The benchmark problems are presented in the page BENCHMAX.

Authors, affiliation

Authors
    Gary COHEN, Marc DURUFLE (numerical methods and computations)

Affiliation
    Inria Rocquencourt (projet POEMS)

Coordinates
    e-mail   Marc.Durufle@inria.fr
    web   https://www.math.u-bordeaux.fr/~durufle

Method and code

The computations are done on isoparametric quadrilateral/hexahedral elements of Nedelec's first family.

To find the eigenvalues and eigenvectors, ARPACK library is used. As a shift-invert mode (cayley mode more precisely) is used, we have to solve several times a large linear system. We assume in these experiences that all the linear systems are solved directly by MUMPS.

Reference 1 :
Marc DURUFLE,
"Numerical integration and high order finite element methods applied to time-harmonic Maxwell equations",2006 Phd Thesis in University Paris Dauphine.

Code Montjoie :
Marc DURUFLE.

Technical data:
Computations in double precision on a PC/Linux (3.2 Ghz, 2 Go of memory)

Results

Maxwell eigenvalues in 2DomA (L-shape)

The reference eigenvalues are computed by nodal elements on a mesh locally refined near the vertex, with a high order of approximation (Q7). There are 59 459 degrees of freedom. The digits of all eigenvalues don't change, when we use a Q8 approximation.

Reference Eigenvalues

It seems that these reference eigenvalues are equal to Monique Dauge''s eigenvalues, except for the first one (only 1 digit differs).

Let us see the errors when regular meshes are used, and NO local refinement is used.

Relative errors with benchmark.

Now, the computations are done on regular meshes, and the lengths of the intervals follow a geometric progression, in order to have a local refinement near the singular vertex

Relative errors with benchmark.

Now, the computations are done on triangular meshes, split in quadrilaterals. The advantage to use triangular meshes is to be able to locally refine near the vertex (without refining other parts of the domain). The main problem on this technique is the large amount of DOFS generated, because each triangle is split in three quadrilaterals.

Relative errors with benchmark.

Maxwell eigenvalues in 2DomB (Cracked domain)

The reference eigenvalues are computed on a mesh locally refined near the vertex (0,0). Nodal elements are used, with a Neumann condition on the boundaries of the domain. The vertices of the crack are doubled, so that the scalar field can be discontinuous across the crack. For a Q6 approximation, the used mesh contains 127 525 degrees of freedom. All the digits of these eigenvalues don't change when the order of approximation is increased (Q7).

Reference Eigenvalues.

These eigenvalues are equal to those computed by Melina, except the first eigenvalue (the four last digits differ).

Let us see the errors when regular meshes are used, and NO local refinement is used.

Now, the computations are done on the same regular meshes than previous, but the lengths of the intervals follow a geometric progression, in order to have a local refinement near the singular vertex

Relative errors with benchmark.

Now, the computations are done on triangular meshes, split in quadrilaterals, with a local refinement.

Relative errors with benchmark.

Maxwell eigenvalues in 2DomC (Part of a ring)

We obtain exactly the same reference eigenvalues than those computed by M. Dauge. We don't remind them. For the computations, isoparametric curved elements are used. This case is perfect for high order approximation.

Relative errors with benchmark.

Maxwell eigenvalues in 2DomD (curved L-shape)

The reference eigenvalues are computed as usual on a mesh locally refined and with an order of approximation Q6. The used mesh contains 92 845 degrees of freedom. All the digits of these eigenvalues don't change when the order is increased (Q7)

Eigenvalues.

Let us see the errors when regular meshes are used, and NO local refinement is used.

Now, the computations are done on the same regular meshes than previous, but the lengths of the intervals follow a geometric progression, in order to have a local refinement near the singular vertex

Relative errors with benchmark.

Now, the computations are done on triangular meshes, split in quadrilaterals, with a local refinement.

Relative errors with benchmark.

Maxwell eigenvalues for piecewise constant permittivity in 2DomE (Square)

The reference eigenvalues are computed with nodal finite elements on a mesh locally refined with a high order approximation (Q7). Only 11 digits don't change when the order of approximation is increased (Q8). For a Q8 approximation, the mesh contains 629 953 degrees of freedom. The 12th digit can be false.

For this case, we don't provide computations with the first family of Nedelec, because Arpack suffer from problems when generalized eigenvalue problems have to be solved, and when the mesh presents a very high ratio between the largest element and the the smallest element. We didn't observe this problem for the solution of standard real symmetric eigenvalue problems.

Maxwell eigenvalues in 3DomA (Thick L-shape)

We consider the reference eigenvalues given by M. Dauge. Let us see the eigenvalues and the errors when regular meshes are used, with a local refinement. The 3-D mesh is generated by extrusion of a 2-D mesh of the L-shaped geometry. In the direction of extrusion, there is a few layers (by exemple one layer for Q8). We compare different orders of approximation with almost the same number of dofs (about 120 000).

Eigenvalues.

Relative errors with benchmark.

Maxwell eigenvalues in 3DomB (Fichera corner)

We don't have any reference eigenvalues for this case. We are not able to have a good accuracy with our code. Our finest mesh contains 177 720 dofs (for a Q5 approximation). we can estimate the accuracy of the computed eigenvalues as following :

8 first eigenvalues.

The first eigenvalue is simple and the corresponding eigenvector has the most singular part possible (the Fichera exponent is visible at the corner).

The eigenvalues 2 and 3 are double. The eigenvalue 4 has a non-smooth eigenvector and is single.

The eigenvalues 5 and 6 are double and the corresponding eigenvectors are "smooth" .

The eigenvalues 7 and 8 are double. If we consider these eigenvalues as reference, one has the following results

Eigenvalues.

Relative errors with benchmark.