next up previous contents
Next: Approximation methods for convective Up: Space discretization of the Previous: Invariant finite volume discretization

Invariant discretization on non-smooth grids

The discretization of the transport equation (5.1) as explained in the previous section is restricted to more or less smooth grids. Moreover, the diffusion coefficient often varies rapidly, so that has large jumps on cell faces. Hence, straightforward discretization of the diffusivity may results in large errors even on smooth grids. Here we shall consider a discretization which is exact for uniform and linear scalar fields, regardless the smoothness of the grid and of the diffusion coefficient. For this purpose, we need an expression for the partial derivative of any quantity with respect to in terms of general coordinates, namely

 

This can be derived using the chain rule and (2.27). We start with the equation (5.1), viz.,

and with (5.7) taken into account, the appropriate form of the transport equation becomes

 

Discretization of (5.9) is obtained by integration over a finite volume with center (i,j,k). Hence, we have

 

whereas the time derivative and the source terms are integrated using the midpoint rule:

So far, no difficulties arose because none of the quantities used in the above formulae are discontinuous. Furthermore, these discretizations are second order accurate and (5.10) is exact for constant . The cell-face values have to be approximated in terms of values of in neighbouring cell centers by means of interpolation. This will be discussed in the next section. The only point left to discuss is the approximation of the diffusion term. First, for the sake of easier manipulations, the diffusion term will hereafter be expressed as

Integration over yields:

 

The physical quantity is everywhere continuous for arbitrary mappings. Hence, when k is discontinuous, is discontinuous. Thus, approximation of using central differences is inaccurate. However, using the integration-path method an accurate approximation of at point (i+1/2,j,k), for example, is obtained:

 

where

 

where C is given by

 

and

We hereby assume that . The two-dimensional version of (5.15) and (5.16) are given by

with

Substitution of (5.14) in (5.13) gives a discretization which is exact when is constant, and hence exact for linear scalar fields. The other cell-face fluxes can be derived in exactly the similar way.


Tatiana Tijanova
Wed Mar 26 10:36:42 MET 1997