In this section we discuss several methods for expressing a cell-face value of a scalar in terms of surrounding nodal values. For clarity, the discussion relates to a uniform infinite staggered grid as depicted in Figure 5.1. For reasons of robustness and algorithmic simplicity, multidimensional central and upwind schemes are treated by one-dimensional decomposition in the normal direction for each cell face. Such schemes are called splitting schemes. Therefore we restrict ourselves to interpolation of the face values along one specific direction in G-space, taking for example the -direction. We need to consider only the cell-face value . The other face values will be treated in the same way.
Figure 5.1: One-dimensional staggered grid showing the nodes involved in the
evaluation of at cell-face .
{
In this scheme, the face value is approximated by means of linear
interpolation in the following way:
resulting in second order accuracy. This scheme is obviously conservative and is not positive when a mesh-Péclet number defined as the ratio of the contributions to the convection and the diffusion exceeds a certain value. For example, when orthogonal grids are employed the scheme may produce wiggles when the mesh-Péclet number given by (see (5.5) and (5.4)):
becomes greater than 1 in magnitude. In the case of skewed grids the expression
for the mesh-Péclet number depends on how the mixed-derivative terms are
approximated. This will be discussed in Section 5.4.
{
The face value is equal to the upstream node value and, for reasons
of efficient vector coding, can be expressed as follows
This scheme is conservative, first order accurate, and unconditionally positive.
However, when the flow direction is oblique to the grid lines this scheme
produces numerical cross-flow diffusion. This may result in a large error in
the solution.
{
In an effort to combine the advantages of both central and upwind schemes
Spalding [28] has proposed the hybrid central/upwind
difference scheme. The approximation to is given by
where is given by (5.22), by (5.24) and s is a switching function depending on the mesh-Péclet number Pe with . With the hybrid scheme the convection terms are approximated with central differences, unless |Pe| > 1 when a switch to the first order upwind scheme is applied:
For convergence reasons, a smooth switching function may be used:
This switching function is obtained by requiring that off-diagonal matrix
elements involving convective and diffusive contributions are non-positive,
which is sufficient for suppression of wiggles.
A disadvantage of the hybrid scheme is that in convection-dominated flows
first order upwind is employed everywhere,
irrespective of whether spurious wiggles arise or not.
A number of higher order convection schemes have been embedded in the
computational procedure. Also, a number of classes of flux-limited schemes based
on the TVD methodology will be presented. All may conveniently
be written in a canonical form involving a few scheme-related parameters. To this
end, the face values are approximated by the first order upwind
scheme, corrected by adding an appropriate anti-diffusive flux controlled by a
limiter. That is, is approximated by
where
Formula (5.28) opens the possibility to incorporate arbitrary upwind biased schemes in an algorithmically simple way. The most interesting schemes are
where , and . These classes of flux limiters bring together most limiters known in the literature. For example, and define the Minmod and Superbee, respectively. Details can be found in [30, 9]. The functions R-0 and R- are Van Leer [9] and ISNAS [41] limiters. The limited scheme proposed by Koren [13] and SMART scheme [6] are identical to PL- with M = 2 and PL- with M = 4, respectively. Finally, by taking , M = 2 and , M = 2 for symmetric PL- limiters the UMIST [16] and MUSCL [35] schemes are recovered. Apart from linear schemes, USR and USPL limiters, the above classes of limiters reduce locally to first order accuracy at physical extrema regardless of the order of accuracy in regions of monotonicity. If USR and USPL limiters are employed uniform second order accuracy is obtained ( ). For more details we refer to [42], where further extensive references can be found.
Generally, the function is nonlinear, and more than 2 immediate neighbouring nodal points per spatial direction may be involved in approximating the convective flux . Difficulties for iterative solution methods can be circumvented by writing the flux in terms of a lower order approximation plus a correction term. This is known as the defect correction technique and was probably first used in the present context by Khosla and Rubin [12]. More specifically, the face value is written as
where stands for the approximation by a lower order scheme, for example, first order upwind, and is the higher-order approximation. The term in brackets is evaluated explicitly using the values from the previous time step, which is indicated by the superscript `o'. It is typically small compared to the implicit part, so that its explicit treatment does not slow down convergence. This approach ensures diagonal dominance for the resulting algebraic equations, thus enhancing iterative rate of convergence while restoring higher-order accuracy at steady-state convergence. The limited anti-diffusive parts of (5.28) may be viewed as deferred corrections to the first order upwind approximation and hence can be put into the source term. Since the stencil associated with first order upwind is maintained, existing codes can easily be modified.