next up previous
Next: Algorithm Up: An Adaptive Finite Previous: Triangle Based Matrices


Finite element representations are accurate and convergent in an integral sense, such as in tex2html_wrap_inline1077 norm. This does not mean, however, that the error at any given point will converge to zero as the mesh is refined. For this reason, the errors in the finite element Laplacian can be a source of noise, unless the mesh is uniform. This is not a problem when one wants to invert the Laplacian, as in eq.(19). This is a smoothing operation and the solution is quite acceptable. On the other hand, calculating the Laplacian, as in eq.(20) can be quite inaccurate locally.

Forming the Laplacian by multiplying by the stiffness matrix and the inverse mass matrix, tex2html_wrap_inline1079 , yields excellent results on a uniform mesh. The Laplacian of a smooth function on this mesh is also smooth. Unfortunately, it is very easy to encounter a non uniform mesh. Fig.13(b) shows


with tex2html_wrap_inline1083 on a mesh with uniformly spaced mesh points, but the triangle vertices have alternately 4 and 8 neighbors. A mesh of this type is used below, and shown in Fig.13(a). Refining the mesh is of no help, if it retains the alternating number of neighbors. This example shows that in some cases, the usual Laplacian is not locally convergent.

A uniform mesh in which all triangle vertices have 6 neighbors can be obtained from this mesh by a simple reconnection operation. One approach is to use vertex reconnection to try to equalize the number of neighbors of each vertex. However, with more complex meshes, it is virtually impossible to do this.

A way to improve the calculation of the Laplacian is to obtain the discretization of tex2html_wrap_inline1085 by applying the gradient operator twice. First, the components of tex2html_wrap_inline1087 are calculated and re expanded in basis functions. Then, the discretized derivatives of these quantities are obtained to get tex2html_wrap_inline1089 which is again expanded in basis functions. The lumped mass is used to make the solution for the derivatives trivial. This method improves the accuracy because the tex2html_wrap_inline1087 components are first averaged from triangles to vertices, smoothing the gradient. The divergence calculation performs a similar smoothing. The method works because taking a single derivative is first order accurate. Taking a second derivative in two steps is formally first order accurate, and the method in practice is found to be convergent, unlike the original method. The discretized version of the gradient is


The effect of applying eq. (36) twice in succession is equivalent to using a new stiffness matrix, S', given by


In the above eqs. (36) and (37), it is sufficient to use the lumped mass matrix (33). This stiffness matrix has the same symmetry properties as the standard stiffness matrix. Fig.13(c) was obtain using the standard lumped mass and tex2html_wrap_inline1095

Although the Laplacian has much better accuracy using this method, there is some loss of numerical stability, because the effective stencil of the Laplacian is now larger.

For adaptive problems, the best method appears to be the use of the current - vorticity formulation. In this approach, the Laplacian does not appear at all on the right hand side of the equations, only on the left hand sides of the two Poisson equations for the potentials. There appears to be no problem with smoothness of the potentials, using the original stiffness matrix, even on an irregular mesh. This because, as remarked above, the inversion of the Poisson equation is a smoothing operation. Both the currents and potentials have sufficient smoothness to obtain acceptable solutions.

next up previous
Next: Algorithm Up: An Adaptive Finite Previous: Triangle Based Matrices

Hank Strauss
Wed Jan 7 14:07:46 EST 1998