next up previous contents
Next: Time integration algorithms Up: Discretizations and Numerical Algorithms Previous: Discretizations and Numerical Algorithms

Spatial Discretization

In a semi-discrete finite element approach, space is discretized using a finite element method, resulting in a system of ODEs. This technique is illustrated using the following multi-dimensional diffusion initial boundary-value problem (IBVP). The diffusion coefficient is assumed to be a function of the impurity concentration.

The concentration of the impurity is an unknown function of space and time ( tex2html_wrap_inline3282 ) on a domain ( tex2html_wrap_inline3284 ). The strong form (S) of the PDE model is

eqnarray1060

Note that it is not necessary to have a Dirichlet boundary ( tex2html_wrap_inline3286 ) for well-posedness. Also, for notational simplicity, D(C) will usually be given as D, even though it depends on C. For the present, we assume that the domain and boundary conditions are fixed (i.e. tex2html_wrap_inline3294 does not change over time nor do the the boundaries where tex2html_wrap_inline3296 and tex2html_wrap_inline3298 are applied).

This PDE model will be transformed to an equivalent semi-discrete weak form of the problem by multiplying by a spatially varying weighting function and integrating over the domain ( tex2html_wrap_inline3294 ). Two classes of functions will be introduced. The first represents possible solutions (trial functions) for the concentration, and the strong form requires that this function space must be restricted by tex2html_wrap_inline3302 on tex2html_wrap_inline3296 . The second represents allowable functions for the weighting space (test functions), and it is restricted by the homogeneous counterpart of the trial space ( tex2html_wrap_inline3306 on tex2html_wrap_inline3296 ). Continuity constraints in the weak form will add additional restrictions to these function spaces. The transformation from strong to weak form proceeds as follows:

eqnarray1068

Note the use of Einstein notation ( tex2html_wrap_inline3310 ) for taking derivatives with respect to an arbitrary number of spatial dimensions.

The trial and test function spaces are constrained as a result of this weak form. Because of the first order derivatives in the weak form, both the test and trial spaces must be first order continuous (i.e. tex2html_wrap_inline3312 ) over the domain. Thus, the test and trial spaces are

eqnarray1090

The weak form (W) of the problem follows directly. Find tex2html_wrap_inline3314 such that, for all tex2html_wrap_inline3316 ,

equation1094

A mesh is introduced consisting of a set of elements ( tex2html_wrap_inline3318 ) and a set of nodes ( tex2html_wrap_inline3320 ). The mesh is a simple tessellation of the domain (i.e. tex2html_wrap_inline3322 and tex2html_wrap_inline3324 for tex2html_wrap_inline3326 ). The trial and test function spaces are restricted to interpolations on this mesh (i.e. tex2html_wrap_inline3328 and tex2html_wrap_inline3330 ). Note that introducing this interpolation may violate tex2html_wrap_inline3332 , as the interpolation is not exact, but tex2html_wrap_inline3334 is assumed to match tex2html_wrap_inline3336 as closely as possible along the prescribed boundary to minimize the impact of this violation. The Galerkin FEM approximation (G) of the weak form follows. Find tex2html_wrap_inline3338 such that, for all tex2html_wrap_inline3340 ,

equation1108

Choose a basis for the set of interpolation functions on the mesh such that the value of a function in this basis is 1 at the node it is associated with ( tex2html_wrap_inline3342 ) and 0 at all other nodes ( tex2html_wrap_inline3344 for tex2html_wrap_inline3346 ). Because the Galerkin approximation must hold for all tex2html_wrap_inline3340 , let tex2html_wrap_inline3350 consist of all linear combinations of functions in tex2html_wrap_inline3352 . That is,

equation1117

where the tex2html_wrap_inline3354 's are unconstrained except for homogeneity constraint on tex2html_wrap_inline3352 which requires that tex2html_wrap_inline3358 for nodes on the dirichlet boundary ( tex2html_wrap_inline3360 ). The trial solutions will be drawn from the same basis, leading to

equation1120

where tex2html_wrap_inline3362 for nodes on the dirichlet boundary. Finally, note that this set of basis functions is only spatially varying. This leads to the following form for tex2html_wrap_inline3364 .

equation1123

Replacing the appropriate functions in (G) will produce a nonlinear system of ordinary differential equations.

equation1127

Linearity of integration allows the exchange of the integration operator with the summation operator, giving

equation1138

or

equation1149

where

equation1154

Because the tex2html_wrap_inline3354 's are unconstrained (for tex2html_wrap_inline3368 ), each corresponding tex2html_wrap_inline3370 must be 0.

equation1163

for tex2html_wrap_inline3368 .

At this point, the use of Einstein notation becomes cumbersome and confusing so we return to the gradient notation. This equation is

  equation1174

or, with tex2html_wrap_inline3374 and tex2html_wrap_inline3376 , 4.15 can be rewritten as

  eqnarray1187

This is a system of nonlinear ordinary differential equations which can be solved using ODE techniques.


next up previous contents
Next: Time integration algorithms Up: Discretizations and Numerical Algorithms Previous: Discretizations and Numerical Algorithms

Dan Yergeau
Wed Aug 20 14:01:54 PDT 1997