grid.
...