Abstract

The solution of complex three-dimensional computational fluid dynamics (CFD) problems in general necessitates the use of a large number of mesh points to approximate directional flow features such as shocks, boundary layers, vortices and wakes. Such large grid sizes have motivated researchers to investigate methods of introducing very high aspect ratio elements to capture these features. In this Thesis, an anisotropic adaptive grid method has been developed for the solution of three-dimensional inviscid and viscous flows by the finite element method. An edge-based error estimate drives a mesh movement strategy that allows directional stretching and re-orientation of the grid with more mesh points introduced along those directions with rapidly changing gradients. The error estimate is built from a modified positive-definite form of the Hessian tensor of a selected solution variable or combination of variables. The resulting metric tensor controls the magnitude as well as, the direction of the grid stretching. The desired directionally adapted anisotropic mesh is constructed in physical space by a coordinate transformation based on this tensor. This research thus seeks a near-isotropic mesh in the transformed metric space and an equidistribution of the error over the mesh edges. The adaptive strategy can be considered to be the first 3-D implementation of an improved spring analogy-based algorithm originally applied on quadrilateral meshes. The adaptive methodology has been validated on various benchmark cases on both hexahedral and tetrahedral meshes. The numerical results obtained span inviscid and viscous flows, as well as internal and external aerodynamics. The effectiveness of the adaptive scheme to equidistribute the interpolation error over the edges of tetrahedral and hexahedral meshes has been gauged on analytical test cases where near-Gaussian distributions of the error were obtained. It was further demonstrated that the error estimate closely follows the true solution error. In analyzing the solution error of different sized non-adapted and adapted grids, one could not only achieve the same level of solution error by adapting and solving on a much coarser grid, but a significant reduction in solution time as well. All test cases revealed that the flow solver required lower amounts of artificial dissipation for solution on the final adapted grids. The current work should convincingly pave the way for its logical extension to unstructured grids, taking further advantage of refinement, coarsening and edge-swapping operations. It is strongly anticipated that this approach will shortly result in "optimal" grids.