Rapid, parallel CFD grid generation using octrees

As Large Eddy Simulation is increasingly applied to flows containing complex geometry, grid generation becomes difficult and time consuming when using software originally developed for RANS flow solvers. The traditional ‘pipeline’ approach of grid generation → solve → visualise entails the time consuming transfer of large files and conversion of file formats. This work demonstrates a grid generation methodology developed specifically to be integrated with parallel LES. The current approach is to use a Cartesian grid with adaptive refinement based upon geometry intersection, surface detail and surface curvature. The grid is defined by an octree data structure with the geometry defined by triangular facets using the STL file format. The result is a set of ‘cubical’ subdomains, each with identical numbers of cells and uniform distributions within the cube. Some subdomains will be entirely fluid and can be solved using straightforward CFD techniques, whilst some cubes will be cut by the surfaces. Individual cells are then tagged as ‘solid’, ‘fluid’ or ‘cut’ with the solver expected to use an immersed boundary approach to model the surface. A key feature is the design of the algorithm to be parallelisable on both shared and distributed memory systems. The distributed memory parallel dynamically partitions the grid as it is being generated, so that the partitioning is suitable for a subsequent flow solver. Grid generation testing has been carried out on a variety of input CAD files ranging up to 350,000 facets. A landing gear case shows how the grid generator correctly finds the fluid inside of the tire and other cavities within the hub. In scalar mode, a grid with 4,916 cubes and 468 million cells is generated in less than 100 seconds, whilst in parallel on 32 processor cores this can be achieved in 4·6 seconds.