MPAS variable-resolution mesh generation

• MPAS-Atmosphere Mesh Generation | Michael G. Duda | NCAR/MMM | MPAS-A and MPAS-JEDI Tutorials, 23-26 October 2023, Taiwan
  • we employ Lloyd’s Method (Originally developed by Stuart Lloyd at Bell Labs for optimal signal quantization)
  • Density function: the key to refinement
  • generating a mesh is still very slow
  • Can take several months on a single multi-core desktop for high-resolution meshes
  • Even with 24 threads on the desktop in my office, generating a 15-3 km mesh with ~6.5 M cells still took months!
• MPAS variable-resolution mesh generation | Feb 2023
  • we don't yet support (or even make available) the software that we use internally to generate MPAS-A meshes
  • Also, note that the lead time to generate this mesh will likely be several months
  • The generation of an SCVT can be viewed as an optimization problem, and the method that we use to perform this optimization -- Lloyd's method -- requires a very large number of iterations (typically O(10^6)), and each iteration can take several seconds. Part of each iteration can be parallelized, but there is still a part that is done in serial with our current software
• 20-km benchmark case | March 2022
  • At present we don't have a 20-km benchmark case, and in order to generate this yourself you would need a 20-km mesh. While we don't actually have a 20-km quasi-uniform mesh at present, we could generate one and make it available on the mesh download page in the near future. I'll follow up in this thread once the 20-km mesh is available.
  • We don't currently provide software for generating new meshes, so generating a 20-km mesh is something that we will do for you.
  • We haven't made our mesh generation software publicly available, so once I have generated the 20-km mesh, I'll follow up here to let you know that the mesh is available.

The generation of an SCVT can be viewed as an optimization problem, and the method that we use to perform this optimization -- Lloyd's method -- requires a very large number of iterations (typically O(10^6)), and each iteration can take several seconds. Part of each iteration can be parallelized, but there is still a part that is done in serial with our current software