immersed finite element method
TRANSCRIPT
Comput. Methods Appl. Mech. Engrg. 193 (2004) 2051–2067
www.elsevier.com/locate/cma
Immersed finite element method q
Lucy Zhang a, Axel Gerstenberger a, Xiaodong Wang b, Wing Kam Liu a,*
a Department of Mechanical Engineering, The Technological Institute, Northwestern University, 2145 Sheridan Road,
Evanston, IL 60208, USAb Department of Mechanical Engineering, Six MetroTech Center, Polytechnic University, Brooklyn, NY 11201, USA
Received 15 April 2003; received in revised form 10 September 2003; accepted 8 December 2003
Abstract
In this paper, the immersed finite element method (IFEM) is proposed for the solution of complex fluid and
deformable structure interaction problems encountered in many physical models. In IFEM, a Lagrangian solid mesh
moves on top of a background Eulerian fluid mesh which spans over the entire computational domain. Hence, the mesh
generation is greatly simplified. Moreover, both fluid and solid domains are modeled with the finite element methods
and the continuity between the fluid and solid sub-domains are enforced via the interpolation of the velocities and the
distribution of the forces with the reproducing kernel particle method (RKPM) delta function. In comparison with the
immersed boundary (IB) method, the higher-ordered RKPM delta function enables the fluid domain to have non-
uniform spatial meshes with arbitrary geometries and boundary conditions. The use of such kernel functions may
eventually open doors to multi-scale and multi-resolution modelings of complex fluid–structure interaction problems.
Rigid and deformable spheres dropping in channels are simulated to demonstrate the unique capabilities of the pro-
posed method. The results compare well with the experimental data. To the authors� knowledge, these are the first
solutions that deal with particulate flows with very flexible solids.
� 2004 Elsevier B.V. All rights reserved.
Keywords: Immersed finite element method; Reproducing kernel particle method; Fluid–structure interaction; Particulate flow;
Immersed boundary method
1. Introduction
For the past few decades, tremendous research efforts have been directed to the development of mod-elling and simulation approaches for fluid–structure interaction problems. Methods [11,12] developed by
Tezduyar and his coworkers are widely used in the simulations of fluid–particle (rigid) and fluid–structure
interactions. To accommodate the complicated motions of fluid–structure boundaries, we often use
qThis paper is an extended version of the paper presented in the Sixth Japan–US International Symposium on Flow Simulation and
Modeling (May 29–31, 2002, Fukuoka, Japan).* Corresponding author. Tel.: +1-847-491-7094; fax: +1-847-491-3915.
E-mail address: [email protected] (W.K. Liu).
0045-7825/$ - see front matter � 2004 Elsevier B.V. All rights reserved.
doi:10.1016/j.cma.2003.12.044
2052 L. Zhang et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 2051–2067
adaptive meshing or the arbitrary Lagrangian Eulerian (ALE) techniques [6,14,17,18,20]. Recently, suchapproaches have also been adopted by Hu et al. in the modelling of fluid–particle (rigid) systems [5].
Nevertheless, mesh update or remeshing algorithms can be time consuming and expensive and a detailed
discussion on this issue is presented in Ref. [7]. In 1999, Glowinski et al. developed the distributed Lagrange
multiplier (DLM) method for the modelling of viscous flows interacting with rigid particles [4]. Their finite
element formulations enable the use of a fixed structured fluid grid along with fast and efficient solvers.
However, none of the aforementioned methods deals with the interaction between fluid and flexible solid
particles with large deformations.
In the 1970�s, Peskin developed the immersed boundary (IB) method [2] to study flow patterns aroundheart valves. The mathematical formulation of the IB method employs a mixture of Eulerian and
Lagrangian descriptions for fluid and solid domains. In particular, the entire fluid domain is represented by
a uniform background grid, which can be solved by finite difference methods with periodic boundary
conditions; whereas the submerged structure is represented by a fiber network. The interaction between
fluid and structure is accomplished by distributing the nodal forces and interpolating the velocities between
Eulerian and Lagrangian domains through a smoothed approximation of the Dirac delta function. The
advantage of the IB method is that the fluid–structure interface is tracked automatically, which removes the
costly computations due to various mesh update algorithms.Nevertheless, one major disadvantage of the IB method is the assumption of the fiber-like one-dimen-
sional immersed structure, which may carry mass, but occupies no volume in the fluid domain. This
assumption also limits accurate representations for immersed flexible solids which may occupy finite vol-
umes within the fluid domain. Furthermore, uniform fluid grids also set limitations in resolving fluid do-
mains with complex shapes and boundary conditions.
In the proposed immersed finite element method (IFEM), we will eliminate the aforementioned draw-
backs of the IB method and adopt parts of the work on the extended immersed boundary method (EIBM)
recently developed by Wang and Liu [16]. With finite element formulations for both fluid and solid do-mains, the submerged structure is solved more realistically and accurately in comparison with the corre-
sponding fiber network representation. The fluid solver is based on a stabilized equal-order finite element
formulation applicable to problems involving moving boundaries [11–13]. This stabilized formulation
prevents numerical oscillations without introducing excessive numerical dissipations. Moreover, in the
proposed IFEM, the background fluid mesh does not have to follow the motion of the flexible fluid–
structure interfaces and thus it is possible to assign sufficiently refined fluid mesh within the region around
which the immersed deformable structures in general occupy and move.
Unlike the Dirac delta functions in the IB method which yield C1 continuity [3,15], the discretized deltafunction in IFEM is the Cn shape function often employed in the meshfree reproducing kernel particle
method (RKPM). Because of the higher-order smoothness in the RKPM delta function, the accuracy is
increased in the coupling procedures between fluid and solid domains [16]. Furthermore, the RKPM shape
function is also capable of handling nonuniform fluid grids.
We first give a detailed description of the IFEM formulations in Section 2. In Sections 3.1 and 3.2, we
discuss the derivations of the governing equations and their discretizations in fluid and solid domains. The
RKPM delta function for nonuniform spacing and its discretized form are presented in Section 4, together
with the fluid–structure interpolation and distribution procedures. Section 5 summarizes the discretizedgoverning equations and provides an outline of the IFEM algorithm. Several numerical examples of
particulate flows are shown in Section 6. To the authors� knowledge, these are the first solutions that deal
with particulate flows with a large number of moving flexible solids under large deformations. In Section 7,
we summarize the IFEM concept and its attributes.
L. Zhang et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 2051–2067 2053
2. Formulations of immersed finite element method
2.1. Nomenclature and kinematics
Let us consider an incompressible three-dimensional deformable structure in Xs completely immersed in
an incompressible fluid domain Xf . Together, the fluid and the solid occupy a domain X, but they do not
intersect:
Xf [ Xs ¼ X; ð1aÞ
Xf \ Xs ¼ ;: ð1bÞIn contrast to the IB formulation, the solid domain can occupy a finite volume in the fluid domain. Since
we assume both fluid and solid to be incompressible, the union of two domains can be treated as one
incompressible continuum with a continuous velocity field. In the computation, the fluid spans over the
entire domain X, thus an Eulerian fluid mesh is adopted; whereas a Lagrangian solid mesh is constructed on
top of the Eulerian fluid mesh. The coexistence of fluid and solid in Xs requires some considerations when
developing the momentum and continuity equations.
As shown in Fig. 1, in the computational fluid domain X, the fluid grid is represented by the time-
invariant position vector x; while the material points of the structure in the initial solid domain Xs0 and the
current solid domain Xs are represented by Xs and xsðXs; tÞ, respectively. The superscript s is used in the
solid variables to distinguish the fluid and solid domains.
In the fluid calculations, the velocity v and the pressure p are the unknown fluid field variables; whereas
the solid domain involves the calculation of the nodal displacement us, which is defined as the difference of
the current and the initial coordinates: us ¼ xs � Xs. The velocity vs is the material derivative of the dis-
placement dus=dt. The definitions of the spatial coordinate, displacement, velocity, and acceleration in both
fluid and solid domains are summarized in Table 1. Since displacement and acceleration are not explicitly
used in the fluid calculations, they are not defined under the column of the fluid domain.
2.2. The equations of motion
In this section, we start by introducing the equations of motion for a continuum that contains both fluid
and solid. The governing equations of both domains are then extracted from the equations of motion along
Fig. 1. The Eulerian coordinates in the computational fluid domain X are described with the time-invariant position vector x; the solid
positions in the initial configuration Xs0 and the current configuration Xs are represented by Xs and xsðXs; tÞ, respectively. The dis-
placement of the solid is defined as us ¼ xs � Xs.
Table 1
IFEM nomenclature
Fluid domain ðXÞ Solid domain ðXsÞSpatial coordinate x xs
Displacement – us ¼ xs � Xs
Velocity v vs ¼ dus=dt ¼ _us
Acceleration – as ¼ d2us=dt2 ¼ €us
2054 L. Zhang et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 2051–2067
with a delta function for the coupling between fluid and solid domains. The detailed mathematical deri-vation and the proof of equivalency between strong and weak forms will be illustrated in a separate paper.
In a continuum, the inertial force of a particle is balanced with the derivative of the Cauchy stress r and
the external force fext exerted on the continuum:
qdvidt
¼ rij;j þ f exti : ð2Þ
If the solid density qs is different from the fluid density qf , i.e. q ¼ qs in Xs and q ¼ qf in Xf , we can divide
the inertial force into two components within the solid domain Xs:
qdvidt
¼qf
dvidt
; x 2 X=Xs;
qfdvidt
þ ðqs � qfÞ dvidt
; x 2 Xs:
8><>: ð3Þ
Note that to illustrate the point of addition and subtraction of sub-domains, we represent the fluid domain
Xf as the entire domain X minus the solid domain Xs, namely, X=Xs. Moreover, in this paper, the externalforce f ext
i is illustrated as the gravitational force and likewise, we obtain
f exti ¼ qfgi; x 2 X=Xs;
qfgi þ ðqs � qfÞgi; x 2 Xs:
�ð4Þ
Since the computational fluid domain is the entire domain X, we ignore the hydrostatic pressure and rewrite
Eq. (4) as
f exti ¼ 0; x 2 X=Xs;
ðqs � qfÞgi; x 2 Xs:
�ð5Þ
Furthermore, the derivative of the Cauchy stress in Eq. (2) can also be decomposed as:
rij;j ¼rfij;j; x 2 X=Xs;
rfij;j þ rs
ij;j � rfij;j; x 2 Xs:
(ð6Þ
Note that in Eq. (6) the fluid stress in the solid domain in general is much smaller than the corresponding
solid stress [8]. Let us now define the fluid–structure interaction force within the domain Xs as f FSI;si , where
FSI stands for fluid–structure interaction:
f FSI;si ¼ �ðqs � qfÞ dvi
dtþ rs
ij;j � rfij;j þ ðqs � qfÞgi; x 2 Xs: ð7Þ
Naturally, the interaction force f FSI;si in Eq. (7) is calculated with the Lagrangian description. Moreover, a
Dirac delta function d is used to distribute the interaction force from the solid domain onto the compu-tational fluid domain:
L. Zhang et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 2051–2067 2055
f FSIi ðx; tÞ ¼
ZXs
f FSI;si ðXs; tÞdðx� xsðXs; tÞÞdX: ð8Þ
Hence, the governing equation for the fluid can be derived by combining the fluid terms and the interaction
force as:
qf dvidt
¼ rfij;j þ f FSI
i ; x 2 X: ð9Þ
Since we consider the entire domain X to be incompressible, we only need to apply the incompressibility
constraint once in the entire domain X:
vi;i ¼ 0: ð10ÞTo delineate the Lagrangian description for the solid and the Eulerian description for the fluid, we
introduce different velocity field variables vsi and vi to represent the motions of the solid in the domain Xs
and the fluid within the entire domain X. The coupling of both velocity fields is accomplished with the Dirac
delta function:
vsi ðXs; tÞ ¼ZXviðx; tÞdðx� xsðXs; tÞÞdX: ð11Þ
Finally, with the variables in the solid domain Xs defined with the Lagrangian description and the fluid
domain X with the Eulerian description, we can summarize the derivations of the governing equations as
well as the interpolation and distribution processes as follows:
qf dvidt
¼ rfij;j þ f FSI
i ; in X; ð12aÞ
vi;i ¼ 0; in X; ð12bÞ
f FSIi ðx; tÞ ¼
ZXs
f FSI;si ðXs; tÞdðx� xsðXs; tÞÞdX; ð12cÞ
f FSI;si ¼ �ðqs � qfÞ dv
si
dtþ rs
ij;j � rfij;j þ ðqs � qfÞgi; in Xs; ð12dÞ
vsi ðXs; tÞ ¼
ZXviðx; tÞdðx� xsðXs; tÞÞdX; ð12eÞ
where the velocity in the solid domain is simply denoted as vs instead of v.
3. Governing equations and discretizations
In this section, we discuss in detail the strong and weak forms of governing equations for both fluid and
solid domains. In particular, using the kinematic and dynamic matching along the fluid–structure interface,
for brevity, we exclude the fluid–structure interface boundary integral terms in the corresponding weak
forms for both fluid and solid domains.
3.1. Fluid dynamics
This section explains in detail the 3D finite element fluid solver used in the IFEM formulation. Thederivations include the strong form, namely, the Navier–Stokes equations, the Petrov–Galerkin weak form,
and the corresponding discretization.
2056 L. Zhang et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 2051–2067
3.1.1. Strong form
For the Eulerian description of the fluid, the material time derivative dvi=dt is expressed as the time
derivative and the convective terms:
dvidt
¼ vi;t þ vjvi;j: ð13Þ
By substituting Eq. (13) into Eq. (12a), the governing equations for the fluid domain X are written as
qfðvi;t þ vjvi;jÞ ¼ rfij;j þ f FSI
i ; ð14aÞ
vi;i ¼ 0: ð14bÞ
Moreover, the Cauchy stress tensor rfij is also defined as
rfij ¼ �pdij þ 2lsij; ð15aÞ
sij ¼ 12ðvi;j þ vj;iÞ; ð15bÞ
where l is the dynamic viscosity and p is the pressure.
In addition to these equations of motion, we stipulate the following boundary conditions:
vi ¼ gi; x 2 Cgi ; ð16aÞ
rfijnj ¼ hi; x 2 Chi ; ð16bÞ
where gi and hi are given functions.
For each spatial degree of freedom i, we assume that the fluid boundary C can be partitioned as
C ¼ Cgi [ Chi , with Cgi \ Chi ¼ ; (i.e. the Dirichlet and Neumann boundaries span the entire fluid surface
but do not intersect).
3.1.2. Petrov–Galerkin weak form
For the fluid domain, transforming the strong form of the conservation equations into the weak formrequires two test functions: the velocity test function dvi and the pressure test function dp for the
momentum equation and the continuity equation, respectively. To reduce or eliminate the numerical
oscillations in the velocity and pressure solutions from the standard Galerkin method for incompressible
materials, we employ the test functions d~vi and d~p along with the stabilization parameters sm and sc. Thealgorithm follows the stabilized equal-order finite element formulation developed in Refs. [11–13]:
d~vi ¼ dvi þ smvkdvi;k þ scdp;i; ð17aÞ
d~p ¼ dp þ scdvi;i; ð17bÞwhere sm and sc are the scalar stabilization parameters dependent of the computational grid, time step size,
and flow variables.The weak form of the continuity equation (14b) is obtained by multiplying the pressure test function d~p
and integrating over X,ZXðdp þ scdvi;iÞvj;j dX ¼ 0: ð18Þ
Similarly, the weak form of the momentum equation (14a) is achieved by multiplying the velocity testfunction d~vi,
L. Zhang et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 2051–2067 2057
ZXðdvi þ smvkdvi;k þ scdp;iÞ qfðvi;t
�þ vjvi;jÞ � f FSI
i
�dX
�ZXdvirf
ij;j dX�Xe
ZXe
ðsmvkdvi;k þ scdp;iÞrfij;j dX ¼ 0: ð19Þ
Applying integration by parts and the divergence theorem, we can rewrite Eq. (19) asZXðdvi þ smvkdvi;k þ scdp;iÞ qfðvi;t
�þ vjvi;jÞ � f FSI
i
�dX�
ZChi
dvihi dC
þZXdvi;jrf
ij dX�Xe
ZXe
ðsmvkdvi;k þ scdp;iÞrfij;j dX ¼ 0; ð20Þ
where individual finite elements are denoted with the subscript e and the stabilization supplementary termsare evaluated at the element interiors.
The detailed illustrations of the stabilized finite element formulations are available in Refs. [11–13]. Let
us assume that there is no traction applied on the fluid boundary, i.e.RChi
dvihi dC ¼ 0, the final weak form
yieldsZXðdvi þ smvkdvi;k þ scdp;iÞ qfðvi;t
�þ vjvi;jÞ � f FSI
i
�dXþ
ZXdvi;jrf
ij dX
�Xe
ZXe
ðsmvkdvi;k þ scdp;iÞrfij;j dXþ
ZXðdp þ scdvi;iÞvj;j dX ¼ 0: ð21Þ
The nonlinear systems are solved with the Newton–Raphson method. Moreover, to improve the compu-
tation efficiency, we also employ the GMRES iterative algorithm and compute the residuals based on
matrix-free techniques [9,21].
3.1.3. Finite element discretization
The discretization process using finite elements follows the procedures in Ref. [10]. The velocity andpressure functions, vi and p, along with the test functions dvi and dp, are interpolated as:
vhi ¼XI
NIdiI ; ð22aÞ
dvhi ¼XI
NIciI ; ð22bÞ
ph ¼XI
NIpI ; ð22cÞ
dph ¼XI
NIqI ; ð22dÞ
where NI is the nodal shape function at node I .Substituting the interpolations into Eq. (21), we derive
XI
ZX
ciINI
�þ smvhkciINI ;k þ scqINI ;i
�qfðvhi;th
þ vhj vhi;jÞ � f FSI;h
i
idXþ
XI
ZXciINI;jr
f ;hij dX
�XI
Xe
ZXe
smvhkciINI ;k
�þ scqINI ;i
�rf;hij;j dXþ
XI
ZXðqINI þ scciINI ;iÞvhj;j dX ¼ 0: ð23Þ
2058 L. Zhang et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 2051–2067
Combining the test function vectors ci and q, we have four set of equations with respect to four unknowns
at each node I :
cTi rwi ¼ 0; ð24aÞ
qT rq ¼ 0; ð24bÞ
where the residual vectors rwiI and rqI are given as follows:
rwiI ¼ZXðNI þ smvhkNI ;kÞ qfðvhi;t
hþ vhj v
hi;jÞ � f FSI;h
i
idXþ
ZXNI ;jr
f;hij dX
�Xe
ZXe
smvhkNI ;krf;hij;j dXþ
ZXscNI ;ivhj;j dX; ð25Þ
rqI ¼ZXscNI;i q
fðvhi;th
þ vhj vhi;jÞ � f FSI;h
i
idX�
Xe
ZXe
scNI;irf;hij;j dXþ
ZXNIvhj;j dX: ð26Þ
The pressure and velocity unknowns are solved by minimizing the residual vectors in Eqs. (25) and (26)
iteratively in each time step.
3.2. Structural mechanics
In this section, we start with the governing equation of the solid presented in Section 2. The weak form
and its discretization are then derived.
3.2.1. Strong form
The solid domain Xs is solved using the Lagrangian description. The material time derivative dvsi=dt cantherefore be directly introduced as €usi and Eq. (7) yields
ðqs � qfÞ€usi ¼ rsij;j � rf
ij;j þ ðqs � qfÞgi � f FSI;si : ð27Þ
The force balance equation in Eq. (27) is the linear momentum equation in the updated Lagrangian
description. The interaction force fFSI;s can be treated as an additional body force in the solid continuum.Obviously, if the densities are the same for both fluid and solid, and if the solid behaves as the fluid, the
entire computational domain would behave as the fluid with f FSI;si ¼ 0.
3.2.2. Weak form
To develop the weak form, we multiply the momentum equation, Eq. (27), by a test function dui andintegrate over the current configuration Xs:Z
Xs
dui ðqs
�� qfހusi �
orsij
oxj� ðqs � qfÞgi þ f FSI;s
i
�dXs ¼ 0: ð28Þ
Note that for brevity we ignore the fluid stress within the solid domain. The transformation of the weak
form from the updated Lagrangian to the total Lagrangian description is to change the integration domain
from Xs to Xs0. Since we consider incompressible fluid and solid, and the Jacobian determinant is 1 in the
solid domain, the transformation of the weak form to total Lagrangian description yieldsZXs0
dui ðqs
�� qfހusi �
oPijoXj
� ðqs � qfÞgi þ f FSI;si
�dXs
0 ¼ 0; ð29Þ
L. Zhang et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 2051–2067 2059
where the first Piola–Kirchhoff stress Pij is defined as Pij ¼ JF �1ik rs
kj and the deformation gradient Fij asFij ¼ oxi=oXj.Using integration by parts and the divergence theorem, we can rewrite Eq. (29) asZ
Xs0
duiðqs � qfÞ€usi dXs0 þ
ZXs0
dui;jPij dXs0 �
ZXs0
duiðqs � qfÞgi dXs0 þ
ZXs0
duifFSI;si dXs
0 ¼ 0: ð30Þ
Note again that the boundary integral terms on the fluid–structure interface for both fluid and solid do-
mains will cancel each other and for brevity are not included in the corresponding weak forms.
3.2.3. Finite element discretization
The discretization of the equilibrium equation (30) is done by introducing the finite element interpola-
tions of the test and trial functions,
duhi ¼XI
NIduiI ; ð31aÞ
uhi ¼XJ
NJuiJ : ð31bÞ
By substituting the interpolation functions into Eq. (30), we can obtain the discretized form:XI
XJ
ZXs0
NIduiIðqs � qfÞNJ€usiJ dXs0 þ
XI
ZXs0
NI;jduiIPij dXs0 �
XI
ZXs0
NIduiIðqs � qfÞgi dXs0
þXI
ZXs0
NIduiIfFSI;si dXs
0 ¼ 0: ð32Þ
Using the arbitrariness of duiI , we deriveXJ
ZXs0
NIðqs � qfÞNJ€usiJ dXs0 þ
ZXs0
NI;jPij dXs0 �
ZXs0
NIðqs � qfÞgi dXs0 þ
ZXs0
NIfFSI;si dXs
0 ¼ 0: ð33Þ
Therefore, at a discrete submerged point I represented with the position vector xsIðX
sI ; tÞ, we have the
following force equilibrium:
f inertiI þ f int
iI � f extiI þ f FSI;s
iI ¼ 0; ð34Þ
where f inertiI , f int
iI , f extiI , and f FSI;s
iI stand for the inertial, internal, external, and fluid–structure interactionforces at node I , respectively.
In the current IFEM implementation, we consider a three-dimensional almost incompressible hyper-
elastic material model with the Mooney–Rivlin material description [16], in which the elastic energy po-
tential W is given as:
W ¼ C1ðJ1 � 3Þ þ C2ðJ2 � 3Þ þ j2ðJ3 � 1Þ2; ð35Þ
where C1, C2, and j are the material constants and J1, J2, and J3 are functions of the invariants of theCauchy–Green deformation tensor C defined as Cij ¼ FmiFmj.
For structures with large displacements and deformations, the second Piola–Kirchhoff stress Sij and the
Green–Lagrangian strain Eij are used in the total Lagrangian formulation:
Sij ¼oWoEij
and Eij ¼1
2ðCij � dijÞ: ð36Þ
2060 L. Zhang et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 2051–2067
The first Piola–Kirchhoff stress Pij can then be obtained from the second Piola–Kirchhoff stress Sij:
Pij ¼ SikFjk; ð37Þwhere the definition of Pij is employed to derive the internal force f int
iI .
Note that if the solid material model is purely incompressible as constrained in Eqs. (11) and (10), we
should have the invariant I3 ¼ 1 and its function J3 ¼ 1 based on the definition of detðCÞ ¼ 1 for incom-
pressible materials. Therefore, the third term in the definition of W should be 0. However, the divergence
free property is not strictly satisfied in the process of interpolating the velocities, which induces accumulated
numerical errors in the fluid and solid calculations and creates possible artificial stress due to the volumechange of the solid. Therefore, in the implementation, we assign the bulk modulus j to be sufficiently small
so that the last term in Eq. (35) has nearly no contribution in the energy potential. A more elaborate study
on this subject will be discussed in a separate paper.
4. Construction of delta function
The coupling between the fluid and the structure consists of two parts: the interpolation of the fluid gridvelocity vðx; tÞ onto the structural nodal velocity vsðXs; tÞ, and the distribution of the solid domain force
fFSI;sðXs; tÞ onto the equivalent fluid domain force fFSIðx; tÞ. Note again that the fluid variables are depicted
in the Eulerian coordinate x; whereas the solid variables are represented in the Lagrangian coordinate Xs.
In this section we describe the construction of the function d that is used in the interactions. Both Dirac
delta functions in the IB method [3] and EIBM [16] are implemented for uniform fluid meshes. However,
the RKPM delta function employed in IFEM can handle both uniform and nonuniform grids [19]. With
this improvement, we are able to deal with fluid domains with arbitrary geometries.
As illustrated in Ref. [19], both wavelet and smooth particle hydrodynamics (SPH) methods belong to aclass of reproducing kernel methods where the reproduced function uRðxÞ is derived as
uRðxÞ ¼Z þ1
�1uðyÞ/ðx� yÞdy; ð38Þ
with a projection operator or a window function /ðxÞ.The reproducing condition requires that up to nth order polynomial can be reproduced, i.e.,
xn ¼Z þ1
�1yn/ðx� yÞdy: ð39Þ
To satisfy the consistency condition, a correction function Cðx; x� yÞ is introduced in the finite domain
of influence or support, so that the window function yieldsZXCðx; x� yÞ 1
r/
x� yr
� dX ¼ 1; ð40Þ
where r is the dilation parameter or refinement of the window function. The additional constant, r�1, scales
the window function so that the integral over the domain of support equals one.The discretized reconstruction of the delta function for nonuniform spacing can be written as
/IðxÞ ¼ Cðx; x� xIÞ1
r/
x� xIr
� DxI : ð41Þ
The window function is used as the delta function for the interpolation of the nodal velocities and the
distribution of the nodal forces between the fluid and solid domains. To ensure the conservation of energy,
the delta functions in both distribution and interpolation procedures must be the same at each time step.
L. Zhang et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 2051–2067 2061
In particular, for each solid point I at its current position xsIðtÞ, we construct an influence domain X/I
that covers a subset of the fluid grid nodes J positioned at xJ . To define the influence domain for each solid
point I , a search algorithm is needed at each time step during the calculation of the RKPM shape functions.
When nonuniform fluid mesh is used, the size of the influence domain for each solid point is dependent of
the definition of the RKPM shape function in Eq. (41) and fluid mesh density.
Therefore, in the interpolation process from the fluid onto the solid grid, the discretized form of Eq. (11)
can be written as
vsiI ¼XJ
viJ ðtÞ/J ðxJ � xsIÞ; xJ 2 X/I : ð42Þ
Here, the solid velocity vsI at node I can be calculated by gathering the velocities at fluid nodes within the
influence domain X/I . A dual procedure takes place in the distribution process from the solid onto the fluid
grid. The discretized form of Eq. (8) is expressed as
f FSIiJ ¼
XI
f FSI;siI ðXs; tÞ/IðxJ � xs
IÞ; xsI 2 X/J : ð43Þ
By interpolating the fluid velocities onto the solid particles in Eq. (42), the fluid within the solid domain
is bounded to solid material points. This ensures not only the no-slip boundary condition on the surface ofthe solid, but also stops automatically the fluid from penetrating the solid, provided the solid mesh is at
least two times denser than the surrounding fluid mesh. This heuristic criterion is based on the numerical
evidence and needs further investigation.
5. Computational algorithm
The governing equations of IFEM in discretized form (except the Navier–Stokes equations, for con-venience) are summarized as follows:
f FSI;siI ¼ �f inert
iI � f intiI þ f ext
iI ; in Xs; ð44aÞ
f FSIiJ ¼
XI
f FSI;siI ðXs; tÞ/IðxJ � xs
IÞ; xsI 2 X/J ; ð44bÞ
qfðvi;t þ vjvi;jÞ ¼ rij;j þ qgi þ f FSIi ; in X; ð44cÞ
vj;j ¼ 0; in X; ð44dÞ
vsiI ¼XJ
viJ ðtÞ/J ðxJ � xsIÞ; xJ 2 X/I : ð44eÞ
An outline of the IFEM algorithm with a semi-explicit time integration is illustrated as follows:
(1) Given the structure configuration xs;n and the fluid velocity vn at time step n,(2) Evaluate the nodal interaction forces fFSI;s;n for solid material points, using Eq. (44a),
(3) Distribute the material nodal force onto the fluid lattice, from fFSI;s;n to fFSI;n, using the delta function as
in Eq. (44b),
(4) Solve for the fluid velocities vnþ1 and the pressure pnþ1 implicitly using Eqs. (44c) and (44d), described in
Section 3.1,
(5) Interpolate the velocities in the fluid domain onto the material points, i.e., from vnþ1 to vs;nþ1 as in Eq.(44e), and
(6) Update the positions of the structure using us;nþ1 ¼ vs;nþ1Dt and go back to step 1.
2062 L. Zhang et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 2051–2067
Note that even though the fluid is solved fully implicitly, the coupling between fluid and solid is explicit.If we rewrite the fluid momentum equation (for clarity only discretized in time), we have
qf vmþ1i � vmi
Dt
�þ vmþ1
j vmþ1i;j
�¼ rf ;mþ1
ij;j þ f FSI;mi : ð45Þ
It is clear that the interaction force is not updated during the iteration, i.e., the solid equations are cal-
culated with values from the previous time step. For a fully implicit coupling, this force must be a function
of the current fluid velocity and the term f FSI;mþ1i should be included into the linearization of the fluid
equations.
6. Numerical examples
In this section, we present the simulations on particulate flows with both rigid and deformable spheres
using the proposed IFEM method. The algorithm requires no mesh update or remeshing and the structures
are not restricted to only rigid cases as explained in Section 2. The demonstrated problems are: a rigid
sphere and a soft sphere dropping between two planes and multiple soft spheres dropping in a square
channel. The sphere flexibility is controlled by adjusting the material properties. Naturally, as the modulus
increases, the sphere behaves more like a rigid body.
6.1. A rigid sphere dropping between two planes
In this section, we discuss the case with a rigid sphere dropping in a channel. Since a rigid structure is
considered, the material modulus is set to be high. The channel has a size of 2 · 2 · 8 cm. The diameter of
the sphere is 0.4 cm. The properties for both fluid and solid as well as the discretization for each mesh are
summarized in Table 2.
As illustrated in Fig. 2, the use of the nonuniform fluid mesh provides the unique ability and the
advantage of IFEM over the IB method as well as EIBM.The interaction force fFSI is calculated using Eq. (44a). As shown in Fig. 3, the pressure and velocities in
the fluid domain are influenced by the movement of the sphere and the vortices and recirculation are clearly
captured near the sphere.
To verify the accuracy of IFEM algorithm, we compared the solution of terminal velocity with the
experimental data. It is explained in Ref. [1] that for a particle moving with steady terminal velocity UT in a
gravitational field, the drag force FD balances the difference between the weight and buoyancy:
FD ¼ gðqs � qfÞðp=6ÞD3 ð46Þand as a consequence the drag coefficient CD is calculated as
Table 2
Properties used for one rigid sphere dropping in a channel
Fluid 4987 nodes qf ¼ 1 g/cm3
26 030 elements l ¼ 0:01 g cm/s
Solid 997 nodes qs ¼ 2:02 g/cm3 C1 ¼ 2:93� 105 g/(cm s2)
864 elements D ¼ 0:4 cm C2 ¼ 1:77� 105 g/(cm s2)
g ¼ 980 cm/s2
Fig. 2. A gradient mesh is constructed with respect to the distance from the submerged sphere.
Fig. 3. Velocity vectors and pressure contours in the fluid domain at different time steps. (a) t ¼ 0:01 s; (b) t ¼ 0:1 s; (c) t ¼ 0:2 s;
(d) t ¼ 0:3 s; (e) t ¼ 0:4 s.
L. Zhang et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 2051–2067 2063
CD ¼ 4ðqs � qfÞgD=3qfU 2T ¼ 4qfðqs � qfÞgD3=3l2Re2T; ð47Þ
where ReT is the Reynolds number at the terminal velocity.
Since both the drag coefficient and Reynolds number are dependent of the terminal velocity UT, an
iterative procedure is needed. It is more convenient to express Re as a function of the so-called best number
ND, which is defined as
ND ¼ CDRe2T ¼ 4qfðqs � qfÞgD3=3l2: ð48ÞFinally, the Reynolds number can be expressed as a function of ND. In this case, for 5806ND 6
1:55� 107, the correlation for Re as a function of ND is expressed as
log10 Re ¼ �1:81391þ 1:34671W � 0:12427W 2 þ 0:006344W 3; ð49Þwith W ¼ log10 ND.
Fig. 4. Comparison between the experimental data and the IFEM solution.
2064 L. Zhang et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 2051–2067
Using the parameters shown in the above example, the experimental terminal velocity is found to be
2.203 cm/s with the terminal Reynolds number 88:11. The calculated terminal velocity from the IFEM
solution, as shown in Fig. 4, is found to be 2.24412 cm/s, which deviates only 1.87% from the experimental
value.
6.2. A soft sphere dropping in a channel
This example shows a soft sphere dropping in a viscous fluid channel. The sphere flexibility is controlledby the parameters C1 and C2 that are defined in Section 3.2 and listed in Table 3.
The deformation of the sphere as well as the stress distribution at different time steps are shown in Fig. 5.
As expected from the simulation, the sphere experiences a higher pressure at the bottom and a lower
pressure on the top as it tries to immerse deeper into the fluid. The deformation can be clearly observed as it
accelerates inside the fluid domain, shown in Fig. 5.
6.3. 20 Soft spheres dropping in a channel
This example solves 20 soft spheres dropping in a channel. These same-sized spheres are placed ran-
domly near the top of the channel. The spheres experience the gravitational and internal forces as well as
the interaction forces with the surrounding fluid. The properties used in this problem can be found in
Table 4.
Table 3
Properties used for one soft sphere dropping in a channel
Fluid 4861 nodes qf ¼ 1 g/cm3
22 557 elements l ¼ 10 g cm/s
Solid 997 nodes qs ¼ 2 g/cm3 C1 ¼ 2:93� 102 g/(cm s2)
864 elements D ¼ 0:5 cm C2 ¼ 1:77� 102 g/(cm s2)
g ¼ 980 cm/s2
Fig. 5. Von Mises stress of the soft sphere at different time steps as it drops in a channel. (a) t ¼ 0 s; (b) t ¼ 7:50 s; (c) t ¼ 11:25 s;
(d) t ¼ 15:00 s; (e) t ¼ 18:75 s.
Table 4
Properties for 20 soft spheres dropping in a channel
Fluid 4861 nodes qf ¼ 1 g/cm3
22 557 elements l ¼ 10:0 g cm/s
Solid 20 · 997 nodes qs ¼ 2 g/cm3 C1 ¼ 2:93� 101 g/(cm s2)
20 · 864 elements D ¼ 0:5 cm C2 ¼ 1:77� 101 g/(cm s2)
g ¼ 980 cm/s2
Fig. 6. The movement of 20 spheres dropping in a channel at different time steps. (a) t ¼ 0 s; (b) t ¼ 1:0 s; (c) t ¼ 2:0 s; (d) t ¼ 3:0 s;
(e) t ¼ 4:0 s.
L. Zhang et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 2051–2067 2065
Fig. 7. The Von Mises stress distribution and deformation of 20 spheres dropping in a channel at different time steps. (a) t ¼ 0:0 s;
(b) t ¼ 1:0 s; (c) t ¼ 2:0 s; (d) t ¼ 3:0 s; (e) t ¼ 4:0 s.
2066 L. Zhang et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 2051–2067
The movements of the spheres at different time steps are shown in Fig. 6. To clearly illustrate the
deformations of the spheres, we present in Fig. 7 the enlarged images of the spheres and the pressure
distributions at different time steps.
7. Conclusions
In this paper, we proposed the IFEM to solve complex fluid–structure interaction problems. In com-
parison with the IB method, we demonstrate the following advantages of IFEM.
The structural models in IFEM are not restricted to one-dimensional volumeless structures such as fi-
bers; instead, they may occupy finite volumes in the fluid. Finite elements are used to discretize the sub-
merged solid and as a consequence the calculated stress are more realistic and accurate. In addition, inIFEM, different material laws can be assigned to the submerged solid.
The higher-ordered RKPM delta function has the ability to handle nonuniform grids and is particularly
useful for fluid domains with arbitrary geometries and boundaries. Finally, unlike the periodic boundary
conditions implemented in most IB applications, essential and natural boundary conditions in the context
of finite element methods are relatively easy to implement in IFEM.
Acknowledgements
We would like to thank Prof. C.S. Peskin in the Courant Institute of Mathematical Sciences at New
York University and Prof. N.A. Patankar at Northwestern University for the helpful discussions and
suggestions. The authors are sponsored by NASA Langley.
References
[1] R. Clift, J.R. Grace, M.E. Weber, Bubbles, Drops, and Particles, Academic Press, 1978.
L. Zhang et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 2051–2067 2067
[2] C.S. Peskin, Numerical analysis of blood flow in the heart, J. Comput. Phys. 25 (1977) 220–252.
[3] C.S. Peskin, The immersed boundary method, Acta Numer. 11 (2002) 479–517.
[4] R. Glowinski, T.W. Pan, T.I. Hesla, D.D. Joseph, J. P�eriaux, A fictitious domain approach to the direct numerical simulation of
incompressible viscous flow past moving rigid bodies: application to particulate flow, J. Comput. Phys. 169 (2001) 363–426.
[5] H.H. Hu, N.A. Patankar, M.Y. Zhu, Direct numerical simulations of fluid–solid systems using the arbitrary Lagrangian–Eulerian
technique, J. Comput. Phys. 169 (2001) 427–462.
[6] A. Huerta, W.K. Liu, Viscous flow with large free surface motion, Comput. Methods Appl. Mech. Engrg. 69 (1988) 277–324.
[7] A. Johnson, T.E. Tezduyar, Advanced mesh generation and update methods for 3d flow simulations, Comput. Mech. 23 (1999)
130–143.
[8] N.A. Patankar, X. Wang, W.K. Liu, A note on immersed boundary formulation for elastic bodies in fluids, preprint, 2003.
[9] Y. Saad, M.H. Schultz, GMRES: a generalized minimal residual algorithm for solving nonsymmetric linear systems, SIAM J. Sci.
Statist. Comput. 7 (3) (1986) 856–869.
[10] T.B. Belytschko, W.K. Liu, B. Moran, Nonlinear Finite Elements for Continua and Structures, John Wiley & Sons, 2000.
[11] T.E. Tezduyar, Stabilized finite element formulations for incompressible-flow computations, Adv. Appl. Mech. 28 (1992) 1–44.
[12] T.E. Tezduyar, Finite element methods for flow problems with moving boundaries and interfaces, Arch. Comput. Methods Engrg.
8 (2) (2001) 83–130.
[13] T.J.R. Hughes, L.P. Franca, M. Balestra, A new finite element formulation for computational fluid dynamics: V. Circumventing
the Babu�ska–Brezzi condition: a stable Petrov–Galerkin formulation of the Stokes problem accommodating equal-order
interpolations, Comput. Methods Appl. Mech. Engrg. 59 (1986) 85–99.
[14] T.J.R. Hughes, W.K. Liu, T.K. Zimmerman, Lagrangian–Eulerian finite element formulations for incompressible viscous flows,
Comput. Methods Appl. Mech. Engrg. 29 (1981) 329–349.
[15] X. Wang, On the discretized delta function and force calculation in the immersed boundary method, in: K.J. Bathe (Ed.),
Computational Fluid and Solid Mechanics, Elsevier, 2003, pp. 2164–2169.
[16] X. Wang, W.K. Liu, Extended immersed boundary method using FEM and RKPM, Comput. Methods Appl. Mech. Engrg., in
press.
[17] W.K. Liu, Finite element procedures for fluid–structure interactions and application to liquid storage tanks, Nucl. Engrg. Des. 65
(1981) 221–238.
[18] W.K. Liu, J. Gvildys, Fluid–structure interaction of tanks with an eccentric core barrel, Comput. Methods Appl. Mech. Engrg. 58
(1986) 51–77.
[19] W.K. Liu, S. Jun, Y.F. Zhang, Reproducing kernel particle methods, Int. J. Numer. Methods Fluids 20 (1995) 1081–1106.
[20] W.K. Liu, D.C. Ma, Computer implementation aspects for fluid–structure interaction problems, Comput. Methods Appl. Mech.
Engrg. 31 (1982) 129–148.
[21] L. Zhang, G.J. Wagner, W.K. Liu, A parallaized meshfree method with boundary enrichment for large-scale CFD, J. Comput.
Phys. 176 (2002) 483–506.