An Embedded Dirichlet Formulation For 3D Continua: A. Gerstenberger and W. A. Wall

INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN ENGINEERING
Int. J. Numer. Meth. Engng 2010; 82:537–563

Published online 23 November 2009 in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/nme.2755
An embedded Dirichlet formulation for 3D continua
A. Gerstenberger and W. A. Wall∗, †

Institute for Computational Mechanics, Technische Universität München,
Boltzmannstr. 15, 85747 Garching, Germany
SUMMARY
This paper presents a new approach for imposing Dirichlet conditions weakly on non-fitting finite element
meshes. Such conditions, also called embedded Dirichlet conditions, are typically, but not exclusively,
encountered when prescribing Dirichlet conditions in the context of the eXtended finite element method
(XFEM). The method’s key idea is the use of an additional stress field as the constraining Lagrange
multiplier function. The resulting mixed/hybrid formulation is applicable to 1D, 2D and 3D problems.
The method does not require stabilization for the Lagrange multiplier unknowns and allows the complete
condensation of these unknowns on the element level. Furthermore, only non-zero diagonal-terms are
present in the tangent stiffness, which allows the straightforward application of state-of-the-art iterative
solvers, like algebraic multigrid (AMG) techniques. Within this paper, the method is applied to the
linear momentum equation of an elastic continuum and to the transient, incompressible Navier–Stokes
equations. Steady and unsteady benchmark computations show excellent agreement with reference values.
The general formulation presented in this paper can also be applied to other continuous field problems.
Copyright q 2009 John Wiley & Sons, Ltd.
Received 19 March 2009; Revised 7 August 2009; Accepted 20 August 2009
KEY WORDS: fluid–structure interaction; extended finite element method; embedded Dirichlet conditions;
fixed grid; domain decomposition; Lagrange multiplier; hybrid/mixed finite elements
1. INTRODUCTION
This paper proposes a new method to weakly embed Dirichlet conditions on non-matching grids. In
the so-called fixed-grid methods, an explicitly or implicitly described interface separates different
regions on a fixed background grid. Hereby, the potentially moving interface usually does not
coincide with element boundaries of elements in the background grid such that prescribing Dirichlet
conditions along the interface cannot be performed by means of normal Dirichlet conditions on
∗ Correspondence to: W. A. Wall, Institute for Computational Mechanics, Technische Universität München,
Boltzmannstr. 15, 85747 Garching, Germany.
†
E-mail: wall@lnm.mw.tum.de, URL: http://www.lnm.mw.tum.de
Contract/grant sponsor: DFG-FG493; contract/grant number: WA 1521/1
Copyright q 2009 John Wiley & Sons, Ltd.

538 A. GERSTENBERGER AND W. A. WALL
nodes. Hence, alternative methods to embed Dirichlet conditions into the background grid are
required. With the increasing interest in fixed-grid methods in general, the development of schemes
to embed such constraints has also gained increasing attention recently.
Our motivation for this research is the goal to develop a 3D fixed-grid fluid–structure interaction
(FSI) scheme [1]. For fixed-grid FSI approaches as proposed in [2, 3], the wet surface of a structure
separates a fluid mesh into the physical fluid domain and a fictitious domain, where no fluid
flow computations are needed. The resulting discontinuities are modelled by the eXtended finite
element method (XFEM) [4–6]. The structural deformation is usually described using a Lagrangean
formulation and the structure surface explicitly defines the interface between fluid and fictitious
domain. The moving interface/discontinuity intersects the fixed fluid elements; hence, application
of FSI coupling condition on nodes of the fixed fluid mesh is not possible and weak formulations
based on Lagrange multipliers have been developed for 2D-FSI problems [2, 3]. Finding appropriate
Lagrange multiplier spaces has been shown to be a challenging and ongoing task also in other 2D
problems, see e.g. [7, 8] and references therein. While extending our 2D-FSI approach [2] to 3D,
it soon became apparent that it is quite hard or even impossible to find an appropriate Lagrange
multiplier space on the embedded fluid–structure interface for arbitrary surface geometries. This
is even more true, if higher-order interface and fluid discretizations are to be used.
As a remedy, methods were proposed, where Lagrange multiplier functions are not defined on
an interface mesh but are embedded into the intersected elements of the background grid. Such
methods include, e.g., the so-called distributed Lagrange multipliers (DLM) as, e.g. in [3, 9], bubble-
stabilized methods [10–12], variations of Nitsche’s method [13–18] or hybrid and discontinuous
Galerkin methods [17, 19, 20]. For a very recent overview on embedding Dirichlet conditions, see
Section 2 in [20]. These methods have the potential to be applied easily in any dimension. All
but the DLM methods have the additional feature that no extra Lagrange multiplier unknowns are
present in the global discrete system.
For our envisioned 3D-FSI approach, we require a method that can be applied to the instationary,
incompressible Navier–Stokes (NS) equations. In particular, it should be applicable to stabilized,
equal-order velocity–pressure-based finite element formulations with piecewise continuous linear,
serendipity and quadratic approximations (hex8, hex20 and hex27). It should further be appli-
cable to any element type like wedge and tetrahedral elements, which can be used to better mesh
complex fluid domains. We also have to keep in mind that fluid problems usually require large
numbers of degrees of freedom (DOFs) and parallel, iterative solvers have to be used. Eventually,
we want to apply algebraic multigrid (AMG) techniques [21, 22] for optimal solution times. In
addition, as the interface moves, the coupling is performed at least every time step. Therefore, the
new coupling approach should be as cheap as possible and also robust. Finally, the method should
be free from user-provided stabilization parameters, if possible.
The method we propose in this paper meets all of our above-listed requirements. The method’s
key idea is the use of an additional stress field to weakly enforce the embedded Dirichlet conditions.
As opposed to the usual Lagrange multiplier, which is an interface vector field, the additional stress
field is introduced as a symmetric tensor field in the domain. Of course, the stress field is only
required within intersected elements. Hence, we choose the hybrid approach using additional stress
unknowns in intersected elements and keep the original, single field approach in uncut elements.
The method does not require stabilization for the Lagrange multiplier/stress unknowns and allows
the complete condensation of stress unknowns on the element level. Furthermore, only non-zero
diagonal-terms are present in the tangent stiffness, which allows the straightforward application
of state-of-the-art iterative solvers, like AMG techniques.
Copyright q 2009 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 82:537–563
DOI: 10.1002/nme
AN EMBEDDED DIRICHLET FORMULATION FOR 3D CONTINUA 539
The idea itself and also several initial 3D-FSI examples were first presented at the Eighth World
Congress on Computational Mechanics in Venice in June/July 2008 [23]. Similar ideas using
hybrid formulations in intersected elements have then also been presented for 2d elliptic problems
in [24].
This paper follows a quite unusual modus operandi. Often, a method is proposed for a compli-
cated problem, but only shown and analyzed for a very much simplified problem. Only when it is
really implemented for the complex case it was intended for, one sometimes finds that it does not
really work for those problems. We believe that doing it the other way around should also be a valid
option for method development. Hence, even though it is the first paper on this approach it applies
the formulation to elasticity (Section 2) and to the complex case—our intended application—
of the incompressible NS equations (Section 3). It also presents details on the algorithm and
implementation, reports 3D flow and FSI computations and verifies it along with 3D benchmark
examples (Section 4). In addition, initial results from a convergence analysis are presented in the
example section. A detailed mathematical analysis of the method will be presented in a separate
paper.
2. EMBEDDED DIRICHLET CONDITIONS FOR LINEAR-ELASTIC PROBLEMS
2.1. Problem definition

In the following, we use the steady, linear momentum equation for a linear-elastic material (small
displacements and small strains) to demonstrate the core features of our proposed method. The
steady-state conservation of momentum in s (with position vector x) is stated as
0 = ∇ ·r in s (1)
Here, r denotes the Cauchy stress tensor. The displacement field is denoted as d. For brevity,
volumetric forces have been omitted in the subsequent derivation.
Assume a Dirichlet boundary condition (DBC) on parts of a boundary of a physical domain
s as depicted in Figure 1(a). This setup can be reformulated as a so-called embedded Dirichlet
problem, where an internal interface i divides a larger domain into a physical and a fictitious
domain named + and − , respectively. The part of the boundary with the DBC becomes now
an internal interface i to the domain as shown in Figure 1(b). This setup is also referred too
as one-sided discontinuity as we have a physical field only on one side of the discontinuity. For
easier writing, we define two additional names for the boundary i , namely + and − , depending
on whether functions are evaluated approaching i from + or − , respectively. Thus, writing d
evaluated at + is equivalent to writing d+ evaluated at i . Note that whether the interface is given
by an analytic function, a level set or a boundary mesh is relevant mainly for the implementation
of the numerical integration and has no effect on the subsequent derivation.
The jump in the displacements [[d]] between the physical values d+ and the void (d− ) can be
expressed as
[[d]] = d+ −
d− in i (2)
=0
DOI: 10.1002/nme
(a) (b)
Figure 1. Figure (a) shows the physical domain with Dirichlet boundary conditions (DBC), while Figure (b)
illustrates how the DBC problem translates to the embedded Dirichlet problem. Physical field s and
interface i along with respective domain normals and variables are shown.
In the case of a void in − , the value of d− is completely irrelevant for the solution in + and
we set it to zero to reduce the complexity of the subsequent derivation. Then the jump height [[d]]
equals the value of d at + and we can pose the kinematic constraints for d+ along i as
i
d+ − d̂ = 0 in i (3)
i
where the interface displacement d̂ is given. Likewise, we can identify a jump discontinuity in
the stress field, which results in a surface traction vector field k as
[[r·n]] = r+ ·n−r −
·n = k in i (4)
=0
In the case of modelling a void, r+ ·n = k can be seen as the surface traction—the ‘reaction
force’—due to the displacement constraint along the interface.
The jump discontinuities can be modelled by the so-called void enrichment [6, 25] using the
XFEM [4, 5]. Here, we can apply the momentum Equation (1) to the entire domain , generate
nodes and elements and then enrich the approximation of intersected elements with the void
enrichment as shown later in Section 2.3. The strong form of momentum equation including the
interface condition at + reads as
−∇ ·r = 0 in + (5)
D
d− d̂ = 0 in D (6)
r·n− ĥ = 0 in N (7)
i
d+ − d̂ = 0 in i (8)
In general, i is not aligned with element boundaries of the fixed grid and Dirichlet conditions
cannot, therefore, be applied directly. Typically, a Lagrange multiplier field k(x) is introduced
along the boundary to weakly enforce the constraint equation (3). The resulting weak form is
i
−(d, ∇ ·r)+ +(d, r·n− ĥ) N −(k, d− d̂ )+ = 0 (9)
DOI: 10.1002/nme
where d and k denote the displacement and the Lagrange multiplier test functions, respectively.
Integration by parts of the stress term yields
−(d, ∇ ·r)+ = −(d, r·n)+ +(∇d, r)+ (10)
Dirichlet and Neumann conditions away from the embedded interface are treated as usual, while
along the interface i , we use Equation (4) to identify the surface traction k
i
(∇d, r)+ −(d, ĥ) N −(d, k)+ −(k, d− d̂ )+ = 0 (11)
The main difficulty for this formulation is to find an appropriate Lagrange multiplier space on the
interface in 2D and—even more difficult—in 3D.
2.2. Embedded Dirichlet formulation using primary stress unknowns

In the following, we propose an approach that removes the need of a surface discretization for
the Lagrange multiplier. The key point is that instead of using a vector traction field k along + ,
we introduce an additional primary stress field r in intersected background elements and use r· n
along + to weakly enforce the interface constraint equation (3). The stress field will be embedded
into the intersected background element as explained in the following.
Assume, we have two primary unknowns, namely the displacement d and the Cauchy stress r
plus their corresponding test functions d and r, respectively. Along the interface, the stress field
will be used to constrain the displacement field as
i
(r·n, d− d̂ )+ (12)
where r·n can be identified as the virtual traction along the interface.
Equation (12) provides only three equations for six unknowns for the symmetric stress tensor
r. Having two primary unknowns in , we add an additional equation to close the set of equations
in by matching the strains computed from both unknowns
e −ed = 0 in (13)
The two strains are defined in terms of the primary unknowns d and r as
ed = 12 (∇d+(∇d)T ) (14)
e = C−1 : r = S : r (15)
The fourth-order tensors C and S represent the material and compliance tensor, respectively. The
new task is then: find d and r such that
−∇ ·rd = 0 in + (16)
e −ed = 0 in + (17)
D
d− d̂ = 0 in D (18)
r ·n− ĥ = 0
d
in N
(19)
i
d+ − d̂ = 0 in i (20)
DOI: 10.1002/nme
The Cauchy stress rd is computed from the displacement unknown via ed using
rd = C : ed (21)
The resulting weak form before integration by parts is
i
−(d, ∇ ·rd )+ −(r, e −ed )+ +(d, rd ·n− ĥ) N −(r·n, d− d̂ )+ = 0 (22)
Integration by parts using Equation (10) and replacing rd with r at i gives

i
(∇d, rd )+ −(r, e −ed )+ −(d, ĥ) N −(d, r·n)+ −(r·n, d− d̂ )+ = 0 (23)
In summary, we shifted the constraining Lagrange multiplier field from the interface to the
background domain, i.e. we replaced k and k by r·n and r·n, respectively. The two primary
fields in + , namely displacement and Cauchy stress, have been weakly coupled by matching the
strain tensors computed from each of the two fields, respectively.
2.3. Spatial discretization

When applying the XFEM, the finite element shape functions are extended or enriched by using
additional DOFs combined with known solution or special enrichment functions. In particular, a
¬
step function (x, t) is used to model the discontinuity between physical domain and void

¬ +1 in +
(x) = (24)
0 in −
As mentioned before, for this paper we constrained ourselves to non-moving interfaces. Hence,
¬
the enrichment function is not time-dependent.
¬
Along i , both displacement and stress field are discontinuous and enriched with (x). The
complete discretization for trial and test functions are given as
¬ ¬
dh (x) = N Id (x)d̃ I + N Jd (x)(x)d J (25)
I J
¬ ¬
dh (x) = N Id (x)d̃ I + N Jd (x)(x)d J (26)
I J
¬
N K (x)r̃ K + N L (x)(x)r L
¬
rh (x) = (27)
K L
¬
N K (x)r̃ K + N L (x)(x)r L
¬
rh (x) = (28)
K L
where the superscript ‘h’ indicates the discretized field function. Values with superscript •˜ represent
¬
the standard unknowns, while values with superscript • indicate additional unknowns due to the
extended approximation.
The displacement shape functions N Id (x) are chosen as piecewise continuous polynomials,
while being C 0 -continuous at interelement boundaries e . The shape functions for the stress
DOI: 10.1002/nme
unknowns N K (x) shall also be polynomial functions inside each element; however, they shall be
C −1 -discontinuous at interelement boundaries e . Hence, stress unknowns are element DOFs and
the discontinuous stress approximations fit the discontinuous strains based on derivatives of the
displacement approximation.
The global system is assembled from element stiffness matrices based on Equation (23) as
N
Kdd Gd d f
A A = A i (29)
e Kd + Gd K e e s e e Gd i d̄
The nodal displacements for this element are denoted as d, element e stress unknowns as s and
discrete forces resulting from Neumann conditions as f N . If the interface is discretized, the known
i
interface displacement is denoted with d̄ . The element matrices K•• correspond to operators
evaluated on the element domain , while all G•• matrices are related to boundary integrals over
e
that part of the interface that crosses the element domain, namely e+ = + ∩e . The element
matrices relate to the weak-form integrals as
Kdd : −(∇d, rd )e (30)

Kd : +(r, ed )e (31)
K : −(r, e )e (32)
and
Gd : −(d, r·n)e+ (33)

Gd : −(r·n, d)e+ (34)
i
Gd i : −(r·n, d̂ )e+ (35)
We can now distinguish between element tangent stiffness matrices for intersected and not
intersected elements. For intersected elements, due to the element-wise discontinuous flux approx-
imation the element flux unknowns can be expressed as
i
s = K−1
(−(Kd + Gd )d + Gd i d̄ ) (36)
This allows the condensation of the stress unknowns at the element level and we are left with only
displacement unknowns as
i
[Kdd − Gd K−1
(Kd + Gd )]e [d]e and [f N − Gd K−1
Gd i d̄ ]e (37)
i
The discretized interface with prescribed discrete interface displacements d̄ result in the term
i
Gd K−1 −1
Gd i d̄ . The term Gd K Gd is clearly symmetric, provided K is symmetric, which
is true for our Bubnov–Galerkin stress ansatz and the linear-elastic material model. However,
Gd K−1
Kd in general is not symmetric. Hence, the global stiffness matrix will have asymmetric
DOI: 10.1002/nme
Figure 2. Void enrichment: dark (red) spheres indicate enriched DOFs on all nodes belonging to
intersected elements, while light (yellow) spheres indicate standard nodal DOFs on all non-enriched
nodes in the physical domain + . Enriched element stress unknowns are applied only to intersected
elements (not shown in this figure).
entries for DOFs attached to intersected elements. Note that it is also possible to construct a fully
symmetric formulation as we will discuss briefly in Section 2.4.
For elements that are not intersected by i , no boundary integrals exist and element tangent
stiffness and right-hand side vector reduce to
N
Kdd 0 f
and (38)
Kd K e 0 e
It follows that the solution of d is independent of the stress solution in such an element and we
are left with
[Kdd ]e and [f N ]e (39)
Hence, we can omit all stress unknowns in uncut elements without effect on the global displacement
solution. Furthermore, as the stress shape functions of an intersected element do not extend into the
neighboring elements, no special treatment (‘blending’) is required for neighbors to an intersected
element.
Figure 2 demonstrates how the DOFs are applied. First, apply enriched nodal DOFs to all nodes
belonging to intersected elements. Then, for the special case of a void enrichment [6], add standard
nodal DOFs to all remaining nodes in the physical domain. As stated before, enriched stress DOFs
are only applied to intersected elements.
For the model problem in this section, we would intuitively choose the stress ansatz function
to be of the same order as the displacement derivatives or higher. Future mathematical analysis
should clarify the optimal choice of ansatz functions. We will specify the used approximation in
Section 3, when we apply this approach to the incompressible NS equations. The related topic of
numerical integration is discussed there as well.
2.4. Relation to the Hellinger–Reissner variational principle and Nitsche’s method

The variational equations have many similarities with hybrid/mixed types of element technology,
for example based on the Hellinger–Reissner (HR) principle. In fact, the underlying idea to this
approach originates from our work on hybrid/mixed element technology. These methods introduce
additional unknowns to improve the stress and/or strain approximations, for example to fight
locking phenomena. In the following, we try to clarify the relation of the proposed weak form to
DOI: 10.1002/nme
the HR variational principle and Nitsche’s method to point out some differences and commonality
between these approaches.
Remember our hybrid/mixed weak-form equation (23), which is given as
i
(∇d, rd )+ −(r, e −ed )+ −(d, ĥ) N −(d, r·n)+ −(r·n, d− d̂ )+ = 0 (40)
Replacing the derived stress rd with the primary stress unknown r only in the first integral, we
get the following variational form:
i
(∇d, r)+ −(r, e −ed )+ −(d, ĥ) N −(d, r·n)+ −(r·n, d− d̂ )+ = 0 (41)
This symmetric weak form corresponds to the variation of the HR functional for linear elasticity.
For specific pairs of functions spaces for displacements and stresses, the HR-formulation is stable
without stabilization.
Nitsches method [18] does not introduce additional primary variables. Writing Nitsche’s method
in a form similar to Equations (40) and (41), we have
i i
(∇d, rd )+ −(d, ĥ) N −(d, rd ·n)+ −(rd ·n, d− d̂ )+ −(d, d− d̂ )+ = 0 (42)
Here, all stress terms are based on nodal displacements and a strain balance equation is not
required. However, the method is reported to be stable only if an additional stabilization term
along the boundary is introduced. For choices of the user-defined stabilization parameter , see
e.g. [13]; for an application of Nitsche’s method to elasticity problems, see e.g. [14].
Comparing these three approaches, all of them result in element stiffness matrices with only
displacement unknowns. However, only the HR-approach and Nitsche’s method result in symmetric
formulations—provided the physical problem itself is symmetric, which, for example, is not the
case for the NS equations. Consequently, as the solution of the NS equations already requires
unsymmetric solvers for the linear system due to convective terms, unsymmetric boundary condi-
tions do not present additional difficulties. Of course, for elastic problems or Poisson equation as
used in heat conduction problems, switching to unsymmetric solvers might not be regarded as a
good choice.
Furthermore, the HR-approach and our mixed/hybrid form do not require a stabilization param-
eter; however, extra computational work for intersected elements is required to condense the stress
unknowns in intersected elements. Without an in-depth mathematical analysis, we can only specu-
late on the stability properties of our proposed approach. Here, as well as for the HR-formulation,
it seems that a proper pair of function spaces leads to a stable formulation, whereas Nitsche’s
formulation is reported to be stable only with proper choices of the stabilization parameter . For
the approximation spaces as given in the next section, we have not observed any stability problems
in our simulations, which also included preliminary tests with the Poisson equation.
Finally, Nitsches method and our approach leave the virtual elastic energy integral untouched
such that the boundary constraint could potentially be added to any advanced existing element
formulation. When choosing an appropriate method for applying weak embedded Dirichlet condi-
tions to the incompressible NS equation, Nitsches method and our approach will likely keep the
fine-tuned equal-order velocity–pressure formulation intact, whereas an equivalent to HR would
require at least a partial rewrite and modification of the element formulation and implementation.
Keeping our existing stabilized equal-order approximation for the incompressible NS equations
without introducing an additional stabilization parameter was the primary motivation to develop
DOI: 10.1002/nme
the mixed/hybrid formulation for embedded Dirichlet conditions. Its application is presented in
the next section.
3. EMBEDDED DIRICHLET FORMULATION FOR THE INCOMPRESSIBLE

NS EQUATIONS
The following section illustrates how the proposed embedded Dirichlet formulation can be applied
to the incompressible NS equations. We consider the fully transient form; however, for this paper,
the interface with the embedded Dirichlet conditions is assumed not to move.
The additional challenge compared with the previous section is the additional pressure unknown
due to the incompressibility constraint. For additional information on applying the XFEM to
incompressible materials, see also [26].
3.1. Definition of the fluid problem

As in the previous section, we want to apply a Dirichlet condition to parts of a fluid boundary
as depicted in Figure 3(a). Again, this setup is reformulated as an embedded Dirichlet problem
with a larger domain consisting of a physical and a fictitious fluid domain named + and − ,
respectively. The fictitious domain − essentially represents a void in the fluid continuum, which
will be modelled via the XFEM during the discretization. The part of the boundary with the
Dirichlet interface condition becomes an internal/embedded interface i to the domain as shown
in Figure 3(b).
The density-scaled, transient, incompressible NS equations are given as
*u
= −u·∇u+∇ ·r (43)
*t
∇ ·u = 0 (44)
Here, u is the fluid velocity and r the Cauchy stress tensor. For incompressible flow, the Cauchy
stress is split into the pressure p and the deviatoric stress s as
r = − pI+s (45)
(a) (b)
Figure 3. Figure (a) shows the fluid domain with Dirichlet boundary conditions (DBC), while Figure (b)
illustrates how DBC problems translate to the embedded Dirichlet problem. The fluid field f and interface
i along with respective domain normals and variables are shown.
DOI: 10.1002/nme
We use the Newtonian material law—with a constant kinematic viscosity —, which defines the
viscous stress as
s = 2e = (∇u+(∇u)T ) (46)
External, velocity-independent volumetric forces b are omitted in the subsequent derivation for
brevity; however, they are included in the actual implementation and their treatment poses no
additional difficulties. The momentum equation was scaled by the fluid density f such that the
Cauchy stress must be multiplied by the density to get the physical stress values.
The velocity conditions along the interface are embedded Dirichlet boundary conditions only
for + such that
u+ = ûi in i (47)
For transient problems, the initial conditions are given as
u(x, t = 0) = u0 in + (48)
+
p(x, t = 0) = p 0
in (49)
−
u(x, t = 0) = 0 in (50)
−
p(x, t = 0) = 0 in (51)
+
Body forces, if present, are only applied to .
In the above form of the NS equations, we have to find a solution for velocity and pressure
fields. To apply the method for embedded Dirichlet conditions sketched in Section 2, we add
the Cauchy stress r as independent variable and add an extra strain rate balance equation. The
resulting task is therefore: find u, p and r such that
*u
+u·∇u−∇ ·(− pI+su ) = 0 in + (52)
*t
∇ ·u = 0 in + (53)
e −eu = 0 in + (54)
u− ûD = 0 in D (55)
ru ·n− ĥ = 0 in N (56)
u+ − ûi = 0 in i (57)
The two strain rates eu and e and the stress tensors ru and su are computed from the primary
unknowns
eu = 12 (∇u+(∇u)T ) (58)
1
e = (r+ pI) (59)
2
DOI: 10.1002/nme
ru = − pI+su (60)
su = 2eu (61)
We want to stress again the fact that we do not replace the viscous stress and pressure in the
momentum equation (52). Our approach is in contrast to 3-field methods as, e.g. in [27], where
an additional stress field is used to improve the stress approximation and to implement advanced
material models. Our additional stress approximation enters the momentum equation via boundary
integrals resulting from partial integration of the weak form, as shown in the following sections.
3.2. Discretization in time

For simplicity of presentation, we describe the one-step- method for time discretization. Here,
the acceleration is defined as a combination of the velocity time derivative at the new and the old
time step
n+1 n
un+1 −un *u *u
= +(1−) (62)
t *t *t
Now *un+1 /*t is replaced by the right-hand side of Equation (43) evaluated at the new time step
n +1

n+1
1
u n+1
+t u·∇u− f ∇ ·(− pI+s ) u
= un +t (1−)u̇n in + (63)

To complete the time-discrete strong form for the fluid flow in + and − , we also require
∇ ·un+1 = 0 in + (64)
From now on, we omit the superscript n +1. No superscript simply means quantities at the new
time step.
The time-discrete strong form is, therefore, given by
u+t[u·∇u−∇ ·(− pI+su )]−un −t (1−)u̇n = 0 in + (65)
e −eu = 0 in + (66)
∇ ·u = 0 in + (67)
u− ûD = 0 in D (68)
ru ·n− ĥ = 0 in N (69)
u+ − ûi = 0 in i (70)
3.3. Weak form of the time-discrete NS equations

The weak form is developed by testing Equations (65)–(70) with the velocity, pressure and stress
test functions v, q and c, respectively. Dirichlet and Neumann conditions on all boundaries but the
interface are employed as usual. Along i , the introduced stress field is used to weakly enforce
DOI: 10.1002/nme
Equation (70). The complete weak form is given as

(v, u)+ +t (v, u·∇u)+ +(∇v, − pI+su )+ +(q, ∇ ·u)+

1
− c, (r+ pI)−eu −(c·n , u)+ −(v, r·n )+
f f
2 +
= (v, un +t (1−)u̇n )+ +t(v, ĥ) N −t(c·nf , ûi )+ (71)
It should be clear from Equation (71) that the coupling condition at i is weakly enforced at t n+1 .
For completeness, the stationary form is also given as
(v, u·∇u)+ +(∇v, − pI+su )+ +(q, ∇ ·u)+

1
− c, (r+ pI)−eu −(c·nf , u)+ −(v, r·nf )+
2 +
= (v, ĥ) N −(c·nf , ûi )+ (72)
In this work we use linear and quadratic equal-order shape functions for velocity and pressure
space discretization. Such a formulation is known to show instabilities for two reasons—dominating
convection and (inf–sup) unstable pairs of velocity and pressure shape functions. The employed
stabilization [28, 29] is characterized by a modification of the discrete test functions vh and q h
given as
vmod,h = vh + M [f uh ·∇vh +∇q h ] (73)

q mod,h = q h + C [∇ ·vh ] (74)
The stabilization parameters M and C for each element K can, e.g., be given as

2 f
h K m K h |uh | p=2 h K
M = min t, , K
and C
= (Re) (75)
2|uh | p=2 4
2
with

Re K , 0Re K <1 m K |uh | p=2 h K f
(Re) = and Re K = (76)
1, Re K 1 2
For tri-linear shape functions we use m K = 13 , for tri-quadratic shape functions m K = 12

1
. The neces-
sary element size h K is evaluated at each integration point using the approximate streamlength.
During all our simulations, we found no need to stabilize the strain balance equation (66). Hence,
no additional user-defined parameter is needed. Future mathematical analysis should completely
clarify this issue.
DOI: 10.1002/nme
3.4. Fluid discretization in space

Along + , velocity, pressure and stress fields are discontinuous within the element and enriched
¬
with (x) from Equation (24). The complete discretization for trial and test functions with poly-
nomial shape functions for velocity, pressure and stress fields is given as
¬ ¬
uh (x, t) = N Iu (x)ũ I (t)+ N Ju (x)(x)u J (t) (77)
I J
¬ ¬
vh (x) = N Iu (x)ṽ I + N Ju (x)(x)m J (78)
I J
p p ¬ ¬
p h (x, t) = N I (x) p̃ I (t)+ N J (x)(x) p J (t) (79)
I J
p p ¬ ¬
q h (x) = N I (x)q̃ I + N J (x)(x)q J (80)
I J
¬
N K (x)r̃ K (t)+ N L (x)(x)r L (t)
¬
rh (x, t) = (81)
K L
¬
N K (x)c̃ K + N L (x)(x) L
¬
ch (x) = (82)
K L
For the stabilized elements as used in the examples, velocity and pressure have the same order
for their ansatz function and are continuous over element boundaries. For the stress, we use the
same shape functions as for the pressure unknowns; however, they are discontinuous over element
boundaries. Hence, velocity and pressure unknowns are located at nodes, while the stress unknowns
are element unknowns. Adopting the notation from [30], the approximation for velocity, pressure
(20) (20) (20)
and stress could be termed as Q1 Q1 Q−1 for tri-linear, Q2 Q2 Q−2 for quadratic serendipity
and Q2 Q2 Q−2 for full tri-quadratic shape functions. Likewise, we choose the approximation for
linear and quadratic tetrahedrons as P1 P1 P−1 and P2 P2 P−2 , respectively. Other element shapes
like pyramids and wedges should approximate the stress in an analogous manner.
3.5. Solution of the fluid equations

For intersected elements, the element stiffness matrix including the additional stress unknowns
becomes
⎡ ⎤ ⎡ ⎤ ⎡ rhs ⎤
--------------
Kuu Kup Gu u f
⎢ K pu ⎥ ⎣ p⎦ = ⎣ 0 ⎥
⎢
e
A⎣ K pp 0 ⎦
- - - - - - - - - - - - - - ------ - - - e
A e -----
---- ⎦ A
(83)
Ku + Gu K p K e s e Gu i ūi e
The submatrices Kuu , Kup , K pu and K pp denote the standard non-linear operators arising in
stabilized FE formulations of the NS equations. Velocity and pressure nodal unknowns are denoted
as u and p, while s denotes the element stress unknowns. The vector f rhs results from Neumann
DOI: 10.1002/nme
conditions and transient terms of the old time step. The remaining submatrices result from boundary
integrals
Gu : −(v, r·nf )+ (84)
Gu : −(c·nf , u)+ (85)
Gu i : −(c·nf , ûi )+ (86)

and domain integrals
Ku : +(c, eu )+ (87)

1
K p : − c, pI (88)
2 +

1
K : − c, r (89)
2 +
as given in Equation (71). Again, we can condense the stress unknowns at the element level
s = K−1 i
(−(Ku + Gu )u − K p p + Gu i ū )|e (90)
With G∗u = Gu K−1

we get

Kuu − G∗u (Ku + Gu ) Kup − G∗u K p u f rhs − G∗u Gu i ūi
Ae Ae = Ae (91)
K pu K pp e
p e 0 e
For uncut elements, no boundary integrals G•• are present, i.e. the stress equation decouples
from velocity and pressure equations such that the element stiffness matrix reduces to
⎡ ⎤
-------------
Kuu Kup 0
⎢ ⎥
⎣ K pu K pp 0 ⎦ (92)
-------------------
Ku K p K e
Consequently, the element stresses can be safely omitted in uncut elements and we integrate and
assemble just the normal stiffness matrix with only velocity and pressure unknowns.
3.6. Numerical integration of domain and surface integrals

All uncut elements are integrated as usual. Elements with discontinuities are subdivided into
subdomains using a constrained Delaunay tetrahedralization algorithm. Domain integrals are eval-
uated over these subdomains [5]. During subtetrahedralization, the interface i is triangulated into
boundary integration cells to perform exact integration of surface integrals. This subdivision process
allows to integrate the approximation functions exactly, but requires higher programming effort
than reduced or approximate integration schemes. The number of integration points for domain
integrals is not increased by the stress approximation, as the stress field has the same approximation
order as velocity and pressure. For testing and verification, we generated and applied various low-
and high-order Gauss rules following [31, 32] to verify that sufficient integration points are used.
DOI: 10.1002/nme
For geometric aspects and the implementation of proper interface localization algorithms, search
trees and creation of linear or higher-order integration cells (tetrahedral cells in the domain and
triangles at the interface) in 3D, see [33].
4. NUMERICAL EXAMPLES
In the following section, simulation results of stationary and instationary fluid flow around
embedded discretized structures are shown. The intention is to demonstrate the flexibility of the
method with respect to structural shapes and to verify or validate the method.
4.1. Shear flow without pressure gradient

The first example shall verify correct viscous stress computations and that a correct viscous stress
solution along the interface leads to correct interface forces. For that purpose, we model a shear
flow with a linear velocity distribution in the y-direction and consequently, a constant viscous stress
xy component. The physical fluid domain has the dimensions 1.0×0.5×0.1 m and is modelled by
a larger computational fluid domain, which is intersected by an embedded plane wall at y = 0.5 m
as shown in Figure 4(a).
Y
Z X
(a) (b)
(c)
Figure 4. Shear flow computation with linear velocity distribution and constant viscous stress:
(a) setup: 9×9×1 hex8 fluid elements are intersected by 2×2 quad4 interface elements; (b)
velocity field with linear distribution of u x and u y = u z = 0 m/s; and (c) nodal forces at the four
quad4 interface/wall elements. Integration is performed over the intersection area (+ ) using
triangular boundary integration cells.
DOI: 10.1002/nme
The fluid density was chosen to be f = 1.0 kg/m3 , the kinematic viscosity is given as
f = 10−3 m2 /s. At the top, the velocity is prescribed as u = (1.0, 0.0, 0.0)T m/s, while on the left
and right sides, the velocity component in y-direction is set to zero to achieve a horizontal inflow.
In addition, no flow in z-direction is allowed at all nodes. The resulting velocity solution is shown
in Figure 4(b).
The chosen setup results in a constant shear rate and viscous stress in the fluid domain. In
particular, all stress components but xy (and, obviously, also yx ) are zero. The pressure and its
gradient are zero as well. The analytic value of xy is

*u x *u y
xy =
f
+ = (2.0 s−1 +0.0 s−1 ) = 0.002 N/m2 (93)
*y *x
The computed values for xy in the intersected elements match the analytic value above up to
numerical precision. Note that we scaled the computed stress field with the density, which for
the given density f changes the unit, but leaves the values of analytic and computed stresses
comparable.
With our envisioned FSI approach in mind, the structure surface is given as a discrete mesh
consisting of four quad4 elements. Lets assume that the interface velocity ui and its test function
i
vi are discretized using the shape function N Ku . Then one can evaluate the integral (vi , r·nf )+ to
get nodal forces for each node K at the interface mesh as

i
fK =
i
N Ku r·n dx (94)
+
The computed nodal forces and the boundary integration cells used for integration are shown in
Figure 4(c). The analytic value for the force in y-direction on the ‘wet’ intersection area Ai = 0.1 m2
is Fy = yx Ai = 0.2·10−3 N. Adding all y-components of the computed nodal forces of the interface
mesh (Figure 4(c)), we get exactly the same value (up to numerical precision).
4.2. Hydrostatic pressure

The next example shall verify a proper pressure solution along the interface and correct interface
forces due to the pressure acting on the wall. For this purpose, again a simple setup was chosen,
where a bodyforce b = f (1.0, 0.0, 0.0)T m/s2 leads to a hydrostatic pressure field in a channel as
depicted in Figure 5(a). Velocity conditions were chosen such that all velocities remain zero and
therefore, no viscous stresses are present in the solution.
The computational fluid domain of the channel has the dimensions 2.0×0.2×0.1 m and is
modelled with 19×2×1 hex8 fluid elements. The channel is intersected at length L W = 1.75 m
with the bigger part being the physical fluid domain f . The intersecting wall is discretized with
two quad4 elements. The fluid density was again chosen to be f = 1.0 kg/m3 . With the given
bodyforce b and zero pressure at the left-hand channel ending, the analytic pressure at the interface
is p = f bx L W = 1.75 N/m2 . The setup and the pressure solution are given in Figure 5(a). The
solution of the element stress component xx in the intersected element is plotted in Figure 5(b),
the resulting interface force in Figure 5(c).
From the absence of viscous forces it follows that
p(x) = −xx (x) = −yy (x) = −zz (x) in + (95)
DOI: 10.1002/nme
(a)
(b) (c)
Figure 5. Hydrostatic pressure distribution and resulting forces on a wall: (a) pressure solution;
(b) discontinuous element stress component xx in the intersected element; and (c) nodal forces at wall.
To match the linear pressure distribution in this example exactly, at least a tri-linear stress ansatz
is required. If one would use element-wise constant ansatz functions for the stress field, it would
be impossible for xx to be 1.75 N/m2 along the interface, since it must also be equal—in a weak
sense—to the linear pressure distribution. Consequently, the interface forces could not be exact
even for this simple example. The inaccuracy would vanish when the element size h goes to zero,
however, slow spatial convergence can be expected and was observed numerically. For the result
shown in Figure 5(a), we use the same stress shape functions (albeit discontinuous over the element
boundaries) as for the velocity and pressure fields. Hence, we are able to represent the linear stress
components exactly.
From this and the previous example we concluded that the stress ansatz should always be
chosen to be of the same or higher polynomial order as the pressure ansatz. At the same time,
the stress ansatz should have the same or higher polynomial order as the velocity derivatives
used for the viscous stress computation. In our implementation with equal order velocity–pressure
approximations, the pressure dictates the minimal polynomial order for the stress approximation.
For other elements like the Taylor–Hood element the velocity derivatives might dictate the minimal
stress approximation. We plan some strict mathematical analysis in the future to provide more
understanding into this ‘rule of thumb’.
4.3. Convergence analysis using the Jeffery–Hamel flow

In the following we use the so-called Jeffery–Hamel flow [34–36] to study the convergence behavior
of the proposed weak Dirichlet constraints. It is one of the few non-trivial analytic solutions to
the steady NS that features convective behavior as opposed, for example, to simple channel flow
with parallel walls.
DOI: 10.1002/nme
(a) (b)
Figure 6. Convergence study using the 2D Jeffery–Hamel flow: (a) setup and (b) convergence rates.
The Jeffery–Hamel flow is a flow between two converging walls, for which a 2D steady-state
solution with only radial velocity components can be established. However, the solution for the
radial velocity u r () is given only implicitly as solution of a non-linear ordinary differential
equation (ODE). The problem can be expressed by ODEs with varying order and corresponding
boundary conditions, which also influence how analytic solutions can be computed. We follow the
discussion on solution strategies in [37] and use the following third-order ODE:

u r +4u r +2u r u r = 0, 0 =
4
The particular setup for the convergence analysis, which has also been used by others, e.g. in
[38], is depicted in Figure 6(a). It requires the following boundary conditions:

u r (0) = 0, u r = −170, u r =0
8 8
The solution is symmetric with respect to = /8 and the maximum velocity at the symmetry
line at radius r = 1 is set to −170. With the kinematic viscosity set to = 1, the Reynolds number
computed by

3
Re = − u r () d
4 0
as defined in [37] is equal to Re ≈ 85. Velocity and pressure solutions on the second refinement
level are depicted Figure 7.
All Dirichlet conditions are applied using the new mixed/hybrid embedded DBC approach. As
all boundaries prescribe velocities, it is necessary to prescribe the pressure at an arbitrary node
DOI: 10.1002/nme
Figure 7. Convergence study using the 2D Jeffery–Hamel flow: example solution on the second refinement
level along with the computational grid: (a) velocity solution and (b) pressure solution.
within the fluid domain. The actual computation is performed using 3D hexahedral elements; the
corresponding convergence rates for the velocity field computed by

ε L 2 = uh −uexact L 2 = |uh −uexact |2 dx (96)
+
are shown in Figure 6(b). It can be seen that optimal convergence rates for all tested element types
have been achieved. Similarly, also for other, simpler test examples like linear–elliptic problems,
optimal convergence rates have been obtained but for brevity of the presentation, these results are
not included here.
4.4. Benchmark computations

For more realistic validation, we use the stationary 2D case 2D-1, the instationary 2D case
2D-3 and the 3D benchmark cases 3D-1Q and 3D-1Z from [39], where incompressible, stationary,
laminar flows (Re = 20–100) over a cylinder and a cuboid are computed. The setup, boundary
conditions and results are taken without modifications from the original publication [39].
DOI: 10.1002/nme
For our fixed-grid approach, the cylinder/cuboid is discretized using a fixed number of 8-node
hexahedral elements, which results in an interface mesh of quad4 elements. For each benchmark,
the number of structural elements is kept constant for all refinement levels.
4.4.1. 2D benchmarks. The 2D benchmarks—calculated as quasi 2D problems with our present

3D implementation—were used to demonstrate that the developed method converges to the correct
solution for linear and quadratic element ansatz functions. Furthermore, we demonstrate that
element-wise condensation of stress unknowns leads—obviously—to less unknowns, while the
same results are achieved. In fact, when comparing the condition number of the condensed and
uncondensed global tangent stiffness matrices, condensation always led to a reduced condition
number (at least for the results presented in this paper).
The benchmark 2D-1 consists of a 2D-channel with no-slip conditions at the upper and lower
walls, a parabolic inflow with a maximum velocity of 0.3 m/s and a do-nothing condition at the
outflow. The 2D-setup is modelled with one-layer of hexahedral elements in the third dimension,
while refinement of the fluid mesh is performed only in the first two dimensions. To make the
number of unknowns comparable to the reference computations, our number of unknowns have to
8
be multiplied by the factor 20 , as we used 3D (hex20) elements instead of 2D (quad8) elements.
4
Likewise a factor of 8 is needed to compare the number of DOFs for the hex8 element with
linear 2D elements in [39]. For illustration, an example solution is given in Figure 8. The results
of the 2D computations are given in Table I.
For all element types, we see a very good agreement with the given reference values. The neces-
sary number of unknowns to achieve this accuracy is within the range of reference computations.
The main conclusion drawn from these results is that condensation of stress unknowns does not
change results as shown for the hex20 element. Hence, we did not show the condensed results
for the hex8 element—they are identical to the results for the hex8 elements with uncondensed
stresses. The computational cost of condensation is negligible when compared with numerical
integration; hence, we conclude that condensation can and should always be performed.
The setup for benchmark 2D-3 is almost identical to case 2D-1. The only difference is a dynamic
inflow condition, where the parabolic inflow over the height (y-direction) is given as
U (x = 0, y, t) = 4Um y(H − y) sin( t/8 s)/H 2 , V =0 (97)
Figure 8. Benchmark case 2D-1: velocity field. Shown is the coarsest mesh with hex8 elements.
DOI: 10.1002/nme
Table I. Results for the stationary 2D benchmark case 2D-1.
Fluid element type NuDOF

,p NrDOF cD cL p
2D-1 hex8, r not condensed 12 040 2832 5.6066 0.0093 0.1147
40 416 5712 5.5893 0.0093 0.1149
146 328 11 376 5.5814 0.0105 0.1156
hex20, r not condensed 5432 1920 5.5309 0.1047 0.1162
14 884 3840 5.5739 0.0063 0.1169
48 020 7440 5.5766 0.0108 0.1159
166 388 14 880 5.5753 0.0107 0.1175
hex20, r condensed 5432 0 5.5309 0.1047 0.1162
14 884 0 5.5739 0.0063 0.1169
48 020 0 5.5766 0.0108 0.1159
166 388 0 5.5753 0.0107 0.1175
Ref. lower bound — — 5.5700 0.0104 0.1172
Ref. upper bound — — 5.5900 0.0110 0.1176
Table II. Results for the instationary 2D benchmark case 2D-3.

,p t cD,max cL,max p
2D-3 hex20, r condensed 3444 0.008 3.1828 0.3944 −0.1078
10 488 0.008 2.7824 0.1308 −0.1141
30 504 0.008 2.9273 0.4861 −0.1077
102 504 0.008 2.9470 0.4733 −0.1093
Ref. lower bound — — 2.9300 0.4700 −0.1150
Ref. upper bound — — 2.9700 0.4900 −0.1050
with H = 0.41 m and Um = 1.5 m/s. The computation has to be performed for 8 s. Measures are
the maximal drag and lift values as well as the pressure difference between front and rear ends
of the cylinder at t = 8 s. For time discretization, we applied the one-step- ( = 0.66) and the
second-order backward differencing (BDF2) scheme.
The results of the instationary 2D computations are given in Table II. Note that only the BDF2
results are given as they were more accurate as expected from a second-order scheme. As expected
after evaluating the stationary simulations, we achieve very good agreement with the reference
results also for the instationary case.
4.4.2. 3D benchmark. The 3D benchmarks 3D-1Q and 3D-1Z have almost identical fluid setups:
the fluid domain is 0.41×0.41×2.5 with ‘no-slip’ conditions on the upper, lower and side walls.
There is a 3D parabolic inflow applied and ‘do nothing’ conditions at the outflow. The two cases
differ in the shape of the immersed structure—a cuboid with dimensions 0.1×0.1×0.41 in case
3D-1Q and a cylinder with radius 0.05 in case 3D-1Z. For illustration, example solutions for both
cases are given in Figure 9. The results of both 3D computations are given in Table III.
When comparing these results with the reference values, drag cD , lift cL and pressure difference
between front and rear ends of the structure p match the reference values very well. The remaining
small disagreements are very likely contributed to the reduced spatial resolution (due to limited
computational resources) when compared with the reference computations in [39]. Nevertheless,
DOI: 10.1002/nme
Figure 9. 3D benchmark cases 3D-1Q and 3D-1Z: (a) pressure field for case 3D-1Q in the cross-section
and (b) velocity field for case 3D-1Z in the cross-section.
Table III. Results for 3D benchmark cases 3D-1Q and 3D-1Z.

,p cD cL p
3D-1Q hex20, r condensed 27 936 7.5465 0.1926 0.1761
88 036 7.5408 0.0759 0.1736
Ref. lower bound — 7.5000 0.0600 0.1720
Ref. upper bound — 7.7000 0.0800 0.1800
3D-1Z tet4, r condensed 66 268 6.1534 0.0356 0.1645
240 708 6.1533 0.0115 0.1744
tet10, r condensed 69 724 6.1979 0.0506 0.1703
246 780 6.1619 0.0064 0.1751
hex20, r condensed 28 112 6.2093 0.1238 0.1645
89 156 6.1572 0.0050 0.1664
hex27, r condensed 74 936 6.2026 0.0736 0.1659
169 592 6.1567 0.0067 0.1663
Ref. lower bound — 6.0500 0.0080 0.1650
Ref. upper bound — 6.2500 0.0100 0.1750
for the given number of unknowns, similar accuracy is achieved when compared with reference
results computed with similar numbers of unknowns. Case 3D-1Q demonstrates that the method
can deal with sharp corners of the interface constraint, even though a pressure singularity exists at
the two front edges of the cuboid. Case 3D-1Z was used to test the new formulation on some of
the more common element shapes. For all element shapes, the solutions tend towards the reference
values.
It is interesting to note that the fluid discretization poses no constraints on how to discretize
the interface mesh. The surface elements of the cubical structure are 2–4 times bigger than the
intersected fluid elements. For the cylinder, these surface elements are usually smaller. The fine
structure resolution is only needed in circumferential direction to approximate the circular shape.
The resolution in axial direction had no influence on the lift and drag values, which are sums
over all nodal forces of the interface mesh. Of course, if one is interested in the force distribution
on the surface as during FSI computations with flexible structures, then a finer interface mesh is
DOI: 10.1002/nme
Figure 10. Stationary flow field through a channel with a flexible structure. The parabolic inflow (right)
and the wall boundaries are standard Dirichlet boundary conditions. The outflow boundary (left) is of
Neumann (zero traction) type. The structure is fixed at the lower wall. The fluid–structure interface is
modelled using the proposed embedded Dirichlet approach.
required. In combination with the example in Sections 4.1 and 4.2, the results indicate that we
have correct forces and no leakage for any ratio of surface and fluid element sizes.
4.5. Channel flow over a deforming thick plate

As an outlook, we show preliminary results of a steady-state 3D FSI simulation, as, among
others, FSI simulations are our main motivation for our research in fixed-grid methods with
implicit boundaries. In this example, a flexible structure (Poison ratio = 0.48, Young’s modulus
E = 90 N/m) is deformed due to a flow through a channel with dimensions 0.5×1.0×3.0 m. The
fully coupled FSI simulation is performed until steady state is reached. Top and bottom channel
walls have zero (no-slip) velocity prescribed, the inflow from the left is prescribed by a parabolic
velocity condition. The Reynolds number based on the mean inflow velocity and the channel height
is 16. The stationary equilibrium solution is shown in Figure 10. A detailed description of the 3D
FSI approach will be the topic of an upcoming paper.
5. CONCLUSION
In this paper, we presented an approach to weakly enforce Dirichlet conditions along discontinuities
modelled by the XFEM. The method promises to be applicable to a wide range of problem classes.
Our ultimate goal is to provide an accurate, robust and simple to implement method that works
reliably in complex FSI computations involving large and very large deformations of structures.
The extra complication is that we require the method to be applicable to the incompressible NS
with its incompressibility constraint and the extra pressure unknown.
This challenge is met by a new weak formulation, where additional element stress unknowns
play the role of the Lagrange multiplier function and are well defined within the fixed domain
and along the interface. Hereby, no extra stabilization parameter is needed unlike in many existing
formulations, e.g. in Nitsche’s method. An element-wise discontinuous approximation allows to
have stress unknowns only along the interface, which can and should be completely condensed at
the element level. With and without condensation, there are no zero entries on the local and global
DOI: 10.1002/nme
tangent stiffness matrix, which simplifies the application of iterative non-linear solvers for the fluid
problem. Furthermore, when stress unknowns are condensed, we could successfully apply AMG
pre-conditioners, which additionally improved solution time.
An important feature with respect to our envisioned 3D FSI implementation is that the fluid
discretization poses no constraint on how to discretize the interface mesh. This fact was clearly
demonstrated by the benchmark examples. If the structure surface is given as a discrete mesh, we
can couple velocity and displacements into the background mesh without introducing additional
intermediate surface meshes as were, for example, required in our original 2D FSI approach [2].
In fact, as the proposed method is applicable to 1D, 2D and 3D problems, our existing 2D FSI
code can be simplified by using the presented method.
The next logical steps here will be the description of the full FSI formulation including moving
boundaries and further mathematical analysis of the embedded Dirichlet approach. Beyond its use
in our FSI scheme, the presented method could directly be applied to 3D mesh tying between
deforming and fixed fluid meshes as sketched for 2D problems in [40, 41].
In our opinion, the presented method differentiates itself, thanks to its relative simplicity and
its clear foundation on a weak form. As the duality of primary field variable and flux field can
be found in many equations describing continua, the new method should be applicable to other
physical fields as we have exemplarily shown for the elasticity problem. Along investigating new
problem types and application areas, the mathematical foundation will be explored in more detail
in future publications.
ACKNOWLEDGEMENTS
The present study is supported by a grant of the ‘Deutsche Forschungsgemeinschaft’ (DFG) through
project WA 1521/1 within DFG’s Research Unit 493 ‘FSI: Modelling, Simulation, and Optimization’.
This support is gratefully acknowledged. Furthermore, the authors would like to thank Ursula M. Mayer
for her continued work on the octree-based interface localization.
REFERENCES
1. Wall WA, Gerstenberger A, Gamnitzer P, Förster C, Ramm E. Large deformation fluid–structure interaction—
advances in ALE methods and new fixed grid approaches. In Fluid–Structure Interaction Modelling Simulation,
Optimisation, Bungartz H-J, Schäfer M (eds). Lecture Notes in Computational Science and Engineering, vol. 53.
Springer: Berlin, 2006.
2. Gerstenberger A, Wall WA. An extended finite element method/Lagrange multiplier based approach for fluid–
structure interaction. Computer Methods in Applied Mechanics and Engineering 2008; 197(19–20):1699–1714.
3. Legay A, Chessa J, Belytschko T. An Eulerian–Lagrangian method for fluid–structure interaction based on level
sets. Computer Methods in Applied Mechanics and Engineering 2006; 195(17–18):2070–2087.
4. Belytschko T, Black T. Elastic crack growth in finite elements with minimal remeshing. International Journal
for Numerical Methods in Engineering 1999; 45(5):601–620.
5. Moës N, Dolbow J, Belytschko T. A finite element method for crack growth without remeshing. International
Journal for Numerical Methods in Engineering 1999; 46(1):131–150.
6. Sukumar N, Chopp DL, Moës N, Belytschko T. Modeling holes and inclusions by level sets in the extended
finite-element method. Computer Methods in Applied Mechanics and Engineering 2001; 190(46–47):6183–6200.
7. Ji H, Dolbow JE. On strategies for enforcing interfacial constraints and evaluating jump conditions with the
extended finite element method. International Journal for Numerical Methods in Engineering 2004; 61(14):
2508–2535.
8. Moës N, Béchet E, Tourbier M. Imposing Dirichlet boundary conditions in the extended finite element method.
International Journal for Numerical Methods in Engineering 2006; 67(12):1641–1669.
DOI: 10.1002/nme
9. Zilian A, Legay A. The enriched space–time finite element method (EST) for simultaneous solution of fluid–
structure interaction. International Journal for Numerical Methods in Engineering 2008; 75(3):305–334.
10. Dolbow J, Franca L. Residual-free bubbles for embedded Dirichlet problems. Computer
Methods in Applied Mechanics and Engineering 2008; 197:3751–3759.
11. Mourad HM, Dolbow J, Harari I. A bubble-stabilized finite element method for Dirichlet constraints on embedded
interfaces. International Journal for Numerical Methods in Engineering 2007; 69(4):772–793.
12. Sanders JD, Dolbow JE, Laursen TA. On methods for stabilizing constraints over enriched interfaces in elasticity.
International Journal for Numerical Methods in Engineering 2009; 78(9):1009–1036.
13. Dolbow J, Harari I. An efficient finite element method for embedded interface problems. International Journal
for Numerical Methods in Engineering 2009; 78(2):229–252.
14. Fernández-Méndez S, Huerta A. Imposing essential boundary conditions in mesh-free methods. Computer Methods
in Applied Mechanics and Engineering 2004; 193(12–14):1257–1275.
15. Hansbo A, Hansbo P. An unfitted finite element method, based on Nitsche’s method, for elliptic interface
problems. Computer Methods in Applied Mechanics and Engineering 2002; 191(47–48):5537–5552.
16. Hansbo A, Hansbo P. A finite element method for the simulation of strong and weak discontinuities in solid
mechanics. Computer Methods in Applied Mechanics and Engineering 2004; 193(33–35):3523–3540.
17. Mergheim J, Kuhl E, Steinmann P. A hybrid discontinuous Galerkin/interface method for the computational
modelling of failure. Communications in Numerical Methods in Engineering 2004; 20(7):511–519.
18. Nitsche J. Über ein Variationsprinzip zur Lösung von Dirichlet-Problemen bei Verwendung von Teilräumen,
die keinen Randbedingungen unterworfen sind. Abhandlungen aus dem Mathematisches Seminar der Universität
1971; 36:9–15.
19. Hemker P, Hoffmann W, van Raalte M. Discontinuous Galerkin discretization with embedded boundary conditions.
Computational Methods in Applied Mathematics 2003; 3:135–158.
20. Lew AJ, Buscaglia GC. A discontinuous-Galerkin-based immersed boundary method. International Journal for
Numerical Methods in Engineering 2008; 76(4):427–454.
21. Gee M, Sala M, Siefert C, Hu J, Tuminaro R. Ml 5.0 smoothed aggregation user’s guide. SAND2006-2649, 2006.
22. Sala M, Tuminaro R. A new Petrov–Galerkin smoothed aggregation preconditioner for nonsymmetric linear
systems. SIAM Journal on Scientific Computing 2008; 31:143–166.
23. Gerstenberger A, Mayer UM, Wall WA. A 3d XFEM/Lagrange multiplier based approach for fluid structure
interaction. Eighth World Congress on Computational Mechanics/Fifth European Congress on Computational
Methods in Applied Sciences and Engineering, ECCOMAS 2008, Venice, July 2008.
24. Zilian A, Fries T-P. A localized mixed-hybrid method for imposing interfacial constraints in the extended
finite element method (XFEM). International Journal for Numerical Methods in Engineering 2009; DOI:
http://dx.doi.org/10.1002/nme.2596.
25. Daux C, Moës N, Dolbow J, Sukumar N, Belytschko T. Arbitrary cracks and holes with the extended finite
element method. International Journal for Numerical Methods in Engineering 2000; 48(12):1741–1760.
26. Legrain G, Moës N, Huerta A. Stability of incompressible formulations enriched with x-FEM. Computer Methods
in Applied Mechanics and Engineering 2008; 197(21–24):1835–1849.
27. Behr MA, Franca LP, Tezduyar TE. Stabilized finite element methods for the velocity–pressure–stress formulation
of incompressible flows. Computer Methods in Applied Mechanics and Engineering 1993; 104(1):31–48.
28. Brooks AN, Hughes TJ. Streamline upwind/Petrov–Galerkin formulations for convection dominated flows with
particular emphasis on the incompressible Navier–Stokes equations. Computer Methods in Applied Mechanics
and Engineering 1982; 32(1–3):199–259.
29. Hughes TJR, Franca LP, Scovazzi G. Multiscale sand stabilized methods. In Encyclopedia of Computational
Mechanics, Stein E, de Borst R, Hughes TJR (eds), vol. 3. Wiley: New York, 2004.
30. Gresho P, Sani R. Incompressible Flow and the Finite Element Method, Volume 2: Isothermal Laminar Flow.
Wiley: New York, 2000.
31. Engels H. Numerical Quadrature and Cubature. Academic Press: London, New York, 1980.
32. Peano A. Gauss–Lobatto integration of high precision tetrahedral elements. International Journal for Numerical
Methods in Engineering 1982; 18(2):311–313.
33. Mayer UM, Gerstenberger A, Wall WA. Interface handling for three-dimensional higher-order XFEM computations
in fluid–structure interaction. International Journal for Numerical Methods in Engineering 2009; 554:4545–5454.
DOI: http://dx.doi.org/10.1002/nme.2600.
34. Hamel G. Spiralförmige Bewegungen zäher Flüssigkeiten. Jahresbericht der Deutschen Mathematiker-Vereinigung
1916; 25:34–60.
DOI: 10.1002/nme
35. Jeffery GB. The two-dimensional steady motion of a viscous fluid. Philosophical Magazine 1915; 29:455–465.
36. Rosenhead L. The steady two-dimensional radial flow of viscous fluid between two inclined plane walls.
Proceedings of the Royal Society of London, Series A, Mathematical and Physical Sciences 1940; 175(963):
436–467.
37. Corless RM, Assefa D. Jeffery-hamel flow with maple: a case study of integration of elliptic functions in a cas.
ISSAC ’07: Proceedings of the 2007 International Symposium on Symbolic and Algebraic Computation. ACM:
New York, NY, U.S.A., 2007.
38. Bagheri B, Scott LR, Zhang S. Implementing, using high-order finite element methods. Finite Elements in
Analysis and Design 1994; 16(3–4):175–189.
39. Schäfer M, Turek S. Flow simulation with high-performance computers II. Benchmark Computations of Laminar
Flow Around a Cylinder. Notes on Numerical Fluid Mechanics, vol. 52. Vieweg: Braunschweig, 1996; 547–566.
40. Gerstenberger A, Wall WA. Enhancement of fixed-grid methods towards complex fluid–structure interaction
applications. International Journal for Numerical Methods in Fluids 2008; 57(9):1227–1248.
41. Wall WA, Gamnitzer P, Gerstenberger A. Fluid–structure interaction approaches on fixed grids based on two
different domain decomposition ideas. International Journal of Computational Fluid Dynamics 2008; 22(6):
411–427.
DOI: 10.1002/nme

An Embedded Dirichlet Formulation For 3D Continua: A. Gerstenberger and W. A. Wall

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

An Embedded Dirichlet Formulation For 3D Continua: A. Gerstenberger and W. A. Wall

Uploaded by

Copyright:

Available Formats

INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN ENGINEERING

Int. J. Numer. Meth. Engng 2010; 82:537–563

An embedded Dirichlet formulation for 3D continua

A. Gerstenberger and W. A. Wall∗, †

Received 19 March 2009; Revised 7 August 2009; Accepted 20 August 2009

Contract/grant sponsor: DFG-FG493; contract/grant number: WA 1521/1

Copyright q 2009 John Wiley & Sons, Ltd.

2. EMBEDDED DIRICHLET CONDITIONS FOR LINEAR-ELASTIC PROBLEMS

2.1. Problem definition

2.2. Embedded Dirichlet formulation using primary stress unknowns

Integration by parts using Equation (10) and replacing rd with r at i gives

2.3. Spatial discretization

Kdd : −(∇d, rd )e (30)

Gd : −(d, r·n)e+ (33)

[Kdd ]e and [f N ]e (39)

2.4. Relation to the Hellinger–Reissner variational principle and Nitsche’s method

3. EMBEDDED DIRICHLET FORMULATION FOR THE INCOMPRESSIBLE

3.1. Definition of the fluid problem

3.2. Discretization in time

3.3. Weak form of the time-discrete NS equations

Equation (70). The complete weak form is given as

(v, u·∇u)+ +(∇v, − pI+su )+ +(q, ∇ ·u)+

= (v, ĥ) N −(c·nf , ûi )+ (72)

vmod,h = vh + M [f uh ·∇vh +∇q h ] (73)

For tri-linear shape functions we use m K = 13 , for tri-quadratic shape functions m K = 12

3.4. Fluid discretization in space

3.5. Solution of the fluid equations

Gu : −(c·nf , u)+ (85)

Gu i : −(c·nf , ûi )+ (86)

With G∗u = Gu K−1

3.6. Numerical integration of domain and surface integrals

4.1. Shear flow without pressure gradient

4.2. Hydrostatic pressure

p(x) = −xx (x) = −yy (x) = −zz (x) in + (95)

4.3. Convergence analysis using the Jeffery–Hamel flow

4.4. Benchmark computations

4.4.1. 2D benchmarks. The 2D benchmarks—calculated as quasi 2D problems with our present

Table I. Results for the stationary 2D benchmark case 2D-1.

Fluid element type NuDOF

Table II. Results for the instationary 2D benchmark case 2D-3.

Fluid element type NuDOF

Table III. Results for 3D benchmark cases 3D-1Q and 3D-1Z.

Fluid element type NuDOF

4.5. Channel flow over a deforming thick plate

You might also like

vmod,h = vh + M [f uh ·∇vh +∇q h ] (73)