Professional Documents
Culture Documents
Isogeometric Fluid-Structure Interaction: Theory, Algorithms, and Computations
Isogeometric Fluid-Structure Interaction: Theory, Algorithms, and Computations
Isogeometric Fluid-Structure Interaction: Theory, Algorithms, and Computations
DOI 10.1007/s00466-008-0315-x
ORIGINAL PAPER
Received: 25 May 2008 / Accepted: 16 June 2008 / Published online: 6 August 2008
© Springer-Verlag 2008
123
4 Comput Mech (2008) 43:3–37
are also referred to as loosely- and strongly-coupled, respec- element needs to be a parameter. The element routines des-
tively. cribed in Hughes [28] possess this property. Some additional
In staggered approaches, the fluid, solid and mesh move- simple data structures are also required. We plan to describe
ment equations are solved sequentially, in uncoupled fashion. these implementational aspects in detail in future work. In
This enables the use of existing well-validated fluid and addition to including finite element analysis as a special case,
structural solvers, a significant motivation for adopting this isogeometric analysis offers the following possibilities:
approach. In addition, for many problems the staggered Precise and efficient geometry modeling; simplified mesh
approach works well and is very efficient. However, diffi- refinement and order elevation procedures; superior approxi-
culties in the form of lack of convergence have been noted in mation properties; smooth basis functions with compact sup-
a number of situations and considerable recent literature has port; and, ultimately, the integration of geometric design and
been devoted to a discussion of these problems and attempts analysis.
to circumvent them. It seems that “light” structures interac- In Sect. 2, we begin with a presentation of general conti-
ting with “heavy” fluids are particularly problematic (see, nuum mechanics on domains undergoing arbitrary motion.
e.g., [50,52,53]). Another situation that has proved difficult The motion is described in terms of a space-time mapping of
is when an incompressible fluid region is fully contained by an arbitrary reference domain. The description is specialized
a solid, such as for the filling of a balloon with water [42]. for the cases of interest, namely, the Lagrangian description
In cases like this, special modifications need to be introdu- of a solid and the arbitrary Lagrangian–Eulerian (ALE) des-
ced to achieve success. On the other hand, Farhat et al. [18] cription of the fluid [32]. In the ALE description the domain is
has reported consistent success for staggered approaches to in motion but the motion does not coincide with the motion
compressible fluids. As of this writing, these claims have not of material particles. The material particles are in relative
been fully reconciled with those for incompressible fluids. motion with respect to the motion of the referential domain
It is somewhat puzzling because incompressible flows may (mobilis mobili). In the derivations, it is found helpful to
be thought of as limiting phenomena within a compressible employ a space-time Piola transformation. Several useful
formulation. forms of the conservation equations are presented, specifi-
In monolithic approaches, the fluid, solid and mesh move- cally, the advective, conservative and mixed forms.
ment equations are solved simultaneously in fully-coupled The mixed form seems to be the preferred one for imple-
fashion. The main advantage is that monolithic solvers tend menting the ALE description in the semi-discrete format.
to be more robust. Many of the problems encountered with the A key relation is identified that influences the ability of the
staggered approach are completely avoided with the mono- formulation to conserve momentum in the discrete case. Spe-
lithic approach. Of course, there is a price to pay in that the cific forms of the fluid and solid equations used in the sequel
monolithic approach necessitates writing a fully-integrated are presented.
fluid-structure solver, precluding the use of existing fluid In Sect. 3 the variational formulation is described. The
and structure software. However, recent attempts have been constituent formulations for the solid, fluid, mesh motion,
made to design schemes that could in principle use existing and coupled problems are presented. In Sect. 4 the discrete
fluid and structure software in the context of fully-coupled spaces and continuity conditions at the fluid-structure inter-
approaches [24]. face are described. The semi-discrete problem is then intro-
Our aim in this work was to develop a robust isogeometric duced and a discussion of conservation properties follows.
analysis formulation for fluid-structure interaction. Conse- The conservation laws of interest are mass, momentum, and
quently, we opted for a monolithic approach, but we note the so-called geometric conservation law. It is argued that all
various staggered techniques can also be obtained from the these are satisfied in the semi-discrete case, and the mass and
fully-coupled formulation at the linearization stage by remo- geometric conservation laws also hold in the fully-discrete
ving certain blocks from the tangent matrix and employing case. Momentum conservation in the fully discrete case
a fixed, small number of Newton steps (see, e.g., [49]). depends on whether or not the time integrator preserves the
Isogeometric analysis was introduced in Hughes et al. key relation alluded to previously. If it does not, then the
[30] and further developed in [2–5,13,14,16,74]. Isogeome- momentum is conserved up to the truncation errors introdu-
tric analysis is based on the technologies used in enginee- ced by the time integration algorithm in this key relation. The
ring design, animation, graphic art, and visualization. It is a time integration is performed by the generalized-α method
generalization of finite element analysis and it is a relatively [12,36]. We present, apparently for the first time, a subfa-
simple matter to develop an isogeometric code from an exis- mily of the generalized-α family of methods that is dissipa-
ting finite element code. The main change involves writing a tive, second-order accurate, and unconditionally stable for
new shape function routine (see Hughes [28, Chap. 3]). It also the coupled fluid-structure case. In Sect. 5 the linearization
requires an element routine that is written in a parameteri- of the system is presented including a discussion of the cal-
zed way, in particular, the number of degrees-of-freedom per culation of shape derivatives, that is, derivatives taken with
123
Comput Mech (2008) 43:3–37 5
123
6 Comput Mech (2008) 43:3–37
The following relationship is also easily verified referring to the time coordinate, and 1, 2, 3 referring to the
(4) space coordinates. Summation over the range of the indices
Jˆ(4) ≡ det F̂ = Jˆ ≡ det F̂, (7) is implied for repeated indices. With these one can compute
(4) as follows:
that is, the determinants of F̂ and F̂ are equal, a conse-
(4) (4)
quence of time synchronization. ∂( Jˆ(4) F̂Aa −1 ) ∂ Jˆ(4) (4) −1 ∂ F̂ −1
(4)
Let γ x : Q x → R4 denote a space-time flux vector = F̂Aa + Jˆ(4) Aa (14)
∂ yA ∂ yA ∂ yA
field defined on the spatial configuration. In order to preserve (4)
∂ Jˆ (4) ∂ Jˆ ∂ F̂bB
(4)
the conservation structure in the reference configuration, we = (4) ∂ y
(4)
define γ y , a space-time flux vector field in the referential ∂ yA ∂ F̂bB A
configuration, as (4)
(4) ∂ 2 φ̂b
= cof F̂bB
γ (4) ˆ(4) F̂ (4) −1 γ (4)
y = J x . (8) ∂ y A ∂ yB
(4)
(4) ∂ 2 φ̂b
This is the space-time Piola transformation. With (8), we can = Jˆ(4) F̂bB −T (15)
prove the following space-time integral theorem: ∂ y A ∂ yB
(4) (4)
∂ F̂Ca −1 (4) −1 ∂ F̂bB (4) −1
∇x · γ x d Q x = ∇ y(4) · γ (4)
(4) (4) = − F̂Cb F̂
y d Qy, (9) ∂ yA ∂ y A Ba
Qx Qy
(4) −1 ∂ 2 φ̂b(4) (4) −1
= − F̂Cb F̂ (16)
where ∂ y A ∂ y B Ba
(4) (4)
∂ ∂ F̂Aa −1 (4) −1 ∂ φ̂b
2
∇x(4) ≡ ∂t (10) = − F̂Ab F̂ (4) −1 (17)
∇x ∂ yA ∂ y A ∂ y B Ba
∂ (4) (4)
∂( Jˆ(4) F̂Aa −1 ) ∂ 2 φ̂b
∇ y(4) ≡ ∂r (11) = Jˆ(4)
(4) (4)
F̂ −1 F̂Aa −1
∇y ∂ yA ∂ y A ∂ y B Bb
(4)
are the space-time gradient operators. The proof of (9) is ∂ 2 φ̂b (4) (4)
based on the Piola identity, namely, − Jˆ(4) F̂ −1 F̂Ba −1
∂ y A ∂ y B Ab
∂ 2 φ̂b
(4) −1 (4) −1
(4) (4)
∇ y(4) · Jˆ(4) F̂ −1 = 0, (12) = Jˆ(4) F̂Bb F̂Aa
∂ y A ∂ yB
the transformation formula for volume elements,
(4) (4)
− F̂Ab −1 F̂Ba −1
d Q x = Jˆ(4) d Q y , (13)
= 0. (18)
and straightforward calculations.
One must be careful in analysis on a space-time manifold This is the component form of (12). The last line of (18)
(4)
∂ 2 φ̂
because many commonly invoked results depend on the exis- follows from the fact that ∂ y A ∂by B is symmetric in A and B,
tence of a Riemannian metric, which in the present case does and the term in parenthesis is skew-symmetric in A and B.
not exist. For example, there is no well-defined unit normal The referential domain can take on several interpreta-
vector to the boundary of ∂ Q x , in contrast with ∂ Q y , for tions. In fluid-structure interaction, it is usually taken to be
which there is a well-defined unit outward normal vector. the initial configuration of the problem domain. When the
However, many important results can be obtained without fluid domain moves, it becomes the so-called arbitrary
the existence of a Riemannian metric, or any metric for that Lagrangian–Eulerian (ALE) description. The material des-
matter. The subject of analysis on manifolds without metric cription is often utilized for the solid domain. By virtue of
structure (i.e., differential topology) is well-developed (see, the fact that the referential domain is arbitrary, it can be spe-
Flanders [21], Spivak [57], Lang [44,45], Guillermin and cialized for the material description. This representation is
Pollack [23], Bishop and Goldberg [8], Marsden and Hughes also important for deriving material forms of the conserva-
[47]). A proof of (9) using exterior calculus methodology is tion laws. A summary of notations and important results for
presented in Appendix A. the material description follows. We set y = X ∈ X ⊂ R3 ,
A way to understand (8) is to assume Cartesian coordinate a “particle” in the material domain, and r = s ∈ ]0, T [,
charts on both Q x and Q y . We will denote these charts as the material time. As before, the differential operator ∂/∂s,
{xa } and {y A }, respectively. We use lower case indices (i.e., needs to be distinguished from ∂/∂r and ∂/∂t. In the case
a, b, c, . . .) to denote the current configuration and upper of ∂/∂s, known as the material time derivative, the material
case indices (i.e., A, B, C, . . .) to denote the reference confi- point (i.e., particle) X ∈ X is held fixed. The material des-
guration. All the indices run over the range 0, 1, 2, 3 with 0 cription is often referred to as the Lagrangian description.
123
Comput Mech (2008) 43:3–37 7
The material and referential clocks are also synchronized. that the domain x changes with time and this must be taken
Let Q X = X ×]0, T [⊂ R4 and φ (4) : Q X → Q x , where into account in the time differentiation of the left-hand sides.
Standard procedures (see [47]) yield the following:
t
Q x = (x, t)| = φ (4) (X, s)
x ∂α
dx = (γ x − αv) n x d ∂x + β dx (32)
T
s ∂t
= ∀(X, s) ∈ Q s (19) x ∂x x
φ(X, s)
⎡ ⎤ and
∂t ∂t
∂s ∂ X 1 0T
F (4) = Dφ (4) = ⎣ ∂φ ∂φ ⎦ = , (20) ∂α
v F dx = ( x −α ⊗ v)n x d ∂x + β dx , (33)
∂s ∂ X ∂t
x ∂x x
∂φ
F= (deformation gradient), (21)
∂X where the partial time derivative ∂/∂t is taken with the spatial
u(X, s) = φ(X, s) − X(particle displacement), (22) coordinate x held fixed. Note that the boundary terms are
modified by the fluxes involving the material particle velocity
∂φ ∂u
v= = (particle velocity), (23) v. Employing the divergence theorem on the boundary terms
∂s ∂s
in (32) and (33) gives
(4) −1 1 0T
F = , (24) ∂α
−F −1 v F −1 + ∇x · (αv − γ x ) − β dx = 0 (34)
∂t
(4) −T 1 −v T F −T x
F = . (25)
0 F −T and
(4) (4)
J ≡ det F = J ≡ det F, (26) ∂α
+ ∇x · (α ⊗ v − x ) − β dx = 0. (35)
γ (4) T
= J γ (4) T (4) −T
F . (27) ∂t
X x x
(4) (4)
∇x(4) · γ (4)
x d Qx = ∇X · γ X d Q X , (28) Equations (34) and (35) represent the scalar and vector master
Qx QX
balance laws on the spatial domain written in a divergence
∂
form. We are going to employ Eq. (10) in (34) and (35) to
(4) ∂s obtain the same balance laws on the referential space-time
∇X ≡ (29)
∇X domain, also in the divergence form. For this purpose we
integrate (34) and (35) in time
2.2 Master balance laws
∂α
+ ∇x · (αv − γ x ) − β d Q x = 0 (36)
In this section we derive generic master balance laws for ∂t
Qx
vectors and scalars, and present their various forms in the
∂α
spatial and referential domains. + ∇x · (α ⊗ v − x ) − β d Q x = 0, (37)
The following master balance laws hold on the spatial ∂t
Qx
domain x (see, e.g., [47] for background):
Scalar case and then change variables, using (10)–(13),
ˆ
d ∂ Jα
−1 −1
α dx = γ x n x d ∂x + β dx ,
T
(30) +∇ y · Jˆα F̂ (v− v̂)− Jˆ F̂ γ x − Jˆβ d Q y = 0
dt ∂r
x ∂x x
Qy
Vector case (38)
d
α dx = x n x d∂x + β dx . ∂ Jˆα −T
(31) + ∇ y · Jˆ α ⊗ (v − v̂) F̂
dt ∂r
x ∂x x Qy
−T
In the above equations α, a scalar, and α, a vector, are the − Jˆ x F̂ − Jˆβ d Q y = 0. (39)
conserved quantities of interest, β, a scalar, and β, a vector,
are the volumetric source terms, and γ Tx n, a scalar, and x n, The advantage of these forms of the master balance law is that
a vector, are the surface fluxes. n is the unit outward normal one is free to choose any referential domain that is convenient
to ∂x , the boundary of x , γ x is a vector, and x is a for a given problem.
second-rank tensor. Note that the unit outward normal vector Particularly useful forms of (38) and (39) are obtained by
to the 3-dimensional spatial slices is well defined. Recall choosing the reference domain to be the material domain. In
123
8 Comput Mech (2008) 43:3–37
this case, we set y = X, v̂ = v, F̂ = F, and Jˆ = J , in (38) derivatives are taken with respect to spatial coordinates,
and (39), and obtain, respectively, leading to a mixed representation. These equations may be
simplified further. Assuming sufficient smoothness of the
∂ Jα
− ∇ X · J F −1 γ x − Jβ d Q X = 0 (40) fields, we compute
∂s
QX
∂ ˆα
J ∂ ˆ
J ∂α
and Jˆ −1
= Jˆ −1
α + Jˆ
∂r ∂r ∂r
∂ Jα
− ∇ X · J x F −T − J β d Q X = 0. (41) ∂α
∂s =J ˆ −1 ˆ
α J ∇x · v̂ + J ˆ
QX ∂r
∂α
As may be noted, this results in a simplification of the general = α∇x · v̂ + , (46)
case due to the fact that v − v̂ = 0. This is an advantage of ∂r
the Lagrangian description. where, going from the first to the second line, we have used
the key identity
Remark 2.3 In order to obtain local forms of the balance
∂ Jˆ
laws, we assume the integrands of any of the previous inte- = Jˆ∇x · v̂, (47)
gral balance laws are continuous and the domain may be ∂r
taken arbitrarily small about any space-time point. In this which we will discuss further in Sect. 4.3. Also note that
case, using standard arguments (see [47]), it follows that the
∇x · α(v − v̂) = (v − v̂) · ∇x α + α∇x · v −α∇x · v̂ . (48)
integrands must vanish pointwise. This is referred to as the
localization argument. Similarly, for a vector quantity, we get
We can also state the mixed form of the master balance ∂ Jˆα ∂α
Jˆ−1 = α∇x · v̂ + , (49)
laws, where “mixed” refers to the fact that time and space ∂r ∂r
derivatives are associated with different descriptions. This and
form is often used as a staring point of ALE
formulations
of
balance laws. We first recognize that both Q y = T y and ∇x · α ⊗ (v − v̂) = (v − v̂) · ∇x α + α∇x · v −α∇x · v̂ .
y do not change in time. Furthermore, using time synchro- (50)
nization and a localization argument with respect to time in
(38) and (39) leads to Substituting (46) and (48) in (44), and (49) and (50) in (45)
leads to simplified forms of the integral balance statements
∂ Jˆα
−1 −1
in the vector and scalar cases,
+∇ y · Jˆα F̂ (v− v̂)− Jˆ F̂ γ x − Jˆβ d y = 0
∂r ∂α
y + (v − v̂) · ∇x α + α∇x · v − ∇x · γ x − β dx = 0
∂r
(42) x
and (51)
∂ Jˆα −T and
+ ∇ y · Jˆ α ⊗ (v − v̂) F̂
∂r ∂α
y + (v − v̂) · ∇x α + α∇x · v − ∇x · x −β dx = 0
∂r
−T x
− Jˆ x F̂ − Jˆβ d y = 0, (43)
(52)
which hold at every time instant. Changing variables y →
respectively. It is important to note the disappearance in (51)
x and using Eq. (10) and (13) yields and (52) of α∇x · v̂ and α∇x · v̂ [i.e., the boxed terms in
(46), (48)–(50)]. This cancellation is due to (47). In the fully-
∂ Jˆα
Jˆ−1 + ∇x · (α(v − v̂) − γ x ) − β dx = 0 (44) discrete case, (47) may not be satisfied identically, which has
∂r
x implications to the discrete conservation of momentum. We
will return to this point in Sect. 4.3.
and
∂ Jˆα 2.3 Discussion of discretization choices for balance laws
Jˆ−1 + ∇x · α ⊗ (v − v̂) − x − β dx = 0. (45)
∂r on moving domains
x
Note that in (44) and (45), partial time derivatives are left In the previous section we have derived integral balance laws
with respect to the referential time variable, while the spatial on the referential and spatial domains. At the continuous or
123
Comput Mech (2008) 43:3–37 9
infinite-dimensional level, all the instantiations of these laws following approach is taken. On the referential domain y
are completely equivalent. The situation changes when one one defines basis functions N̂ A ( y), A ∈ I, and assumes the
tries to numerically approximate the equations emanating following expansion for the solution variable as a function
from the balance laws. In this section we discuss the suitabi- of the referential domain variables
lity of the existing computational approaches for partial diffe-
rential equations arising from different forms of the balance û( y, r ) = Û A (r ) N̂ A ( y), (59)
laws. We focus on ALE and space-time methods (see, e.g., A∈ I
[31,34,41,64,65,69]). which, in turn, results in the following expression for the
In the space-time finite element method the approximation referential time derivative
space consists of basis functions that explicitly depend on ∂ Û A (r )
∂ û( y, r )
space and time, denoted N A (x, t), where A spans the index = N̂ A ( y). (60)
set I of functions on a space-time mesh defined on Q x . Let- ∂r ∂r
A∈ I
ting u = u(x, t) denote a generic space-time field, its partial
The basis in the spatial configuration x is the push forward
time and spatial derivatives are expressed as follows:
of N̂ A ( y) defined by
u(x, t) = U A N A (x, t) (53) −1 −1
N A (x, t) ≡ N̂ A (φ̂ (x, t)) = N̂ A ◦ φ̂ (x, t) (61)
A∈ I
∂u ∂ NA In the spatial configuration, a solution field is now defined as
(x, t) = UA (x, t) (54)
∂t ∂t −1 −1
A∈ I u(x, t) = û(φ̂ (x, t), t) = Û A (t) N̂ A (φ̂ (x, t))
and A∈ I
∂u ∂ NA = Û A (t)N A (x, t), (62)
(x, t) = UA (x, t) (55)
∂x ∂x A∈ I
A∈ I
where the U A ’s are real coefficients. Note that the partial time and its gradient with respect to the spatial coordinates is easily
derivative in (54) is, by definition, taken with x fixed. With computed as
this observation, the forms of the balance equations given by ∂u(x, t) ∂ N A (x, t)
= Û A (t) (63)
(36) and (37) are well-suited for space-time treatment. ∂x ∂x
One may also employ the space-time technique for discre- A∈ I
tizing the balance equations on the referential domain, (38) Finally, on the spatial domain, the referential time derivative
and (39). Just as before, let N̂ A ( y, r ) be the basis functions of the solution field becomes
associated with the space-time discretization of Q y . Now ∂ Û A (t)
∂ û −1 −1
the solution field and its partial time and space derivatives (φ̂ (x, t), t) = N̂ A ◦ φ̂ (x, t)
∂r ∂r
become A∈ I
∂ Û A (t)
û( y, r ) = Û A N̂ A ( y, r ) (56) = N A (x, t) (64)
A∈ I ∂t
A∈ I
∂ û ∂ N̂ A
( y, r ) = Û A ( y, r ) (57) In (44) and (45) the spatial gradient and the referential time
∂r ∂r derivative are evaluated according to (63) and (64), respecti-
A∈ I
vely. This is the essence of the discrete ALE approach. These
and
particularly simple expressions explain in part the popularity
∂ û ∂ N̂ A of ALE methods for moving domain problems. Comparing
( y, r ) = Û A ( y, r ). (58)
∂y ∂y expressions (62) and (64) we note that a referential time
A∈ I
derivative of the solution in the spatial configuration may
As before, the Û A ’s are real coefficients, but the partial time be obtained by simply taking a time derivative of its coeffi-
derivative in (57) is now taken with the referential coordinate cients, thus rendering the semi-descrete equations amenable
y held fixed. to finite-difference-in-time treatment. As a result, we may
The mixed form of the balance laws, as expressed by think of ALE as an extension of the classical semi-discrete
Eq. (44) and (45), is not amenable to space-time discreti- approach to moving domain problems.
zation because partial time and space derivatives employed The semi-discrete approach may also be applied to the
in the formulation of the conservation equations are taken equations emanating from the master balance law written
with respect to different descriptions, namely, the referen- with respect to the referential domain [see Eqs. (42) and (43)].
tial and spatial description, respectively. In this case, the An expression of the form (59) may be employed in this case
123
10 Comput Mech (2008) 43:3–37
due to the orthogonality of space and time in the referential tensors, respectively, as
description.
P = J σ F −T (70)
123
Comput Mech (2008) 43:3–37 11
of ψdil precludes material instabilities for states of strong discretized, it trivially satisfies the so-called Discrete Geo-
compression. For this definition of the elastic potential, the metric Conservation Law (DGCL). The DGCL states that in
second Piola-Kirchhoff stress tensor becomes the absence of body forces and surface tractions, the discrete
scheme must preserve a constant velocity solution. For a dis-
1 1
S = µs J −2/3 I − trC C −1 + κ s (J 2 − 1)C −1 , (80) cussion of the importance of conservation and satisfaction of
3 2
the DGCL for moving domain problems see [17,22,46].
and the fourth-order tensor of elastic moduli is Substituting α = ρ, γ = 0 and β = 0 in (50), we arrive
at
∂ 2ψ 2 s −2/3
C=4 = µ J trC + κ J C −1 ⊗ C −1
s 2
∂ρ
∂ C∂ C 9 + (v − v̂) · ∇x ρ + ρ∇x · v dx = 0. (86)
∂r
2 s −2/3
+ µ J trC − κ (J − 1) C −1 C −1
s 2 x
3
Assuming that the fluid has constant mass density (i.e., the
2
− µs J −2/3 (I ⊗ C −1 + C −1 ⊗ I). (81) flow is incompressible) and localizing the above equation to
3 a point in space and time, we obtain the following form of
In (81) the ⊗ symbol is used to denote the outer product of the mass conservaton equation,
two second-rank tensors, that is,
∇x · v = 0, (87)
−1 −1 −1 −1
(C ⊗C ) IJKL ≡ (C ) IJ (C ) KL , (82) which manifests the incompressibility constraint. To arrive
and at the conservation of linear momentum we use (52) and set
α = ρv, = σ , and β = ρ f
(C −1 ) IK (C −1 ) JL + (C −1 ) IL (C −1 ) JK
(C −1 C −1 ) IJKL ≡ ∂ρv
2 + (v − v̂) · ∇x ρv
(83) ∂r
x
Parameters µs
and κs
may be determined by the Lamé + ρv∇x · v − ∇x · σ − ρ f dx = 0. (88)
constants of the linear elastic model, denoted µl and λl , by Using the assumption of constant mass density, (87) and loca-
considering the case when the current and the reference confi- lizing the result to a point in space and time, we obtain
gurations coincide. Then, by inspection,
∂v
ρ + ρ(v − v̂) · ∇x v − ∇x · σ = ρ f . (89)
µs = µl (84) ∂r
2 To complete the specification of the fluid problem, we assume
κ s = λl + µl . (85)
3 that the flow is Newtonian with the following definition of
Thus, µs and κ s are the classical shear and bulk moduli, the Cauchy stress tensor
respectively. σ = − p I + 2µ∇xs v, (90)
where p is the pressure, µ is the dynamic viscosity, I is the
2.4.2 Formulation of the fluid problem
second-rank identity tensor, and ∇xs = 1/2(∇x + ∇xT ) is the
symmetric gradient.
While the solid problem is written using a Lagrangian des-
cription, an ALE approach is adopted for the fluid problem. Remark 2.4 We note that if f = 0 and σ · n = 0 on ∂x ,
Although ALE is widely used for fluid flow in moving then v = constant identically satisfies (88).
domains, the authors wish to point out the recent works on the
Particle Finite Element Method (PFEM, see, e.g., [35] and
references therein), which makes use of a Lagrangian des- 3 Variational formulation of the coupled fluid-structure
cription of fluid mechanics, enabling straight-forward solu- interaction problem at the continuous level
tion of very complicated flows and fluid-structure interaction
problems. To arrive at the formulation of the fluid problem Let 0 ≡ y ⊂ Rd , d = 2, 3, represent the combined
employed in this work, Eq. (51) and (52) are taken as a depar- fluid and solid domain in the initial configuration, which
ture point. Note that we work with the so-called advective serves simultaneously as the reference configuration. Let
forms of the master balance laws rather than their conserva- φ̂ : 0 → t ≡ x |t ⊂ Rd denote the motion of the
tive counterparts given by (44) and (45). We will show later fluid-solid domain, as before.
in this paper that our final semi-descrete formulation satis- The domain 0 admits the decomposition
fies global conservation of mass and linear momentum. Fur-
f
thermore, the advantage of the advective form is that, when 0 = 0 ∪ s0 , (91)
123
12 Comput Mech (2008) 43:3–37
where
2
s∂ u
B (w , u) = w
s s s
, ρ0 2 + ∇ X w s , F S s , (96)
∂s s 0
0
and
where
0s,N is the Neumann part of the solid boundary, hs is
the boundary traction vector, ρ0s is the density of the solid in
Fig. 2 Abstract setting for the fluid-structure interaction problem.
Depiction of the initial and the current configurations related through the initial configuration, f s is the body force per unit mass,
the ALE mapping. The initial configuration also serves as the reference and (·, ·)D is the L 2 inner product with respect to domain D.
configuration The above relations are written over the initial configuration
s0 , which is also the material configuration. The subscript
f
X on the partial derivative operators indicates that the deri-
where 0 is the subset of 0 occupied by the fluid, and s0 vatives are taken with respect to the material coordinates X.
is the subset of 0 occupied by the solid. The decomposition
is non-overlapping, that is
3.2 Motion of the fluid subdomain
f
0 ∩ s0 = ∅. (92)
This section gives a weak formulation of the motion of the
Likewise, fluid subdomain. Partial differential equations of linear elas-
tostatics subject to Dirichlet boundary conditions coming
f
t = t ∪ st , (93) from the displacements of the solid region are employed to
define the ALE mapping φ̂( y, r ) of the fluid domain. For pre-
with cise conditions on the regularity of the ALE map, see Nobile
[51]. In the discrete setting, the fluid subdomain motion pro-
f
t ∩ st = ∅. (94) blem is referred to as “mesh moving.” The fluid subdomain
motion problem may be thought of as a succession of ficti-
fs
Let
0 denote the interface between the fluid and the tious linear elastic boundary-value problems designed simply
solid regions in the initial configuration, and, analogously, to produce a smooth evolution of the fluid mesh.
fs
let
t be its counterpart in the current configuration. The We write φ̂ r ( y) = φ̂( y, r ). Consequently, φ̂ r : 0 → r
−1
setup is illustrated in Fig. 2. It is important to emphasize that and φ̂ r : r → 0 . Likewise, we define the displacement
the motion of the fluid domain is not the particle motion of of the reference domain as
the fluid. It does, however, conform to the particle motion of
the solid at the fluid-solid interface because the Lagrangian û( y, r ) = φ̂( y, r ) − y (98)
description is adopted for the solid.
and write ûr ( y) = û( y, r ). Note that ûr is defined on 0 and
represents the displacement of the reference configuration at
3.1 Solid problem time r .
Let t˜ be the configuration of 0 at t˜ < t. We think of
This section gives a weak formulation of the solid in the this as a configuration “nearby” t that in numerical com-
Lagrangian description. Let V s = V s (s0 ) denote the trial putations will typically represent the final configuration of
solution space for displacements and let W s = W s (s0 ) the previous time step. We wish to push forward the func-
denote the trial weighting space for the linear momentum tions defined on 0 to t˜. We write φ̂ t˜ : 0 → t˜ and
equations. Let u denote the displacement of the solid with −1 −1 −1
φ̂ t˜ : t˜ → 0 . Then ût ◦ φ̂ t˜ and ût˜ ◦ φ̂ t˜ are the dis-
respect to the initial configuration and let w s be the weighting
placements of the reference domain at time t and t˜, respec-
function for the momentum equation. We also assume that
tively, but both are defined with respect to the configuration
the displacement satisfies the boundary condition, u = g s on
of the reference domain at time t˜, namely t˜. We write
123
Comput Mech (2008) 43:3–37 13
We would like to remind the reader that φ̂ t˜ and ût˜ are consi- Note that due to (100), the ALE map φ̂ in (105) is continuous
−1
dered known when we solve for ût ◦ φ̂ t˜ . at the fluid-solid interface. Recall also that the velocity of the
f
Let V m = V m (t˜ ) denote the trial solution space of dis- fluid domain is obtained by taking a partial time derivative
f of û with y held fixed, that is, v̂ = ∂ û/∂r .
placements and let W m = W m (t˜ ) denote the weighting
space for the elastic equilibrium equations. As usual, kine-
3.3 Fluid problem
matic boundary conditions are built into the definitions of the
spaces, namely,
In this section we give a weak formulation of the incompres-
f d −1 fs sible Navier–Stokes fluid on a moving domain in the ALE
V = u | u ∈ H (t˜ ) , u = ut ◦ φ̂ t˜ on
t˜
m m m 1 m
description. The motion of the fluid domain was construc-
f
(100) ted in the previous section. Let V f = V f (t ) denote the
d trial solution space of velocities and pressures and let W f =
f f s f
Wm = wm | wm ∈ H 1 (t˜ ) , wm = 0 on
t˜ (101) W f (t ) denote the trial weighting space for the momen-
tum and continuity equations. Let {v, p} denote the particle
where ut is the particle displacement at time t. Note that in velocity-pressure pair and {w f , q f } the weighting functions
our formulation ut will be an unknown and will be solved for the momentum and continuity equations. We also assume
for simultaneously along with ût in a coupled fashion. The that the fluid particle velocity field satisfies the boundary
f,D
variational formulation of the problem is stated as follows: condition, v = g f on
t , the Dirichlet part of the fluid
−1 boundary. The variational formulation is stated as follows:
Find ût ◦ φ̂ t˜ ∈ V m such that ∀w m ∈ W m ,
Find {v, p} ∈ V f such that ∀{w f , q f } ∈ W f ,
−1
B m (wm , ût ◦ φ̂ t˜ ) = F m (wm ), (102) B f ({w f , q f }, {v, p}; v̂) = F f ({w f , q f }) (106)
where where
123
14 Comput Mech (2008) 43:3–37
with the following auxiliary relations holding in the sense of which, together with (110), produces proper fluid-solid inter-
traces: face conditions.
∂u −1
v
f s = ◦ φ̂ , (110)
t ∂t
t
fs
4 Formulation of the fluid-structure interaction
−1
w f
f s = w s ◦ φ̂ f s . (111) problem at the discrete level
t
t
Relationship (110), the kinematic constraint, equates the fluid In this section we give a formulation of the fluid-structure
particle velocity with that of the solid at the fluid-solid boun- interaction Eq. (109) in the discrete setting. We begin by
dary. Equation (111) leads to the compatibility of the Cauchy defining the spatial discretization of the problem. It is exactly
stress vector at the fluid-solid interface. To demonstrate this the same for finite elements and NURBS-based isogeome-
fact, we first set w m = 0 and focus on the fluid and solid tric analysis. Having defined the semi-discrete forms, we
parts of the coupled problem (109). Integrating by parts in present the time stepping algorithm, which is the generalized-
(109) and assuming sufficient regularity of the solution fields α method of Chung and Hulbert [12].
gives
4.1 Approximation spaces and enforcement of kinematic
0 = w f , L f (v, p; v̂) − ρ f f f f + q f , ∇x · v f compatibility conditions
t t
f f f f
+ w , σ nt − h
f f
f,N + w ,
f
σ nt fs We begin by considering the discretization of the reference
t
t
s s domain 0 . Here, and in what follows, we will use the same
+ w , L (u) − ρ0s f s s + ws , P ns0 − hs
s,N
s 0 0 notation for the discrete objects as for their continuous coun-
+ w , P n0
f s ,
s
(112) terparts to simplify the presentation. Let N̂ A denote a set of
0
basis functions defined on 0 , as in Sect. 2.3, and let I denote
f
where nt and ns0 are the unit outward normal vectors to the the index set of all basis functions defined on 0 . These func-
fluid and solid domains, in the current and reference confi- tions do not depend on time, they are “fixed” in space on the
gurations, respectively. In (112) the following definitions are reference domain. Consider a discrete ALE mapping φ̂( y, r )
used: which can be expressed as
∂v φ̂( y, r ) = φ̂ A (r ) N̂ A ( y) = (Û A (r )+ y A ) N̂ A ( y) (120)
L f (v, p; v̂)=ρ f+ρ f (v − v̂) · ∇x v − ∇x · σ f , (113)
∂r A∈I A∈I
σ f = −∇x p I + 2µ f ∇xs v, (114) The mapping pertains to the entire fluid-structure domain.
∂2u The motion of the fluid subdomain is obtained from (120) by
Ls (u) = ρ0s 2 − ∇ X · P. (115)
∂s restricting the index set to the fluid control variables (or nodal
P = F S, (116) variables in the case of finite elements). We write I = I f ∪ Is ,
where I f and Is are the index sets of the fluid and solid
Standard variational arguments imply that the fluid and the control variables, respectively. Note that I f ∩ Is
= ∅ due to
solid momentum equations and the fluid incompressibility the kinematic continuity conditions imposed at the fluid-solid
constraint hold in the interior of the appropriate subdomains. interface.
The Neumann boundary conditions are also satisfied on the The displacement field of the solid is written as
appropriate parts of the fluid and solid domain boundaries.
Selecting test functions that vanish everywhere in the domain, u(X, s) = U A (s) N̂ A (X), (121)
except at the fluid-solid interface in (112), gives A∈Is
−1
w f , σ f nt
f In contrast to the solid problem, the fluid problem (106)
f s + w ◦ φ̂ , σ s nst f s = 0.
s
(118)
t
t is posed over the current configuration with unknown fields
Using (111) we arrive at the weak continuity of surface trac- expressed as functions of the spatial coordinates x. In order
tions at the fluid-solid interface to approximate the unknown fields in the current domain,
−1
we employ {N A (x, t) = N̂ A ◦ φ̂ t (x)} A∈I f , as defined in
f
w f , σ f nt + σ s nst f s = 0, (119) Sect. 2.3, Eq. (61), to approximate the fluid velocity and
t
123
Comput Mech (2008) 43:3–37 15
123
16 Comput Mech (2008) 43:3–37
The following definitions are employed in (129): We note that selecting the advective form was an impor-
tant constituent in obtaining this result. Furthermore, assu-
v = τ M (L f (v, p; v̂) − ρ f f f ) (131) ming that a time integrator is chosen such that it “respects”
Ct a constant solution, that is, if the velocity field is constant
τM = + (v − v̂) · G(v − v̂)
t 2 in time, the discrete approximation to the time derivative is
f 2 zero, then the formulation satisfies the geometric conserva-
µ
+C I G : G )−1/2 (132) tion law at a fully discrete level. Any consistent time integra-
ρf
tor, including the one employed in this work and described
τC = (τ M g · g)−1 (133) in the following section, should satisfy this condition.
−1/2
τ = (v · Gv ) (134) To demonstrate global mass conservation, we set w f = 0,
d
∂ξk ∂ξk w = 0, wm = 0, and q f = 1 in (128). This choice leaves
s
Gi j = (135) us with
∂ xi ∂ x j
k=1
d (1, ∇x · v) f = (1, v · n f )
f = 0, (140)
t t
G:G= Gi j Gi j (136)
i, j=1 which is precisely the statement of global mass conservation
on the fluid domain. Note that in the solid region the mass-
d
(v − v̂) · G(v − v̂) = (vi − v̂i )G i j (v j − v̂ j ) (137) conservation is satisfied a priori due to the choice of the
i, j=1 Lagrangian description.
Let ei , i = 1, 2, 3, be the ith Cartesian basis vector.
d
∂ξ j
gi = , (138) To show global momentum conservation, we set w f =ei ,
∂ xi ws =ei , wm =0 in (128) and assume that there are no body or
j=1
g · g = gi gi . (139) surface forces present. In this case (128) reduces to
∂ξ ∂2u ∂v
In the above, ∂ x is the inverse Jacobian of the mapping bet- ei , ρ0s 2 + ei , ρ f
∂s s ∂r tf
ween the isoparametric, or parent, and physical domains,
t
0
is the time step size, and C I is a positive constant, inde- + ei , ρ f (v − v̂) · ∇x v f +(ei , v · ∇x v) f = 0 (141)
pendent of the mesh size, derived from an element-wise t t
inverse estimate (see, e.g., Johnson [40]). In (128) the sym- 1
˜ tf is used to denote the fact that integrals are taken over (q f , ∇x · v) f + ∇x q f , f v =0 (142)
bol t ρ t
f
element interiors.
Galerkin’s method is employed for the solid and mesh Integration-by-parts in the third term of (141) gives
motion problems. The fluid formulation (129) emanates from
∂2u ∂v
the variational multiscale residual-based turbulence mode- ei , ρ0s 2 + ei , ρ f
∂s s ∂r tf
ling paradigm [2,4,7,11,29]. The residual-based formulation
0
of fluid flow may be viewed as an extension of well-known + ei , ρ f {(v − v̂) · n f }v f
stabilized methods, such as SUPG [10]. However, the last
t
term of (129) is not motivated by multiscale arguments, but − ei , ρ {∇x · (v − v̂)}v f
f
merely provides additional residual-based stabilization (see t
4.3 Discussion of conservation Using (142) with a specific choice of the test function q f =
ρ f vi gives
In this section we focus on the semi-discrete formulation
(ei , ρ f {∇x · v}v) f + (ei , v · ∇x v) f = 0 (144)
(128) restricted to the fluid. We show that the formulation t t
satisfies the discrete geometric conservation law and is glo- which, in turn, simplifies (143) to
bally mass-conservative. We also show that our formulation 2
globally conserves momentum under semi-discretization. s∂ u f ∂v
ei , ρ0 2 + ei , ρ
The discrete geometric conservation law is satisfied if the ∂s s ∂r tf
formulation preserves a constant fluid velocity in space and
0
time when there are no body forces and the stress tensor + ei , ρ f {(v − v̂) · n f }v f
t
is self-equilibrating. Indeed, if a constant particle velocity
field is assumed, it is easily seen to satisfy (128) identically. + ei , ρ {∇x · v̂}v f = 0
f
(145)
t
123
Comput Mech (2008) 43:3–37 17
Note that q f = ρ f vi is a valid choice because velocity and finite elements are employed. On the other hand, lack of satis-
pressure are approximated by the same discrete spaces. This faction of the identity induced by a time-integration method
would not be the case if Vh = Wh
⊃ Vhv component-wise.
p p
will, in general, prevent us from obtaining (147) from (146)
We proceed by changing variables in the second and fourth and in this case momentum may not be conserved in the
terms of Eq. (145) to the referential description as fully-discrete case.
2
s∂ u f ∂v ˆ
ei , ρ0 2 + ei , ρ J 4.4 Time integration of the fluid-structure interaction
∂s s ∂r 0
f
0
system
+ ei , ρ f {(v − v̂) · n f }v f
t
In this section, we present the time integration algorithm for
+ ei , ρ f { Jˆ∇x · v̂}v f = 0 (146) semi-discrete Eq. (128). The method is the generalized-α
0
algorithm proposed by Chung and Hulbert [12] for the equa-
Recognizing that Jˆ∇x · v̂ = ∂ Jˆ/∂r [see also Eq. (47)] and tions of structural dynamics, and extended to the equations of
combining the second and fourth terms in (146) gives fluid mechanics by Jansen et al. [36]. In the context of fluid-
2
structure interaction, the generalized-α method was used in
s∂ u f ∂v J
ˆ
ei , ρ0 2 + ei , ρ [15,43]. In this section we give details of the method as it
∂s s ∂r f applies to the semi-discrete formulation (128).
0 0
Let U, U̇, Ü, and P denote the vectors of nodal or control
+ ei , ρ f {(v − v̂) · n f }v f = 0 (147) variable degrees of freedom of displacements, velocities,
t
velocity time derivatives, and pressure, respectively, of the
Defining u̇ = ∂ u/∂s to be the velocity of the solid, taking
fluid-structure system. Let V , V̇ , and V̈ denote the vectors of
the partial time derivative outside of the inner product and
nodal or control variable degrees of freedom of mesh displa-
changing variables back to the spatial domain in the first two
cements, velocities, and accelerations, respectively. Note that
terms of (147) gives
there is no “fluid displacement” in our formulation; U̇ and
d
d Ü are simply labels that we chose for fluid time derivative
ei , ρ0s J −1 u̇ s + (ei , ρ f v) f
dt t dt t degrees of freedom. Also note that fluid and solid degrees of
+ ei , ρ f {(v − v̂) · n f }v f = 0 (148) freedom are combined in one solution vector. We define three
t residual vectors corresponding to the momentum, continuity,
By conservation of mass in the Lagrangian description, and mesh motion equations by substituting individual basis
ρ0s J −1 = ρts [see also Eq. (67)]. Also note that v = v̂ on functions in place of w f , ws , q f , and wm in (128) as follows:
fs
simultaneously employed, for example, when space-time ×V n+α f , V̇ n+α f , V̈ n+αm ) = 0, (157)
123
18 Comput Mech (2008) 43:3–37
Rmesh (U n+α f , U̇ n+α f , Ü n+αm , while for the first order system Jansen et al. [36] give
× P n+1 , V n+α f , V̇ n+α f , V̈ n+αm ) = 0, (158)
U n+α f = U n + α f (U n+1 − U n ), (159) j 1 3 − ρ∞
j
αm = , (173)
U̇ n+α f = U̇ n + α f (U̇ n+1 − U̇ n ), (160) 2 1 + ρ∞
j
123
Comput Mech (2008) 43:3–37 19
Predictor stage. Set Solve this linear system using a preconditioned GMRES
algorithm (see Saad and Schultz [54]) to a specified
U̇ n+1,(0) = U̇ n , (177) tolerance.
(γ − 1) (3) Having solved the linear system, update the iterates as
Ü n+1,(0) = Ü n , (178)
γ
t 2 Ü n+1,(l) = Ü n+1,(l−1) +
Ü n+1,(l) (194)
U n+1,(0) = U n +
t U̇ n + ((1 − 2β)Ü n
2 U̇ n+1,(l) = U̇ n+1,(l−1) + γ
t
Ü n+1,(l) (195)
+ 2β Ü n+1,(0) ), (179)
U n+1,(l) = U n+1,(l−1) + β(
t)
Ü n+1,(l)
2
(196)
P n+1,(0) = P n , (180)
P n+1,(l) = P n+1,(l−1) +
P n+1,(l) (197)
V̇ n+1,(0) = V̇ n (181)
V̈ n+1,(l) = V̈ n+1,(l−1) +
V̈ n+1,(l) (198)
(γ − 1)
V̈ n+1,(0) = V̈ n (182) V̇ n+1,(l) = V̇ n+1,(l−1) + γ
t
V̈ n+1,(l) (199)
γ
V n+1,(l) = V n+1,(l−1) + β(
t)
V̈ n+1,(l)
2
(200)
V n+1,(0) = V n +
t V̇ n
t 2
+ ((1 − 2β) V̈ n + 2β V̈ n+1,(0) ). (183) Remark 4.4 In the context of solution strategies employed
2 to solve the coupled fluid-structural equations, Tezduyar and
The subscript 0 on the left-hand side quantities is the ite- co-workers (see, e.g., [71]) introduced the following termi-
ration index. Note that the predictor is consistent with the nology: the block-iterative solution strategy refers to the case
generalized-α Eq. (165)–(168). when the solution for the three fields (i.e., the solid, fluid and
Multi-corrector stage. Repeat the following steps for mesh) is obtained in a fully segregated manner; the quasi-
l = 1, 2, . . . , lmax . direct solution strategy refers to the case when the fluid and
solid equations are solved in a coupled fashion, while the
(1) Evaluate iterates at the intermediate time levels as mesh motion is solved for separately; finally, the direct stra-
tegy refers to the case when the three fields are solved for
in a coupled fashion and all the influences of the fields on
U n+α f ,(l) = U n + α f (U n+1,(l−1) − U n ) (184)
each other are reflected in the tangent matrix. Adopting the
U̇ n+α f ,(l) = U̇ n + α f (U̇ n+1,(l−1) − U̇ n ) (185) terminology of Tezduyar, the method outlined in this section
Ü n+αm ,(l) = Ü n + αm (Ü n+1,(l−1) − Ü n ) (186) may be classified as a direct solution strategy.
V n+α f ,(l) = V n + α f (V n+1,(l−1) − V n ) (187)
V̇ n+α f ,(l) = V̇ n + α f ( V̇ n+1,(l−1) − V̇ n ) (188)
5 Linearization of the fluid-structure interaction
V̈ n+αm ,(l) = V̈ n + αm ( V̈ n+1,(l−1) − V̈ n ) (189) equations: a methodology for computing shape
P n+1,(l) = P n+1,(l−1) (190) derivatives
(2) Use the intermediate solutions to assemble the residuals Derivatives of the momentum, continuity, and mesh motion
of the continuity and momentum equations and the cor- residuals with respect to solution variables define the so-
responding matrices in the linear system called tangent matrices. In particular, derivatives of the
momentum and continuity residuals with respect to the mesh
∂ Rmom ∂ Rmom motion variables are referred to as shape derivatives. The
Ü n+1,(l) +
P n+1,(l) computation of shape derivative matrices, required for a
∂ Ü n+1 ∂ P n+1
consistent linearization of the fluid-structure system, has not
∂ Rmom
+
V̈ n+1,(l) = −Rmom (l) (191) been extensively studied in the fluid-structure interaction lite-
∂ V̈ n+1 rature, although recently a few references have appeared on
∂R con ∂R con
the subject (see, e.g., [15,19]). In this section, we present
Ü n+1,(l) +
P n+1,(l)
∂ Ü n+1 ∂ P n+1 a detailed methodology for deriving shape derivatives and
∂ Rcon provide their explicit expressions. Although presented in the
+
V̈ n+1,(l) = −Rcon(l) (192) ALE context, this methodology is applicable to other fluid-
∂ V̈ n+1
structure formulations.
∂ Rmesh ∂ Rmesh
Ü n+1,(l) +
P n+1,(l) We begin by introducing notation. Let x = x(ξ ) denote
∂ Ü n+1 ∂ P n+1 the isoparametric mapping at a particular time instant. Let ∂ x
∂ξ
∂ Rmesh ∂ξ −1
+
V̈ n+1,(l) = −Rmesh
(l) (193) be the Jacobian of this mapping, let ∂ x = ∂ x denote its
∂ V̈ n+1 ∂ξ
123
20 Comput Mech (2008) 43:3–37
123
Comput Mech (2008) 43:3–37 21
where Ndof is the number of element degrees of freedom. Changing variables back to the physical domain, taking the
The above implies sum over the elements in the fluid mesh, and accounting for
the second term of (218), we get
∂ ∂∂ξxkl ∂ξ ∂ N B ∂ξl
l
= α f β
t 2 . (215)
Nel
∂ V̈B, j ∂ xk ∂ξl ∂ x j ∂vi ∂ N B
α f β
t 2 N A ρ f (vk − v̂k ) de
∂ xk ∂ x j
In (215) we made use of Newmark update formulas (164) and e=1 e
(168). Inserting (215) into (210), changing variables back to
Nel
the physical domain, and summing over the elements of the ∂vi ∂ N B
− α f β
t 2 N A ρ f (vk − v̂k ) de
fluid mesh, we finally get ∂ x j ∂ xk
e=1 e
Nel
∂vi ∂ N B
Nel
∂vi
α f β
t 2 N Aρ f de . (216) − α f γ
t NA ρ f N B de . (220)
∂t ∂ x j ∂x j
e=1 e e=1 e
Matrix (216) is the contribution to the shape derivative matrix
Pressure stabilization term
(205) from the acceleration term present in the discrete
As a final example we present the derivation of the shape
momentum equations of the incompressible Navier–Stokes
derivative contribution from the discrete continuity equation,
system. It is form-identical to and has the same sparsity struc-
that is
ture of the matrices that contribute to the tangents in the ana-
lysis of fluids and solids, and its implementation in finite ∂ R cont
element and isogeometric codes is standard.
A
. (221)
∂ V̈B, j
Advection term
In the acceleration term the coupling between the momen- Consider the pressure contribution from the stabilizing terms,
tum residual and mesh motion variables occurs through Jξ . that is
Other terms of the discrete incompressible Navier–Stokes Nel
system exhibit more complex coupling. For example, consi- ∂ N A τM ∂ p
de . (222)
der the advective contribution to the momentum residual ∂ xi ρ f ∂ xi
e=1
e
Nel
∂vi
N A ρ f (vk − v̂k ) de As in previous developments, restricting to a single element
∂ xk and changing variables in (222) gives
e=1
e
Nel
∂vi N
el
∂vi ∂ N A ∂ξk τ M ∂ p ∂ξl
ˆ e.
Jξ d (223)
= N A ρ f vk de − N A ρ f v̂k de . ∂ξk ∂ xi ρ f ∂ξl ∂ xi
∂ xk ∂ xk
e=1
e
e=1
e
ˆe
(217)
Taking the derivative with respect to the acceleration degrees
Restricting the sum to a single element, changing variables of freedom and isolating terms that are independent of the
to the parent domain, and taking the derivative with respect mesh motion in (223) we get
to the mesh acceleration degrees of freedom gives
∂ τ ∂ξk ∂ξl
J
∂ NA 1 ∂p M ∂ xi ∂ xi ξ
∂v ∂ ∂∂ξxkl Jξ ˆe
d (224)
N A ρ f (vk − v̂k )
i ˆe
d ∂ξk ρ f ∂ξl ∂ V̈B, j
∂ξl ∂ V̈B, j ˆe
ˆe
∂ξk ∂ξl
∂vi ∂ξl ∂ v̂k ∂ N A τM ∂ p ∂ ∂ xi ∂ xi Jξ
− NA ρ f ˆe
Jξ d (218) = ˆe
d
∂ξl ∂ xk ∂ V̈B, j ∂ξk ρ f ∂ξl ∂ V̈B, j
ˆe ˆe
∂ N A 1 ∂ p ∂ξk ∂ξl ∂τ M
Using relation (203) in the first term of (218) gives + Jξ ˆ e.
d
⎛
∂ξk ρ ∂ξl ∂ xi ∂ xi ∂ V̈B, j
f
∂ ∂ xn ˆe
∂vi ⎝ ∂ξm ∂ξm ∂ξl
N A ρ (vk − v̂k )
f
∂ξl ∂ xn ∂ V̈B, j ∂ xk The last term on the right-hand side of the above expression
ˆe
involves the derivative of τ M with respect to the mesh acce-
⎞
∂ xn leration degrees of freedom. It is, in principle, present in the
∂ξl ∂ ∂ξm ∂ξm
− ⎠ Jξ d
ˆe (219) tangent matrix and is computable, but in this work it is omit-
∂ xn ∂ V̈B, j ∂ xk ted. In order to handle the first term on the right-hand side of
123
22 Comput Mech (2008) 43:3–37
123
Comput Mech (2008) 43:3–37 23
123
24 Comput Mech (2008) 43:3–37
E m = E 0m Jξ−1 , (240)
123
Comput Mech (2008) 43:3–37 25
123
26 Comput Mech (2008) 43:3–37
123
Comput Mech (2008) 43:3–37 27
123
28 Comput Mech (2008) 43:3–37
Flowrate (cm3/s)
volume (circles). (b) Plot of the
0.4
relative error in the flow rates
that is attributable to
convergence tolerances
employed in the calculations 0.2
0 4 8 12 16 20 24 28
Time (s)
(b) x 10−4
10
8
(dVf /dt−Oin)/max(Q in)
0 4 8 12 16 20 24 28
Time(s)
123
Comput Mech (2008) 43:3–37 29
⎡ ⎤
σ 0 0
σ = ⎣ 0 0 0⎦ (243)
0 0 0
and, by the transformation of stress formula (71),
⎡ −2 ⎤
J σ F11 0 0
S = J F −1 σ F −T = ⎣ 0 0 0⎦
0 0 0
⎡ −1 ⎤
J σ C11 0 0
=⎣ 0 0 0⎦, (244)
0 0 0
123
30 Comput Mech (2008) 43:3–37
to evaluate the effective tangent modulus. Note that the equi- Elastic energy at
300 physiological stress due
valent tangent stiffness is higher at the level of deformation to intramural pressure
caused by the physiological stress as compared to the tangent
200
stiffness in the undeformed configuration. Also note that, as
expected, the value of the tangent stiffness in the undeformed
91.3
configuration is exactly the value of Young’s modulus used to
42.3
define the solid model parameters. Figure 18 shows the plot 0
of elastic energy ψ versus strain . We believe that this ana- 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45
Strain: (lx−l 0x)/l x
lysis justifies using the model in fluid-structure interaction
analysis of arteries. We also feel that it supports the hypo- Fig. 18 Elastic energy plotted against strain for the material model
theses behind the coupled momentum method of Figueroa et used in the abdominal aorta simulation. Physiological range of energies
al. [20]. Obviously, significant improvements are possible. and strains due to intramural pressure are indicated by dashed lines
123
Comput Mech (2008) 43:3–37 31
Fig. 19 Flow in a
patient-specific abdominal aorta
with aneurysm.
a Patient-specific imaging data;
b Skeleton of the NURBS mesh;
c Smoothed and truncated
NURBS model and mesh. In c,
every NURBS patch is assigned
a different color. For more
details of geometrical modeling
for isogeometric analysis of
blood flow the reader is referred
to [74]
123
32 Comput Mech (2008) 43:3–37
After a nearly periodic-in-time solution was attained, Fig. 20 Flow in a patient-specific abdominal aorta with aneurysm.
simulation data was collected for postprocessing. Figures 20 Large frame: fluid velocity vectors colored by their magnitude in the
and 21 show snapshots of the velocity field plotted on the vicinity of the celiac and superior mesenteric arteries. Left small frame:
volume rendering of the velocity magnitude. Right small frame: fluid
moving domain at various times during the heart cycle. The velocity isosurfaces
flow field is quite complex and fully three-dimensional, espe-
cially in diastole. Velocity magnitude is largest near the inflow
and
and is significantly lower in the aneurysm region. This occurs
in part due to the fact that a significant percentage of the flow T
1
goes to the upper branches of the abdominal aorta. There τabs = |τ s |dt. (260)
T
is also an increase in the cross-sectional area of the vessel 0
associated with the aneurysm, resulting in a decreased flow Note that, in contrast to the wall shear stress, OSI is largest
velocity in this region. in the aneurysm region, especially along the posterior wall,
Figure 22 shows the magnitude of the wall shear stress suggesting that wall shear stress is highly oscillatory there
(i.e., the tangential component of the fluid or solid traction due to the recirculating flow. Low time-averaged wall shear
vector) averaged over a cycle. Note that the wall shear stress stress, in combination with high shear stress temporal oscil-
is significantly lower in the aneurysm region than in the upper lations, as measured by the OSI, are indicators of the regions
portion of the aorta and at branches. Also note the smoothness of high probability of occurrence of atherosclerotic disease.
of the stress contours due to the NURBS representation of Figure 24 shows the distribution of flow among the branches.
the geometry and solution fields. The outflow lags the inflow due to the distensibility of the
Figure 23 shows the oscillatory shear index (OSI) at the arterial wall. The overall flow distribution and the time lag
luminal surface. OSI is defined as (see, e.g., [62,63]), are in qualitative agreement with results in [3,20].
1 τmean
OSI = 1− , (258)
2 τabs 9 Conclusions
where, denoting by τ s the wall shear stress vector,
We have presented a monolithic isogeometric fluid-
T structure interaction formulation. The fluid subdomain was
1
τmean = τ s dt , (259) assumed to be governed by the incompressible Navier–Stokes
T equations and the solid subdomain was assumed to be
0
123
Comput Mech (2008) 43:3–37 33
123
34 Comput Mech (2008) 43:3–37
Fig. 22 Flow in a
patient-specific abdominal aorta
with aneurysm. Magnitude of
wall shear stress averaged over a
cycle plotted in three different
views
Fig. 23 Flow in a
patient-specific abdominal aorta
with aneurysm. Oscillatory
shear index (OSI) plotted in
three different views
With these preliminaries we have In the above we have utilized the commutativity of pull-
back and exterior differentiation, and the fact that the exte-
∇ y(4) · γ (4)
y d Q y = Lγ (4) d Q y
y rior derivative of a 4-form, closed in four dimensions, is
(4) (4)
= d(i γ (4) d Q y ) zero. That Lγ (4) d Q y = ∇ y · γ y d Q y can be proved by
y y
(4) ∗ assuming cartesian coordinates on R4 ⊃ Q y and d Q y =
= d(φ̂ i γ (4) d Q x )
x dy0 ∧ dy1 ∧ dy2 ∧ dy3 , where ∧ is the wedge product. Like-
(4) ∗ wise, the same construct can be used to prove Lγ (4) d Q x =
= φ̂ d(i γ (4) d Q x ) x
x
(4) (4)
(4) ∗ ∇x ·γ x d Q x . Curvilinear spatial coordinates can be accom-
= φ̂ Lγ (4) d Q x modated by using the usual tensor transformation formulas
x
(4) ∗
= φ̂ (∇x · γ (4)
(4)
x d Qx )
on these coordinates (see Marsden and Hughes [47]). The
(4) (4) (4) (4) ∗ time coordinate is unaffected in this case.
= ∇x · γ x ◦ φ̂ φ̂ d Qx
(4) (4) Remark A.1 With differential forms, correct boundary inte-
= ∇x(4) · γ (4)
x ◦ φ̂ Jˆ d Qy. (A.3)
gral expressions can be obtained without the necessity of unit
This results and the change of variables formula yields (9). normal vectors by using Stokes theorem and the change of
123
Comput Mech (2008) 43:3–37 35
Fig. 24 Flow in a
patient-specific abdominal aorta
with aneurysm. Flow rate
distribution among the arterial
branches during a heart cycle
variables formula. For example, d Q x and d Q y , respectively, in keeping with the standard
usage.
i γ (4) d Q y = d(i γ (4) d Q y )
y y
∂ Qy Qy
References
(4) ∗
= φ̂ d(i γ (4) d Q x )
x 1. Abedi P, Patracovici B, Haber RB (2006) A spacetime discon-
Qy tinuous Galerkin method for linearized elastodynamics with
element-wise momentum balance. Comput Methods Appl Mech
= d(i γ (4) d Q x ) Eng 195:3247–3273
x
(4)
2. Akkerman I, Bazilevs Y, Calo VM, Hughes TJR, Hulshoff
φ̂ (Q y ) S (2008) The role of continuity in residual-based variational mul-
tiscale modeling of turbulence. Comput Mech 41:371–378
= d(i γ (4) d Q x ) 3. Bazilevs Y, Calo VM, Zhang Y, Hughes TJR (2006) Isogeometric
x fluid-structure interaction analysis with applications to arterial
Qx
blood flow. Comput Mech 38:310–322
4. Bazilevs Y, Calo VM, Cottrell JA, Hughes TJR, Reali A, Scovazzi
= i γ (4) d Q x . (A.4) G (2007) Variational multiscale residual-based turbulence mode-
x
∂ Qx ling for large eddy simulation of incompressible flows. Comput
Methods Appl Mech Eng 197:173–201
5. Bazilevs Y, Beiraoda Veiga L, Cottrell JA, Hughes TJR, Sangalli
Remark A.2 The notations d Q x and d Q y are used under the
G (2006) Isogeometric analysis: approximation, stability and error
integration sign in the main body of the paper, in contrast with estimates for h-refined meshes. Math Models Methods Appl Sci
the differential form notations to represent volume elements, 16:1031–1090
123
36 Comput Mech (2008) 43:3–37
6. Bazilevs Y, Hughes TJR (2007) Weak imposition of Dirichlet 27. Holzapfel GA (2000) Nonlinear solid mechanics, a continuum
boundary conditions in fluid mechanics. Comput Fluids 36:12–26 approach for engineering. Wiley, Chichester
7. Bazilevs Y, Michler C, Calo VM, Hughes TJR (2007) Weak Diri- 28. Hughes TJR (2000) The finite element method: linear static and
chlet boundary conditions for wall-bounded turbulent flows. Com- dynamic finite element analysis. Dover Publications, Mineola
put Methods Appl Mech Eng 196:4853–4862 29. Hughes TJR, Calo VM, Scovazzi G (2004) Variational and multis-
8. Bishop R, Goldberg S (1980) Tensor analysis on manifolds. Dover, cale methods in turbulence. In: Gutkowski W, Kowalewski TA (eds)
New York In Proceedings of the XXI International Congress of Theoretical
9. Brezzi F, Fortin M (1991) Mixed and hybrid finite element and Applied Mechanics (IUTAM), Kluwer, pp 153–163
methods. Springer, Berlin 30. Hughes TJR, Cottrell JA, Bazilevs Y (2005) Isogeometric analysis:
10. Brooks AN, Hughes TJR (1982) Streamline upwind/Petrov- CAD, finite elements, NURBS, exact geometry, and mesh refine-
Galerkin formulations for convection dominated flows with par- ment. Comput. Methods Appl Mech Eng 194:4135–4195
ticular emphasis on the incompressible Navier–Stokes equations. 31. Hughes TJR, Hulbert GM (1988) Space-time finite element
Comput Methods Appl Mech Eng 32:199–259 methods for elastodynamics: formulations and error estimates.
11. Calo VM (2004) Residual-based multiscale turbulence mode- Comput Methods Appl Mech Eng 66:339–363
ling: finite volume simulation of bypass transistion. Ph.D. The- 32. Hughes TJR, Liu WK, Zimmermann TK (1981) Lagrangian–
sis, Department of Civil and Environmental Engineering, Stanford Eulerian finite element formulation for incompressible viscous
University flows. Comput Methods Appl Mech Eng 29:329–349
12. Chung J, Hulbert GM (1993) A time integration algorithm for 33. Hughes TJR, Wells GN (2005) Conservation properties for the
structural dynamics with improved numerical dissipation: The Galerkin and stabilised forms of the advection-diffusion and
generalized-α method. J Appl Mech 60:371–75 incompressible Navier-Stokes equations. Comput Methods Appl
13. Cottrell JA, Reali A, Bazilevs Y, Hughes TJR (2006) Isogeometric Mech Eng 194:1141–1159
analysis of structural vibrations. Comput Methods Appl Mech Eng 34. Hulbert GM, Hughes TJR (1990) Space-time finite element
195:5257–5297 methods for second order hyperbolic equations. Comput Methods
14. Cottrell JA, Reali A, Hughes TJR (2007) Studies of refinement and Appl Mech Eng 84:327–348
continuity in isogeometric structural analysis. Comput Methods 35. Idelsohn SR, Oñate E, Del Pin F, Calvo N (2006) Fluid-structure
Appl Mech Eng 196:4160–4183 interaction using the particle finite element method. Comput
15. Dettmer W, Perić D (2006) A computational framework for fluid- Methods Appl Mech Eng 195:2100–2123
structure interaction: finite element formulation and applications. 36. Jansen KE, Whiting CH, Hulbert GM (1999) A generalized-α
Comput Methods Appl Mech Eng 195:5754–5779 method for integrating the filtered Navier-Stokes equations with
16. Elguedj T, Bazilevs Y, Calo VM, Hughes TJR (2008) B-bar and a stabilized finite element method. Comput Methods Appl Mech
F-bar projection methods for nearly incompressible linear and non- Eng 190:305–319
linear elasticity and plasticity using higher-order NURBS elements. 37. Johnson AA, Tezduyar TE (1994) Mesh update strategies in paral-
Comput Methods Appl Mech Eng 197:2732–2762 lel finite element computations of flow problems with moving
17. Farhat C, Geuzaine P, Grandmont C (2001) The discrete geometric boundaries and interfaces. Comput Methods Appl Mech Eng
conservation law and the nonlinear stability of ALE schemes for 119:73–94
the solution of flow problems on moving grids. J Comput Phys 38. Johnson AA, Tezduyar TE (1997) Parallel computation of incom-
174(2):669–694 pressible flows with complex geometries. Int J Numer Methods
18. Farhat C, van der Zee K, Geuzaine P (2006) Provably second-order Fluids 24:1321–1340
time-accurate loosely-coupled solution algorithms for transient 39. Johnson AA, Tezduyar TE (1999) Advanced mesh generation and
nonlinear computational aeroelasticity. Comput Methods Appl update methods for 3D flow simulations. Comput Mech 23:130–
Mech Eng 195:1973–2001 141
19. Fernandez MA, Moubachir M (2005) A Newton method using 40. Johnson C (1987) Numerical solution of partial differential equa-
exact jacobians for solving fluid-structure coupling. Comput Struct tions by the finite element method. Cambridge University Press,
83:127–142 Sweden
20. Figueroa A, Vignon-Clementel IE, Jansen KE, Hughes TJR, Taylor 41. Johnson C, Nävert U, Pitkäranta J (1984) Finite element methods
CA (2006) A coupled momentum method for modeling blood flow for linear hyperbolic problems. Comput Methods Appl Mech Eng
in three-dimensional deformable arteries. Comput Methods Appl 45:285–312
Mech Eng 195:5685–5706 42. Küttler U, Förster C, Wall WA (2006) A solution for the incom-
21. Flanders H (1963) Differential forms with applications to the phy- pressibility dilemma in partitioned fluid-structure interaction with
sical sciences. Academic Press, London pure Dirichlet fluid domains. Comput Mech 38:417–429
22. Formaggia L, Nobile F (2005) Stability analysis of second-order 43. Kuhl E, Hulshoff S, de Borst R (2003) An arbitrary Lagrangian
time accurate schemes for ALE-FEM. Comput Methods Appl Eulerian finite element approach for fluid-structure interaction phe-
Mech Eng 193:4097–4116 nomena. Int J Numer Methods Eng 57:117–142
23. Guillermin V, Pollack A (1974) Differential topology. Prentice- 44. Lang S (1972) Differential manifolds. Addison-Wesley, Reading
Hall, Englewood Cliffs 45. Lang S (1995) Differential and riemannian manifolds (graduate
24. Heil M, Hazel A, Boyle J (2008) Solvers for large-displacement texts in mathematics, vol. 160). Springer, Heidelberg
fluid-structure interaction problems: segregated vs. monolithic 46. Le Tallec P, Mouro J (2001) Fluid structure interaction with
approaches. Comput Mech. doi:10.1007/s00466-008-0270-6 large structural displacements. Comput Methods Appl Mech Eng
25. Heywood JG, Rannacher R, Turek S (1996) Artificial boundaries 190:3039–3068
and flux and pressure conditions for the incompressible Navier- 47. Marsden JE, Hughes TJR (1993) Mathematical foundations of
Stokes equations. Int J Numer Methods Fluids 22:325–352 elasticity. Dover Publications Inc., New York
26. Holzapfel GA (2004) Computational biomechanics of soft biologi- 48. Masud A, Hughes TJR (1997) A space-time Galerkin/least-
cal tissue. In: Stein E, De Borst R, Hughes TJR (eds) Encyclopedia squares finite element formulation of the Navier-Stokes equations
of computational mechanics, Solids and structures, chap 18, vol 2. for moving domain problems. Comput Methods Appl Mech Eng
Wiley, London 148:91–126
123
Comput Mech (2008) 43:3–37 37
49. Michler C, van Brummelen EH, de Borst R (2006) Error- 64. Tezduyar TE, Behr M, Liou J (1992) A new strategy for finite
amplification analysis of subiteration-preconditioned GMRES for element computations involving moving boundaries and inter-
fluid-structure interaction. Comput Methods Appl Mech Eng faces. The deforming-spatial-domain/space-time procedure. I. The
195:2124–2148 concept and the preliminary numerical tests. Comput Methods
50. Michler C, van Brummelen EH, Hulshoff SJ, de Borst Appl Mech Eng 94:339–351
R (2003) The relevance of conservation for stability and accu- 65. Tezduyar TE, Behr M, Mittal S, Liou J (1992) A new strategy
racy of numerical methods for fluid-structure interaction. Comput for finite element computations involving moving boundaries and
Methods Appl Mech Eng 192:4195–4215 interfaces. The deforming-spatial-domain/space-time procedure.
51. Nobile F (2001) Numerical Approximation of Fluid-Structure II. Computation of free-surface flows, two-liquid flows, and flows
Interaction Problems with Application to Haemodynamics. Ph.D. with drifting cylinders. Comput Methods Appl Mech Eng 94:353–
Thesis, EPFL 371
52. Piperno S, Farhat C (2001) Partitined procedures for the transient 66. Tezduyar TE, Sathe S (2007) Modelling of fluid-structure interac-
solution of coupled aeroelastic problems. Part II: Energy trans- tions with the space-time finite elements: solution techniques. Int
fer analysis and three-dimensional applications. Comput Methods J Numer Methods Fluids 54:855–900
Appl Mech Eng 190:3147–3170 67. Tezduyar TE, Sathe S, Cragin T, Nanna B, Conklin BS, Pause-
53. Piperno S, Farhat C, Larrouturou B (1995) Partitined procedures wang J, Schwaab M (2007) Modelling of fluid-structure interac-
for the transient solution of coupled aeroelastic problems. Part I: tions with the space-time finite elements: Arterial fluid mechanics.
Model problem, theory and two-dimensional application. Comput Int J Numer Methods Fluids 54:901–922
Methods Appl Mech Eng 124:79–112 68. Tezduyar TE, Sathe S, Schwaab M, Conklin BS (2008) Arterial
54. Saad Y, Schultz MH (1986) GMRES: A generalized minimal resi- fluid mechanics modeling with the stabilized space-time fluid-
dual algorithm for solving nonsymmetric linear systems. SIAM J structure interaction technique. Int J Numer Methods Fluids
Sci Stat Comput 7:856–869 57:601–629
55. Simo JC, Hughes TJR (1998) Computational inelasticity. Springer, 69. Tezduyar TE (2003) Computation of moving boundaries and inter-
New York faces and stabilization parameters. Int J Numer Methods Fluids
56. Babus̆ka I (1973) The finite element method with Lagrange multi- 43:555–575
pliers. Numer Math 20:179–192 70. Tezduyar TE, Behr M, Mittal S, Johnson AA (1992) Computation
57. Spivak M (1965) Calculus on manifolds. Benjamin, New York of unsteady incompressible flows with the stabilized finite element
58. Stein K, Tezduyar T, Benney R (2003) Mesh moving techniques methods—space–time formulations, iterative strategies and massi-
for fluid-structure interactions with large displacements. J Appl vely parallel implementations. In: New methods in transient ana-
Mech 70:58–63 lysis, pVP, vol 246/AMD, vol 143. ASME, New York, pp 7–24
59. Stein K, Tezduyar TE, Benney R (2004) Automatic mesh update 71. Tezduyar TE, Sathe S, Keedy R, Stein K (2006) Space-time finite
with the solid-extension mesh moving technique. Comput Methods element techniques for computation of fluid-structure interactions.
Appl Mech Eng 193:2019–2032 Comput Methods Appl Mech Eng 195:2002–2027
60. Texas Advanced Computing Center (TACC). http://www.tacc. 72. Vignon-Clementel IE, Figueroa CA, Jansen KE, Taylor
utexas.edu CA (2006) Outflow boundary conditions for three-dimensional
61. Taylor CA, Hughes TJR, Zarins CK (1998) Finite element mode- finite element modeling of blood flow and pressure in arteries.
ling of blood flow in arteries. Comput Methods Appl Mech Eng Comput Methods Appl Mech Eng 195:3776–3796
158:155–196 73. Wall W (1999) Fluid-Struktur-Interaktion mit stabilisierten Finiten
62. Taylor CA, Hughes TJR, Zarins CK (1998) Finite element mode- Elementen. Ph.D. Thesis, Institut für Baustatik, Universität Stutt-
ling of three-dimensional pulsatile flow in the abdominal aorta: gart
relevance to atherosclerosis. Ann Biomed Eng 26:975–987 74. Zhang Y, Bazilevs Y, Goswami S, Bajaj C, Hughes
63. Taylor CA, Hughes TJR, Zarins CK (1999) Effect of exercise on TJR (2007) Patient-specific vascular NURBS modeling for
hemodynamic conditions in the abdominal aorta. J Vasc Surg isogeometric analysis of blood flow. Comput Methods Appl Mech
29:1077–1089 Eng 196:2943–2959
123