Beruflich Dokumente
Kultur Dokumente
DOI 10.1007/s00466-007-0173-y
ORIGINAL PAPER
Received: 21 August 2006 / Accepted: 8 March 2007 / Published online: 3 April 2007
© Springer Verlag 2007
Abstract A simple yet effective modification to the The formulation in this paper is applied to linear elasticity
standard finite element method is presented in this paper. problems and examined for 2D cases, although the concepts
The basic idea is an extension of a partial differential equa- are generally valid.
tion beyond the physical domain of computation up to the
boundaries of an embedding domain, which can easier be Keywords Finite cells · Embedded domain ·
meshed. If this extension is smooth, the extended solution p-extension · Meshless methods
can be well approximated by high order polynomials. This
way, the finite element mesh can be replaced by structured
or unstructured cells embedding the domain where classical
h- or p-Ansatz functions are defined. An adequate scheme for 1 Introduction
numerical integration has to be used to differentiate between
inside and outside the physical domain, very similar to strate- Since the invention of the finite element method in 1950s
gies used in the level set method. In contrast to earlier works, numerous researches have been carried out to justify the
e.g., the extended or the generalized finite element method, mathematical foundations of the method, to develop tech-
no special interpolation function is introduced for enrichment niques in order to increase the accuracy of the results or to
purposes. Nevertheless, when using p-extension, the method decrease the cost of computations, and to make the applica-
shows exponential rate of convergence for smooth prob- tion of the method as user-friendly as possible. However, for
lems and good accuracy even in the presence of singularities. mesh generation as one key step in using the method even
today not all major questions are fully answered. Problems
arise when pure hexahedral meshes are desired, when struc-
tures are defined only implicitly by measurements like X-ray
The first author would like to appreciate the financial support of his tomography, when the domain of computation is strongly
stay in Germany, where this research has been carried out, by the
Alexander von Humboldt foundation.
distorted like in metal forming, when the material is ripped
away or torn apart through processes like explosion or impact,
J. Parvizian (B) or when free and moving boundary problems are
Isfahan University of Technology, Isfahan, 84156 83111, Iran considered.
e-mail: japa@cc.iut.ac.ir
Many attempts have been made over the past decades
J. Parvizian · A. Düster · E. Rank to relieve from the necessity of an exact meshing of the
Technische Universität München, domain of computation. A vast literature is available, where
Lehrstuhl für Bauinformatik, München, Germany very often different denotations are used for similar tech-
e-mail: japa@inf.bv.tum.de
niques. An early suggestion to embed the (a priori unknown)
A. Düster domain of computation into a larger computational mesh
e-mail: duester@bv.tum.de was suggested in [16] for a seepage flow problem, based
E. Rank on an adaptive h-version finite element method. The flow
e-mail: rank@bv.tum.de regime was only defined implicitly by imposing appropriate
123
122 Comput Mech (2007) 41:121–133
integration techniques and transforming the free boundary the boundary. Stiffness matrices for all other cells are com-
problem formally into a formulation with nonlinear material puted using classic FE techniques. This idea of using a “soft”
properties. Instead of a sharp boundary of the seepage regime material in void regions of a domain has been frequently used
a smooth transition from the flow to the non-flow domain in the optimization literature, see for example [1].
was used, very similar to the recently suggested fat bound- Using high order Ansatz functions to approximate the
ary method [10] of Ismail. This technique, being related to extended variables is the most important aspect of the cur-
approaches which are well-known today as level set methods rent work, not mentioned before in the relevant literature to
(see, e.g. [12,20]), was extended in the late 1980s to industrial the best of the authors’ knowledge. A mathematical proof
applications for the simulation of moving boundary problems is sketched, and numerical studies are presented showing
in semiconductor process simulation [13,17,18]. that the method yields exponential rate of convergence for
In the 1990s, a wide variety of papers on meshless meth- smooth problems until the integration error dominates the
ods appeared with the objective to construct approximations discretization error. The error of the method is investigated
entirely in terms of nodes but not node connectivities [3]. in the energy norm and also pointwise, showing that, even
Major, and up to now only partially solved, problems are for “porous” domains with many inclusions, very accurate
optimal distribution patterns of the nodes, possible disconti- solutions can be obtained. The numerical convergence study
nuities of parameters within the domain of computation, and for solid mechanics problems is also extended to non-smooth
proper and efficient implementation of boundary conditions. problems with a detailed look into possibilities of improving
Parallel to efforts to develop meshless methods, exten- accuracy in the neighborhood of singularities.
sion of the domain of computation as well as of the approx-
imation spaces was considered by Duarte et al. [5] and also
by Strouboulis et al. [22]. The essential aspect of similar 2 Finite cell method
versions of the extension, such as eXtended Finite Element
Method (XFEM) or Generalized Finite Element Method 2.1 Basic formulation
(GFEM) reported so far, is the enrichment of the approx-
imation, using a partition of unity method, by the a priori The finite cell method is an extension to the classical finite
knowledge of the solution in special cases, such as singular element method. Therefore, the main concepts and notations
corners. These methods rely on so called library functions of finite elements are used. Without losing generality, this
and adaptive integration schemes. paper concentrates on 2D linear elasticity problems.
Another line of research has been directed toward embed- Let us assume that on a domain Ω with the boundary ∂Ω
ding the domain while the field variables go continuously a problem of linear elasticity is described in weak form by
beyond the physical boundary, [2,4,11,19]. Application of
these methods has been mainly confined to optimal design B(u, v) = F(v) (1)
problems, and low order elements were used in all cases. Sim-
ilar techniques used in domain embedding methods are devel- where the bilinear form is
oped under different versions of fictitious domain methods.
The core idea is again to immerse the original domain into
B(u, v) = [L v]T C [L u] dΩ (2)
a geometrically bigger and simpler one [15]. Finite differ-
ence, finite volume and low order finite element methods have Ω
been used to discretize the fictitious domain, see for example
[4,14,21,25]. To treat the boundary conditions, the litera- in which u is the displacement, v is the test function, L is the
ture provides a wide scope of ideas and techniques similar standard strain–displacement operator and C is the elasticity
in nature but different in names, such as Lagrange multiplier matrix. Without loss of generality we assume homogeneous
method [9] or fat boundary method [10]. Dirichlet boundary conditions along Γ D and a Neumann
boundary Γ N with prescribed tractions, ∂Ω = Γ D Γ N ,
In the present work, to discretize the domain of computa-
tion and the embedding domain, a finite number of compu- and Γ D Γ N = ∅.
tational cells is used, which cover the domain and may have The linear functional
simple shapes like squares or cubes. The field variable, e.g.
the displacement, is smoothly extended beyond the cells cut F(v) = vT f dΩ + vT t N dΓ (3)
by the original boundary. Numerically, the fictitious mate- Ω ΓN
rial is set extremely soft outside the physical domain, thus
guaranteeing that the strain energy of the solutions in the considers the volume loads f and prescribed tractions t N .
original and the extended domains remains the same. Spe- The domain of computation Ω is now embedded in the
cial integration techniques are employed to treat cells cut by domain Ωe with the boundary ∂Ωe , Fig. 1. The interface
123
Comput Mech (2007) 41:121–133 123
+ =
Ω
between Ω and the extension is defined as Γ I = ∂Ω\(∂Ω Note that in the case of a “zero extension”, where C∗ = 0,
∂Ωe ), and it is first assumed that the Dirichlet boundary Γ D the bilinear functional (8) turns to
of Ω does not intersect Γ I . This is, e.g. the case for struc-
tures where the extended domain covers stress free voids in Be (u, v) = [L v]T C [L u] dΩ + [L v]T 0 [L u] dΩ
the interior of Ω. The more general case of extension across Ω Ωe \Ω
Dirichlet boundary conditions is discussed in Sect. 2.2.
Following [11], the displacement variable, u is defined as: = [L v]T αC [L u] dΩ = B(u, v) (10)
Ωe
u1 in Ω
u= (4) in which
u2 in Ωe \ Ω
1.0 in Ω
while the transition conditions guarantee the continuity at the α= (11)
0.0 in Ωe \Ω
interface between Ω and Ωe \ Ω:
The linear functional
t1 = t2 on Γ I
(5)
u1 = u2 on Γ I
Fe (v) = vT f dΩ + vT t N dΓ + vT t̄ dΓ (12)
Boundary conditions are set for Ωe : Ωe ΓN Γ1
123
124 Comput Mech (2007) 41:121–133
To apply boundary conditions, when cell boundaries and Fig. 3 When a cell is cut by the boundary, the integration scheme
the boundary ∂Ω do not conform, several approaches, such considers only the part interior to Ω
123
Comput Mech (2007) 41:121–133 125
uniformly in the cell. Also Gaussian integration with higher for any w in the finite element space of the extended
number of integration points on the original cell or on sub- domain.
cells is applicable. This approach is used in some of the exam-
ples presented later in this paper. Another alternative is to use Using these observations, we can distinguish between the
a quadtree technique subdividing the cut cell to non-uniform following cases:
smaller cells for adaptive integration. The question of the
most efficient integration scheme for cut cells is however – Case 1: There is an analytic extension of uex to Ωe . Obvi-
still open. ously this is true if and only if uex is analytic itself, i.e.
To avoid ill conditioning of the global stiffness matrix, if the original problem has smooth solution up to the
cells completely outside Ω can be ignored for integration boundary ∂Ω.
and assembly, or a small non-zero α can be used for such – h-extension: The convergence rate of the FCM is the
cells. Our numerical experiences show that any α as small same as that of FEM on the extended domain, i.e.
as 10−10 can replace zero. The domain integral in the lin- algebraic and identical to a finite element h-extension
ear functional (12) can be dealt with in the extended domain on Ω, (see Example 3.1.1).
by introducing α and using the same procedure as for the – p-extension: As ue,ex is smooth, convergence is expo-
bilinear form. nential on Ωe and therefore also on Ω, (see
Examples 3.1.2 and 3.2).
– Case 2: uex has singular points at the boundary of Ω
2.4 Convergence properties
which are located at nodes of the finite cells. uex can be ex-
tended to Ωe , yet only with singularities in the interior of
For simplicity, we will only consider the case of a traction
the extended domain. The finite element approximation is
free boundary on Γ I . Typical examples fulfilling this con-
affected by singularities at nodes of its elements. Conver-
dition are structures, where interior voids are covered by
gence rate is algebraic for h- and p-extension and equal
the extended domain (for instance, see examples in the next
to that of a classical finite element approximation on Ω.
section). In this case, is C∗ = 0 in (9).
p-extension will converge at twice the rate of h-extension
By construction, the finite cell approximation uFC on Ω
(see [23,24]) and a preasymtotic exponential rate of con-
to the exact solution uex of (1) is a restriction of the finite ele-
vergence may be observed, (see Example 3.3.2).
ment approximation uFE on the extended domain Ωe to ue,ex
– Case 3: uex has singular points at the boundary of Ω
of (7). Therefore, convergence properties of the finite cell
which are not located at nodes of the finite cells. Now,
method can be derived from those of the associated finite
uFE must approximate a solution with singularities in
element computation. The following first remarks can be
the interior of elements. For h- and p-extension conver-
made:
gence rate is algebraic and lower than in Case 2, (see
Example 3.3.1). Yet, the FCM-solution can be signifi-
– ue,ex is not determined uniquely in Ωe \ Ω, as, due to cantly improved by a strictly local modification of the cell
C∗ = 0, no energy is contributed to the bilinear form (7). pattern, (see Example 3.3.4).
Yet, the energy norm of ue,ex is uniquely determined.
– The restrictions on each exact extended solution to Ω and
the solution to (1) are identical: 3 Numerical examples
123
126 Comput Mech (2007) 41:121–133
formulation. GiD [8] is used for post-processing aims. A number of cells in each direction is chosen as odd so that
fine post processing mesh is used, if necessary. any chance of cells matching the boundary points of the cir-
cular hole is omitted. A simple integration scheme with a
3.1 Perforated plate high Gaussian order, e.g. 20 or more, is used. The conver-
gence rate is then compared with a reference h-finite-element
3.1.1 h-extension for smooth problems extension, using a quasi-uniform, yet unstructured, boundary
matching mesh. Table 1 gives the convergence data.
In the first example, symmetry conditions are applied on As shown in Fig. 5 the rate of convergence and even the
the left and lower boundaries (see Fig. 4), and traction free convergence constants (e.g. the offset of the convergence
boundary conditions are assumed at the void. The reference curves) are nearly the same for FCM and FEM. When sub-
solution is achieved by a finite element p-extension on a mesh tracting all degrees of freedom of cells being completely
of four elements which describe the exact geometry using the inside the void (which need not be assembled to the global
blending function method [23,24]. The interpolation degree, equation system and therefore add no significant cost to the
p, is increased from 1 to 20 yielding an estimate for the strain computation) the convergence curve of FCM would even
energy as 0.7021812127. be shifted down by 20%. On the other hand, the conver-
Different strategies are used to inspect the FCM. First of gence constant of the FEM-h-extension may be improved by
all, h-extension convergence in energy norm is checked by rearranging the elements to minimize the discretization error
refining from a coarse cell distribution, while p = 1. The from distortion of elements. One may conclude however, that
the error introduced by integrating over cells covering both
material and void does not dominate the discretization error
ty = 100 of the h-extension process. In Fig. 5, the slope of the lines
for both solutions is 0.5, as expected.
100
h-extension, uniform refinement
FCM
y
Error in Energy Norm, %
x
10
1
10 100 1000 10000 100000
Degrees of Freedom
Fig. 4 The perforated square plate (4.0×4.0) under symmetric bound- Fig. 5 Convergence of the strain energy for h-refinement of FEM and
ary conditions. The plate is divided to 2 × 2 cells FCM, p=1
123
Comput Mech (2007) 41:121–133 127
1 work software [6]. Since the FCM inherits the fast conver-
gence property of the p-extension, the CPU time is much
shorter than of the linear h-extension. For the current exam-
0.1
ple, a finite element approximation with 57,212 linear ele-
ments result in an error of 0.8% in the energy norm, in 106 s.
0.01 A finite cell solution with only four cells ( p = 8, integra-
tion subcells:10 × 10, Gaussian points: 10 in each direction)
0.001
results in an error of 0.43% in only 7 s. One may notice, more-
10 100 1000 10000 over, that the finite cell method saves also the time required
Degrees of Freedom
for meshing and pre-processing, which is not relevant for the
Fig. 6 A comparison between the FCM- and FEM-p-extension. 2 × 2 given example but which can be considerable for complex
quadrilateral cells are used for the perforated square domains.
123
128 Comput Mech (2007) 41:121–133
1000
Reference
von Mises stress FCM
100
10
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2
x position along the diagonal cutline
1000
Reference
FCM
von Mises stress
100
10
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2
x position along the horizontal cutline
Fig. 8 von Mises stress for the symmetric perforated plate: Refer- Fig. 9 Contours of von Mises stress for the perforated plate, FCM (top)
ence solution and FCM approximation along the diagonal cut-line from with four cells, p = 10, and reference solution (bottom)
(−2, −2) to (2,2) and horizontal cut-line from (−2, 0) to (2,0), top and
bottom, respectively
123
Comput Mech (2007) 41:121–133 129
1000
10
Reference
FCM
1
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2
x position along the horizontal cutline
1000
Reference
FCM
10
1
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2
x position along the diagonal cutline
Fig. 11 von Mises stress along the horizontal (top) and diagonal (bot-
tom) cut-lines for the perforated plate embedded in an extended domain
while cells match the singular corners
Fig. 10 The perforated plate [(−2, −2), (2, 2)] (top) is embedded in
an extended domain [(−3, −4), (3, 2)] covering the Dirichlet boundary. 3.3.1 Cells do not conform to singularities
von Mises contours are given for the domain Ω, only (bottom)
If the cells happen to conform to the edges and singular
corners, the rate of convergence for FCM would be the same
as the rate of a uniform finite element p-extension [23,24].
necessary. The reference solution is obtained by a fine quasi-
Let us instead continue with 11 × 11 cells that do not match
uniform unstructured mesh of 6,106 elements while p is
any singular corner or edges of the problem. Figure 16 shows
increased to 8 to ensure convergence. Figure 14 gives the
(top) this discretized domain and the convergence history of
stress contours and Fig. 15 compares the stress distribution
the energy norm using a FCM-p-extension to an error of
along the diagonal cut-line, which is in very good agreement
about 4% (bottom). The convergence rate is now, as stated
with the reference solution while the maximum error is less
in Sect. 2.4 slower than for a FEM-h-extension.
than 1%.
To show the accuracy of pointwise stress values let us con-
sider a critical cut-line passing through the singular points of
the square corners from (−2, −2) to (2, 2). The stress dis-
3.3 Singularities due to reentrant corners tribution along this cut-line is given in Fig. 17 for p = 8
and an integration with 4 × 4 sub-cells in the cut cells.
The capacity of the method when strong singularities exist This figure suggests that the only region of the problem
in the domain will now be investigated. In this example the that contributes significantly to the error in strain energy,
circular hole of Fig. 4 is replaced by a square inclusion, thus to the slow convergence observed in Fig. 16, is in the
yielding four reentrant corners to the domain of computa- very close vicinity of the singular point. However the finite
tion. A reference solution is obtained from a p-extension on cell method reveals well the singular nature of the problem
a geometrically refined mesh and yields a strain energy of although the cells do not match the singular points or edges
0.8618383790 with 66,334 DOF (p = 16). at all.
123
130 Comput Mech (2007) 41:121–133
1000
Reference
FCM
von Mises stress
100
10
1
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2
x position along the diagonal cutline
Fig. 12 von Mises stress contours for the perforated plate embedded
in an unstructured cell (top), and along the diagonal cut-line (bottom)
123
Comput Mech (2007) 41:121–133 131
1000 10000
Reference Reference
von Mises stress FCM 11*11 Cells, p=8
100
100
10 10
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2
x position of nodes along the diagonal cutline X position of nodes along the diagonal cutline
Fig. 15 Porous domain. von Mises stress distribution by FCM and the Fig. 17 Singular problem. von Mises stress distribution by FCM and
reference solution along the diagonal cut-line from (−2, −2) to (2,2). the reference solution along the diagonal cut-line from (−2, −2) to
Since the results enjoy almost perfect agreement, it is difficult to differ- (2,2)
entiate between them at this scale
uniform h-extension
30 unstructured cells, matching
10
In another attempt, we examined a macro cell pattern enclos-
ing the singular points without necessarily conforming to
them, Fig. 21. This macro is applied to produce a mesh of
57 cells, Fig. 22. The convergence history for two different
1
number of integration sub-cells is given in Fig. 23. As shown,
10 100 1000 10000 100000 exponential preasymptotic convergence is achievable even
Degrees of Freedom
though the nodes and singular corners do not conform.
Fig. 16 The singular perforated plate is discretized using 11 × 11 cells
(top). The error in the energy norm versus DOF (bottom)
4 Conclusions
convergence rate and high accuracy of the solution can be The finite cell method, as described in this paper, is based
observed. on a simple concept which can be implemented with little
123
132 Comput Mech (2007) 41:121–133
Fig. 19 A mesh of 11 × 11 cells is used and the sharp corners are Fig. 22 The macro is used to create a mesh of 57 cells for the singular
replaced by filleted corners of radius r problem
100
r= 0.2 100
r= 0.1 Macro Cell, Subcells: 4*4
r=0.05 Macro Cell, Subcells: 8*8
r=0.02
10
error in energy norm, %
10
1 0.1
0.01
100 1000 10000 100000
DOF
0.1
100 1000 10000 100000
DOF Fig. 23 Preasymptotic exponential convergence rate for different
numbers of integration sub-cells
Fig. 20 The convergence history for various radii
123
Comput Mech (2007) 41:121–133 133
References 14. Ramiére I (2006) Méthodes de domaine fictif pour des problèmes
elliptiques avec conditions aux limites générales en vue de la sim-
1. Allaire G, Jouve F, Toader A (2004) Structural optimization ulation numrique d’écoulements diphasiques. PhD Thesis, UNI-
using sensitivity analysis and a level-set method. J Comp Phys VERSITÉ DE PROVENCE - AIX-MARSEILLE I, France
194/1:363–393 15. Ramiére I, Angot P, Belliard M (2007) A fictitious domain approach
2. Badea L, Daripa P (2001) On a boundary control approach to with spread interface for elliptic problems with general boundary
domain embedding methods. SIAM J Control Optim 40(2):421– conditions. Comput Methods Appl Mech Eng 196(4–6):766–781
449 16. Rank E, Werner H (1986) An adaptive finite element approach
3. Chen Y, Lee J, Eskandarian A (2006) Meshless methods in solid for the free surface seepage problem. Int J Numer Methods Eng
mechanics. Springer, Heidelberg 23:1217–1228
4. Del Pino S, Pironneau O (2003) A fictitious domain based gen- 17. Rank E (1989) Local oxidation of silicon—a finite element
eral PDE solver. In: Kuznetsov Y, Neittanmaki P, Pironneau O (ed) approach. In: Bank RE, Bulirsch R, Merten K (eds) Mathematical
Numerical methods for scientific computing variational problems modelling and simulation of electrical circuits. Birkhuser Verlag,
and applications. CIMNE, Barcelona Berlin
5. Duarte CAM, Babuška I, Oden JT (2000) Generalized finite ele- 18. Rank E (1991) Finite-element-simulation of local oxidation in
ment method for three-dimensional structural mechanics problems. semiconductor processing, In: Whitemann JR (ed) The mathemat-
Comput Struct 77(2):215–232 ics of finite elements and applications VII. Academic, New York
6. Düster A, Bröker H, Heidkamp H, Heißerer U, Kollmannsberger S, 19. Rusten T, Vassilevski PS, Winther R (1998) Domain embedding
Krause R, Muthler A, Niggl A, Nübel V, Rücker M, Scholz D, preconditioners for mixed systems. Numer Linear Algebra Appl
AdhoC 4 -user’s guide, Lehrstuhl für Bauinformatik, Technische 5:321–345
Universität München 20. Sethian JA (1999) Level-Set Methods and Fast Marching Methods.
7. Düster A, Parvizian J, Yang Z, Rank A (2007) The finite cell method Cambridge University Press, Cambridge
for 3D problems of solid mechanics, under preparation 21. Singh KM, Williams JJR (2005) A parallel fictitious domain mul-
8. http://www.gid.cimne.upc.es tigrid preconditioner for the solution of Poisson’s equation in com-
9. Glowinski R, Pan T-W, Hesla TI, Joseph DD (1999) A distributed plex geometries. Comput Methods Appl Mech Eng 194:4845–4860
lagrange multiplier/fictitious domain method for particulate fows. 22. Strouboulis T, Babuška I, Copps K (2000) The design and analysis
Int J Multiph Flow 25:755–794 of the generalized finite element method. Comput Methods Appl
10. Ismail M (2004) The fat boundary method for the numerical resolu- Mech Eng 181:43–69
tion of elliptic problems in perforated domains. Application to 3D 23. Szabó BA, Düster A, Rank E (2004) The p-version of the finite
Fluid Flows. PhD thesis, University Pierre and Marie Curie-Paris element method. In: Stein E, de Borst R, Hughes TJR (eds) Ency-
VI, France clopedia of computational mechanics, Vol. 1, Chapter 5. Wiley,
11. Neittaanmäki, Tiba D (1995) An embedding domains approach in New York, pp. 119–139
free boundary problems and optimal design. SIAM Control Optim 24. Szabó BA, Babuška I (1991) Finite element analysis. Wiley,
33(5):1587–1602 New York
12. Osher S, Fedkiw R, Level-Set Methods and Dynamic Implicit Sur- 25. Zhou YC, Wei GW (2006) On the fictitious-domain and interpo-
faces. Springer, Heidelberg lation formulations of the matched interface and boundary (MIB)
13. Paffrath M, Jacobs W, Klein W, Rank E, Steger K, Weinert U, method. J Comp Phys 219:228–246
Wever U (1993) Concepts and algorithms in process simulation.
Surv Math Ind 3, S.149–183
123