Available online at www.sciencedirect.

com

Mathematics and Computers in Simulation 82 (2012) 1645–1666

Original articles

A finite volume method on NURBS geometries and its application in

isogeometric fluid–structure interaction
Ch. Heinrich a,∗ , B. Simeon b , St. Boschert c
a Technische Universität München, Zentrum Mathematik, Boltzmannstraße 3, 85748 Garching, Germany
b Technische Universität Kaiserslautern, Felix-Klein-Zentrum für Mathematik, Paul-Ehrlich-Straße, 67663 Kaiserslautern, Germany
c Siemens AG, CT T DE TC 3, Otto-Hahn-Ring 6, 81739 München, Germany

Received 23 February 2011; received in revised form 10 December 2011; accepted 28 March 2012
Available online 19 April 2012

Abstract
A finite volume method for geometries parameterized by Non-Uniform Rational B-Splines (NURBS) is proposed. Since the
computational grid is inherently defined by the knot vectors of the NURBS parameterization, the mesh generation step simplifies
here greatly and furthermore curved boundaries are resolved exactly. Based on the incompressible Navier–Stokes equations, the
main steps of the discretization are presented, with emphasis on the preservation of geometrical and physical properties. Moreover,
the method is combined with a structural solver based on isogeometric finite elements in a partitioned fluid–structure interaction
coupling algorithm that features a gap-free and non-overlapping interface even in the case of non-matching grids.
© 2012 IMACS. Published by Elsevier B.V. All rights reserved.

Keywords: Finite volume method; NURBS; Exact geometry; Navier–Stokes equations; Fluid–structure interaction; Matching interface

1. Introduction

In numerical partial differential equations, the role of the geometry description has been mostly neglected or
underestimated for many years. Only recently, with the advent of isogeometric analysis [18], it has been demonstrated
that geometry and numerics may go hand in hand with substantial mutual benefits. In this contribution we discuss the
class of finite volume methods (FVM) from the same perspective and introduce a discretization technique for geometries
parameterized by Non-Uniform Rational B-Splines (NURBS). The discretization is able to preserve free-form surfaces
and is particularly attractive for the combination with an isogeometric solver in fluid–structure interaction problems.
The isogeometric approach extends isoparametric finite elements to more general basis functions such as B-splines
and NURBS [8]. In this way, exact geometries at the coarsest level of discretization are obtained and geometry errors
are eliminated from the very beginning. The resulting discretization still fits into the variational framework of the finite
element method (FEM), and it is possible to equip established FEM codes with isogeometric elements [4].
In computational fluid dynamics, however, the FVM is still the method of choice. It can be interpreted as a discon-
tinuous Galerkin method of lowest order, i.e., with constant approximations in each cell or control volume. A major

∗ Corresponding author.
E-mail addresses: heinrich@ma.tum.de (Ch. Heinrich), simeon@mathematik.uni-kl.de (B. Simeon), stefan.boschert@siemens.com
(St. Boschert).

0378-4754/$36.00 © 2012 IMACS. Published by Elsevier B.V. All rights reserved.

http://dx.doi.org/10.1016/j.matcom.2012.03.008
1646 Ch. Heinrich et al. / Mathematics and Computers in Simulation 82 (2012) 1645–1666

feature of isogeometric FEM is the optional usage of higher-order geometry basis functions in the Galerkin projection
with higher smoothness compared to classical FEM. Therefore, the numerical solution, in general, also exhibits a higher
smoothness, which can be advantageous as, e.g., in stress analysis [30]. Obviously, the transfer of this feature to the
finite volume framework is not possible as we seek for a discontinuous piecewise constant solution. In computational
fluid dynamics, however, it is often desirable to allow for discontinuous solutions, which can occur in reality, e.g., when
it comes to shocks. We will show in this contribution that it is straightforward to directly define an FVM on NURBS
geometries without the need for an additional mesh generation step where the original geometry is approximated by
polygonal or simplicial shapes.
More specifically, we transform the underlying integral formulation of the FVM to the parametric domain of the
geometry description and apply the standard discretization in parametric space, enhanced by higher-order quadrature
rules to integrate non-linear geometrical information with a controllable error. The latter is not only important to
incorporate the NURBS data exactly but also to satisfy the first geometric conservation law (GCL) that comes into play
when we consider control volumes with curved boundaries in physical space [23]. In that way the presented approach
can be seen as a superparametric method. As opposed to structural mechanics and classical finite elements the usage of
the latter is common practice in computational fluid dynamics and the discontinuous Galerkin method. When it comes
to curved boundaries this technique can have a significant impact on the quality of the numerical solution [15].
A signature feature of the present approach is that the computational mesh is inherently defined by the knot vectors
of the NURBS parameterization, and therefore a mesh generation step can be omitted. Furthermore, we present a
partitioned fluid–structure interaction (FSI) coupling algorithm combining the FVM on NURBS geometries with a
structural solver based on isogeometric finite elements. While classical coupling methods have the drawback that gaps
and overlaps occur if different mesh resolutions are used in the single fields at the interface, this problem can be
overcome in the isogeometric setting, leading to an approach with gap-free and non-overlapping interfaces even in the
case of non-matching grids.
Among the various recent publications in the emerging field of isogeometric analysis, we refer to the monograph [8]
and the tutorial [29], while standard references on splines and CAD are [14,26]. Tri-variate volume parameterization
techniques are presented in [1], whereas [6] deals with saddle point problems and [3,2] with isogeometric FSI and a
monolithic coupling algorithm. References for FSI and partitioned solution approaches are, e.g., [22,31]. In the field
of the FVM we mention here [12,13,16] for a general treatment and the theoretical underpinnings.
This article is organized as follows: Section 2 introduces NURBS at a glance and discusses the connection to
grid generation. The details of our approach, based on a discontinuous Galerkin formulation and elaborated for the
incompressible Navier–Stokes equations, are presented in Section 3. Thereafter, in Section 4, the FSI problem and
an isogeometric coupling algorithm are described. Furthermore, we introduce a sufficient condition for the fluid and
structure parameterizations such that an exact or matching interface can be guaranteed in the sense that it is gap-free
and non-overlapping. Finally, Section 5 presents numerical results for flow simulation problems and an FSI example.
The paper closes with the conclusions and an outlook.

2. NURBS and grid generation

The focus of this contribution are computational domains that are defined by means of NURBS, one of the standard
representations in computer aided design (CAD). We therefore summarize first some of the basics on NURBS and
introduce further geometric quantities to be used below. Moreover, the corresponding computational mesh, which
allows for curved boundaries, is defined.

2.1. NURBS and geometric quantities

We start with the quite general concept of a geometry function

F : 0 → , F(␰) = x (␰) , (1)

which maps a parametric domain 0 to a physical domain . While we aim at solving systems of partial differential
equations in the physical domain, the geometry function allows us to transform the problem at hand and the numerical
scheme to the parametric domain.
 Ch. Heinrich et al. / Mathematics and Computers in Simulation 82 (2012) 1645–1666 1647

Fig. 1. Mapping of the parametric domain to the physical space. A control volume i0 in parametric space and its image i = F(i0 ) are highlighted
in gray.

For ease of presentation, we restrict ourselves here to the planar case where the corresponding coordinates read
ξ = (ξ, η)T and x = (x, y)T , cf. Fig. 1. But we point out that the extension to the 3D case follows exactly the same
methodology. The geometry function F is now and for the remainder of this paper assumed to be a NURBS, a
Non-Uniform Rational B-Spline.
Such spline functions are typically first formulated in a univariate framework and then extended to the multivariate
case. To this end, let  = (ξ0 , . . . , ξl ) ∈ Rl+1 be the knot vector of a NURBS of degree p consisting of non-decreasing
real numbers. We assume that the first and the last knot have multiplicity p + 1, which means that their values are
repeated p times. Therefore, the endpoints are interpolatory, which is important for representing the computational
domain exactly. The ith B-spline basis function of p-degree Ni,p is defined recursively as

1 ifξi ≤ ξ < ξi+1 ,
Ni,0 (ξ) = (2)
0 otherwise;
ξ − ξi ξi+p+1 − ξ
Ni,p (ξ) = Ni,p−1 (ξ) + Ni+1,p−1 (ξ). (3)
ξi+p − ξi ξi+p+1 − ξi+1
Note that the quotient 0/0 is assumed to be zero.
In one dimension, a NURBS of degree p is then given by
wi Ni,p (ξ)
Ri,p (ξ) =  (4)
j∈J wj Nj,p (ξ)

with B-splines Ni,p , weights wi ∈ R, and an index set J = {0, . . . , l − p − 1} labeling the B-splines. If the multiplicity
of a knot ξ i is mi ≥ 1, the smoothness of the spline is Cp−mi in this knot. E.g., the multiplicity mi = p reduces the
smoothness to C0 . By specifying single or multiple knots, one may thus change the smoothness of a NURBS.
Bivariate NURBS are constructed via (suppressing the degrees pξ and pη )
wk Nk (ξ)N (η)
Rk (ξ, η) =   . (5)
i∈I j∈J wij Ni (ξ)Nj (η)

Though this representation requires knot vectors ξ = (ξ0 , . . . , ξl1 ) and η = (η0 , . . . , ηl2 ) for each parameter direc-
tion, it does not exactly have a tensor product structure due to the weights wij , which can be altered separately. The index
sets are I = {0, . . . , l1 − pξ − 1} and J = {0, . . . , l2 − pη − 1}. Note that the continuity in each parameter direction
is determined by the knot multiplicities of the corresponding knot vectors.
A single patch NURBS parameterization of the physical domain consists of the geometry function

F(ξ, η) = Rij (ξ, η)cij (6)
i∈I j∈J

with bivariate NURBS Rij defined on 0 = [ξ0 , ξl1 ] × [η0 , ηl2 ] and control points cij ∈ R2 . The shape of  is thus
defined by the position of the control points, the weights, the knot vectors ξ , η , and the degrees pξ , pη . Changing
any of these results in a different geometry.
One of the main advantages of a NURBS parameterization is that conic intersections (ellipses, hyperbolas, parabolas)
can be represented exactly. Typical engineering shapes such as circular pipes are thus very easy to model. In general,
however, these shapes are just defined by outer and inner hulls, i.e., by a surface description. Techniques for extending
CAD models to a volume parameterization by means of bi-variate NURBS (planar case) or tri-variate NURBS (3D
1648 Ch. Heinrich et al. / Mathematics and Computers in Simulation 82 (2012) 1645–1666

case) are currently under investigation in the context of isogeometric analysis, see, e.g., [1]. It has been shown in
[32] that the kind of parameterization can have a significant impact on the accuracy of the numerical solution. We
also add that complicated shapes are usually designed as composite objects and parameterized by several geometry
functions. The resulting individual components are then glued together by means of certain continuity requirements.
Such multi-patch parameterizations are beyond the scope of this contribution but will be discussed in a forthcoming
article.
Partial derivatives of the geometry function will play an important role later on when discussing the finite volume
method for a NURBS geometry. The following notation proves useful in this context, see [17,25] for more details. If
F is a C1 -continuous function, its Jacobian is written as
 
xξ xη
J= , (7)
yξ y η

where xξ denotes the partial derivative of x with respect to ξ, and

J = xξ · yη − xη · yξ (8)
is its determinant. The covariant metric tensor is defined as
 
xξ2 + yξ2 xξ xη + yξ yη
g= , (9)
xξ xη + yξ yη xη2 + yη2
where g12 = g21 = 0 in the case of orthogonal grids.

2.2. Grid generation

A key ingredient of most numerical algorithms for solving partial differential equations is the computational mesh.
Practical experience shows that for complex domains, generating a mesh may actually become a bottleneck of the
whole numerical simulation chain. This gets apparent when the geometry is created in a CAD format but the grid
generator applies a different format and needs to communicate with the CAD system.
Isogeometric analysis promises to overcome or at least significantly improve this bottleneck by using the same
description both for the geometry and the simulation. Extending this approach to the finite volume case, it is quite
natural to make use of the mesh which is inherently defined by the knot vectors of the NURBS description. This meshing
of the parametric domain consists of rectangles in the planar case and cuboids in 3D. In this way, the parametric domain
0 = ∪ ˙ i0 is divided into subsets (or cells, elements) i0 , which are connected to their counterpart in the physical
domain via the NURBS geometry function such that i = F(i0 ) for all i, cf. Fig. 1. While the grid is rectilinear with
respect to the parametric domain, its image in the physical domain is, in general, curved.
Adopting the usual terminology for finite volume methods, we call the cells i0 control volumes, and a discretization
is admissible if the control volumes in the parametric domain satisfy the following conditions, cf. [11,12]:

Definition 2.1. (Admissible discretization)

Let 0 be an open bounded rectangular subset of Rd (d = 2, 3), and ∂0 = 0 \0 its boundary. An admissible
finite volume discretization of 0 is given by D = (M, E, P) where:

1 M is a finite family of non-empty open rectangular disjoint subsets of 0 (the control volumes) such that 0 =
∪K∈M K.
2 E is a finite family of disjoint subsets of 0 (the edges of the mesh), such that, for all σ ∈ E, there exists an affine
hyperplane E of Rd and K ∈ M with σ ⊂ ∂K ∩ E and σ is a non-empty open subset of E. We assume that, for all
K ∈ M, there exists a subset EK of E such that ∂K = ∪σ∈EK σ. We also assume that, for all σ ∈ E, either σ ⊂ ∂0
or σ = K ∩ L for some (K, L) ∈ M × M.
3 P is a family of points of 0 with coordinates (ξ K )K∈M (in 2D ξ K = (ξK , ηK )T ) and such that, for all K ∈ M,
ξ K ∈ K. Furthermore, the straight line connecting the points ξ K and ξ L of two neighboring control volumes K and L
has to be orthogonal to their common edge.
 Ch. Heinrich et al. / Mathematics and Computers in Simulation 82 (2012) 1645–1666 1649

Fig. 2. Control volume i0 with center of gravity P and parametric coordinates.

If we define the points in P to be the centers of gravity of the control volumes, then the mesh induced by the knot
vectors of the geometry function obviously fulfills Definition 2.1. A control volume i0 = [ξ1i , ξ2i ] × [ηi1 , ηi2 ] is shown
in Fig. 2, together with its center of gravity P.
To summarize, the geometry function (6) automatically defines an admissible computational mesh. Though this
initial mesh, in general, will be rather coarse, its refinement involves almost no extra effort since it can be performed by
adding knots where desired. A knot insertion algorithm in the sense of computing new control points is not necessary
as we still use the original knot vectors to define the geometry function and introduce distinct knot vectors dξ and dη to
define the corresponding refined mesh. Distinct knot vectors neglect repeated knots and can be refined by adding knots
where desired. If, e.g., the knot vectors are ξ = (0, 0, 0, 0.5, 1, 1, 1) and η = (0, 0, 0, 1, 1, 1) their distinct counterparts
are dξ = (0, 0.5, 1) and dη = (0, 1). Based on these distinct knot vectors we get the coarsest computational mesh as
depicted in Fig. 3 on the left. The uniformly refined distinct knot vectors and the resulting mesh can be seen in the
same figure on the right. Thereby, the original NURBS data is left unchanged.
Obviously, the overall geometric shape is exactly preserved during this process. However, we remark that, due to
the tensor product structure of the NURBS parameterization, a local refinement of only one of the knot vectors creates
a horizontal (or vertical) refined stripe in the mesh and thus has a global effect. In the context of isogeometric analysis
promising approaches have emerged to remedy this drawback [28]. Our approach, so far, requires rectilinear and
matching grids in parametric space, which precludes locally refined meshes. But, we tackle this problem in connection
with the extension of our approach to geometries defined by multiple patches, where non-matching grids are allowed
at the interfaces between patches. As mentioned before, this will be addressed in our future work.
Finally, as our approach is based on the transformation of integrals from physical space to parametric space, we
state here the required transformation rules.
For volume integrals it holds
 
f (x) dx = f (F(ξ))|J| dξ, (10)
i i0

where f : i → R is in L1 (i ) and F a C1 diffeomorphism.

Fig. 3. Computational mesh in the parametric domain for distinct knot vectors dξ = (0, 0.5, 1) and dη = (0, 1) (left) and after one uniform
h-refinement step (right).
1650 Ch. Heinrich et al. / Mathematics and Computers in Simulation 82 (2012) 1645–1666

Surface integrals over one face ∂ij of a control volume i are transformed via
 
h(x) ds = ij
h(F(ξ))|Ḟ(ξ)| ds(ξ), (11)
∂ij ∂0

where h : ∂ij → R is continuous, F ∈ C1 and Ḟ denotes the partial derivative of F with respect to the coordinate that
is not constant on the face.

3. The finite volume method on NURBS geometries

This section presents the extension of the finite volume method to NURBS geometries in detail. We concentrate
on the incompressible Navier–Stokes equations as standard application in fluid dynamics, outline the corresponding
discretization steps, and discuss the properties of the method.

3.1. Discretization

Consider the incompressible Navier–Stokes equations

1
ut − ν u + (u · ∇) u + ∇p = 0, (12)
ρ
∇ ·u = 0 (13)
with balance of momentum (12) and continuity equation (13) on a domain  ⊂ Rd
defined by a NURBS geometry
function. Here, u :  × [0, T ] → Rd where d = 2 or d = 3 represents the unknown velocity field and p :  × [0, T ] →
R the unknown pressure. Furthermore, ν ∈ R and ρ ∈ R in the momentum equation (12) are material parameters of the
fluid, denoting the kinematic viscosity and the mass density, respectively, and the temporal derivative of u is indicated
by a subscript t.

3.1.1. Weak formulation

We partition the computational domain  into N finite volumes i , i = 1, . . ., N, and follow the formalism of the
Discontinuous Galerkin (DG) methods as given by, e.g., [24]. Multiplying (12) and (13) by test functions v and q,
respectively, integrating over a control volume and applying partial integration leads to the weak formulation for each
i
    
ut · v dx + ν ∇u : ∇vdx − ν (∇u · n) · v ds − (u ⊗ u) : ∇v dx + ((u ⊗ u) · n) · v ds
i i ∂i i ∂i
 
1 1
− p∇ · v dx + pn · v ds = 0, (14)
ρ i ρ ∂i

 
− u · ∇q dx + u · nq ds = 0, (15)
i ∂i
where n denotes the outward pointing unit normal and ⊗ represents the dyadic operator.

3.1.2. Integral formulation

Next, the exact solution components u and p are approximated by their numerical counterpart uh and ph . We
assume constant test functions in each i such that the weak form (14) and (15) simplifies to the well-known integral
formulation
   
h · n) ds + 1
uh,t dx − ν (∇u (uh ⊗ uh ) · n ds + ph n ds = 0, (16)
i ∂i ∂i ρ ∂i

uh · n ds = 0, (17)
∂i
 Ch. Heinrich et al. / Mathematics and Computers in Simulation 82 (2012) 1645–1666 1651

which is the basis of the finite volume method. Here, the flux variables ∇uh , uh and ph need to be approximated
via interpolation. We will come back to this point in the next subsection. Clearly, the finite volume method can be
interpreted as a DG method of lowest order.
We now transform all arising integrals to the parametric domain 0 , based on the rules that have been already
introduced in (10) and (11). Noteworthy, this is equivalent to first transforming the PDE system to the parametric
domain and deriving the integral formulation thereafter. For ease of presentation, we do not distinguish between
variables in parametric or physical coordinates.
The time derivative term and the convective term become
 
uh,t dx = uh,t |J| dξ, (18)
i i0
 
(uh ⊗ uh ) · n ds = |J|(uh ⊗ uh ) · J−T · n0 ds (ξ) . (19)
∂i ∂i0

The pressure term and the continuity equation lead to

 
1 1
ph n ds = |J|ph J−T · n0 ds(ξ), (20)
ρ ∂i ρ ∂i0
 
uh · n ds = |J|uh · (J−T · n0 ) ds(ξ), (21)
∂i ∂i0

and the diffusive term in matrix–vector notation yields

 
ν h · n) ds = ν
(∇u h · J−1 · J−T · n0 ds(ξ)
|J|∇u (22)
∂i ∂i0

where n0 denotes the outward pointing unit normal in parametric space. In the planar case, the last expression consists
of the components
   
|J| g22 u1h,ξ − g12 u1h,η −g21 u1h,ξ + g11 u1h,η
h · n) ds =
(∇u · n0 ds(ξ), (23)
∂i ∂i0 J
2 g22 u2h,ξ − g12 u2h,η −g21 u2h,ξ + g11 u2h,η

with u1h,ξ denoting the partial derivative of the first component of the diffusive flux with respect to ξ, and so forth for
the other terms.
For planar problems, the control volumes are quadrilaterals, and we can thus split up the boundary integrals in
ij
(19)–(22) into four integrals over the southern, the eastern, the northern and the western face ∂0 where j ∈ {s, e, n, w}.
The diffusive term then reads
  
h · J−1 · J−T · n0 ds (ξ) =
|J|∇u h · J−1 · J−T · n0 ds (ξ) ,
|J|∇u (24)
ij
∂i0 j∈{s,e,n,w} ∂0

and the other boundary integrals follow in the same fashion.

3.1.3. Flux definition

To approximate the flux variables ∇u h , uh , and ph , we introduce computational nodes in the centers of gravity of
the control volumes 0 , i.e., the points in P (cf. Section 2.2), and apply linear interpolation, which is equivalent to the
i

central difference scheme (CDS). For the eastern face, assuming both the control volume and its center of gravity have
the same notation, see Fig. 4, this results in
ukh (E) − ukh (P)
ukh,ξ = , (25)
ξE − ξP
γ 1−γ
ukh,η = (uk (NE) − ukh (SE)) + (uk (N) − ukh (S)), (26)
(ηNE − ηSE ) h (ηN − ηS ) h
1652 Ch. Heinrich et al. / Mathematics and Computers in Simulation 82 (2012) 1645–1666

Fig. 4. Control volumes and variables needed to discretize the boundary integrals over the eastern face (with the exception of the western neighbor)
of control volume P (gray) with the central difference scheme (CDS). Both the control volumes and the centers of gravity are named after their
direction seen from P (e.g., E stands for ‘East’).

ukh = γukh (E) + (1 − γ)ukh (P), ph = γph (E) + (1 − γ)ph (P) (27)
with γ = (ξ e − ξ P )/(ξ E − ξ P ) and k = 1, 2.
Of course, also other interpolation schemes like the upwind difference scheme (UDS) can be employed.
By inserting these approximations into (19)–(22) and by assuming that uh and ph are constant in each control
volume, we can write the unknowns in front of the integrals. In this way, only geometric quantities remain in the
integrands. The discretized continuity equation, e.g., becomes
 1    
 uh,ij |J| yη −yξ
· · n0 ds (ξ) = 0, (28)
j∈{s,e,n,w}
u2h,ij ∂0 J
ij −xη xξ

where u1h,ij denotes u1h at the jth face of i0 .

3.1.4. Stabilization
The discretization outlined so far is nothing else than the finite volume method for the Navier–Stokes equations
defined on the parametric domain, with the midpoint rule as basic quadrature scheme and linear interpolation.
It is well-known that this scheme tends to oscillations in the pressure solution, called checkerboard pressure modes.
The reason for their occurrence can be found when considering the pressure term (20) and a control volume P with
corresponding neighbors, cf. Fig. 4. If we assume F to be the identity, for simplicity, and quadratic control volumes of
width l, we get
  
1 ph (E) − ph (W)
p̂h n ds = l . (29)
∂i 2 ph (N) − ph (S)

This formula reveals a lack of coupling of adjacent control volumes. Since we want to use a collocated (non-staggered)
grid, a stabilization term is hence required for the continuity equation. Our choice is the stabilization of “Brezzi-
Pitkäranta type” [5,13], which means that the continuity equation is replaced by
 
|J|uh · (J−T · n0 ) ds(ξ) − λ size(M)α h · (J−1 · J−T · n0 ) ds(ξ),
|J|∇p (30)
∂i0 ∂i0

where size (M) = sup diam(i ), i ∈ {1, . . . , N} . The stabilization parameters λ > 0 and α ∈ (0, 2) have to be tuned
to get both a stable and an accurate solution [13]. The definition of ph,ξ and ph,η is straightforward according to (25)
and (26) (but with p instead of u).
 Ch. Heinrich et al. / Mathematics and Computers in Simulation 82 (2012) 1645–1666 1653

3.1.5. Boundary conditions

Boundary conditions are incorporated as in the standard FVM. There, in contrast to FEM, we have to impose
boundary conditions for both physical fields. Consider, e.g.,
u=g and ∇p · n = h on 0 , (31)
i.e., a Dirichlet condition for the velocity and a Neumann condition for the pressure which are imposed on a certain
part 0 of the boundary ∂ in physical space. The type of boundary condition (31) is used to model two of the most
important boundary conditions in viscous fluid flow, an inlet condition (with h = 0) and a no-slip condition (with h = 0,
g = 0), where the fluid adheres to a wall. Assume now that the eastern face of an involved control volume lies on the
corresponding part of the boundary in parametric space, cf. Fig. 4. The flux variables on the eastern face change to

gk (F(e)) − ukh (P)

ukh,ξ = , ukh = gk (F(e)). (32)
ξ e − ξP
The quantity ukh,η can be computed exactly or approximated by finite differences, whereas ph has to be determined from
the transformed Neumann boundary condition. In Section 5, apart from the above-mentioned boundary conditions, we
also employ boundary conditions of the form
p=g and ∇u · n = h on 0 , (33)
by the help of which an outlet (with g = 0, h = 0) can be modeled.

3.2. Properties of the method

Next, we discuss some of the properties of the finite volume method on NURBS geometries. We concentrate
in particular on the conservation of physical and geometric quantities. Furthermore, we outline results concerning
existence, uniqueness and an error estimate for the approximate solution of Poisson’s equation.

3.2.1. The geometric conservation law

In connection with control volumes with curved boundaries in physical space, we have to take care of the geometric
conservation laws (GCL), a term originally coined in [27]. It stands for the requirement that the numerical scheme
should preserve a constant flow with u(ξ, t) = u∗ and p(ξ, t) = p∗ for all (ξ, t). As we deal with non-moving grids, only
the first GCL has to be considered, cf. [23].
Inserting constant solutions into the discretized Navier–Stokes equations, the time derivative term, the diffusive
term and the stabilization term vanish, and the condition that remains is

|J|J−T · n0 ds (ξ) =0
!
∀i. (34)
∂i0

This condition is equivalent to


!
n ds=0 ∀i, (35)
∂i

which means that the integrated outward pointing unit normal over each control volume has to vanish. This property
has to be preserved by the numerical scheme.
In the case of polygonal control volumes in physical space this is obviously fulfilled by using simple quadrature
rules. To be more precise, we study the error of the Gauss-Legendre quadrature
 1 
n  1
1
f (ξ) dξ − gi f (ξi ) = D2n f (γ) Ln (ξ)2 dξ, (36)
−1 (2n)! −1
i=1

where f ∈ C2n [−1, 1], ξ i are quadrature points, gi are the weights (i = 1, . . ., n), γ ∈ [−1, 1] and Ln are the Legendre
polynomials of degree n, see, e.g., [9].
1654 Ch. Heinrich et al. / Mathematics and Computers in Simulation 82 (2012) 1645–1666

In case of polygonal control volumes, the normals are constant on each face and therefore the error in (36) equals zero
for the midpoint rule, i.e., n = 1 is already enough. This does not hold in general when it comes to curved boundaries.
Consequently, we need more accurate quadrature rules.
A simple way out is to increase the number of Gauss points. In case of a geometry based on B-splines, the necessary
number of Gauss points depends on the degree p. E.g., if we consider p = 2 in both directions, we obtain linear functions
in the integrand in (34), which implies that the second derivative vanishes and therefore the midpoint rule is still exact.
But if the degree is higher, we need more Gauss points to fulfill the GCL exactly according to (36).
NURBS contain rational functions, and therefore in case of a NURBS geometry we have to increase the number of
Gauss points until the GCL is fulfilled up to a desired tolerance. Though the number of unknowns to solve for is not
affected by a change of the quadrature rule, the extra evaluations are the price to be paid if one wants to satisfy the
GCL for a NURBS geometry. The simulation results below will demonstrate that for coarse meshes, the choice of the
quadrature rule makes indeed a difference, and a more accurate rule leads to better overall convergence.
Note that in the context of isogeometric analysis new quadrature rules emerge that exploit the higher smoothness of
basis functions across element boundaries and are thus more efficient than applying Gauss quadrature on each element
separately [19]. But this technique does not apply to our case as we are interested in the integration of geometry
information over control volume edges except for the time derivative term.

3.2.2. Physical conservation

An advantage of the FVM is its inherent conservation property as opposed to the FEM, which is not conservative
in general.

Definition 3.2. A finite volume scheme is conservative, if

FK,σ = −FL,σ (37)
holds for every face σ, which is shared by two neighboring control volumes K and L. FK,σ stands for the numerical
approximation of the flux through the face σ that arises in the underlying partial differential equation.

In case of the incompressible Navier–Stokes equations and matching grids, this condition is obviously fulfilled as
it even holds for the single components of the flux
⎧
⎪
⎪ |J|uTh · J−T · n0 ds (ξ) ,
⎪
⎪
⎪
⎪
⎪
⎪
σ

⎪
⎪
⎪
⎪ h · J−1 · J−T · n0 ds (ξ) ,
⎪
⎪ −ν |J|∇u
⎪
⎪
⎪
⎪
σ
⎪
⎨  
FK,σ = |J| uh ⊗ uh · J−T · n0 ds (ξ) , (38)
⎪
⎪
⎪
⎪
σ
⎪
⎪ 
⎪1
⎪ −T
⎪
⎪ ρ σ |J|ph J · n0 ds ,
⎪ (ξ)
⎪
⎪
⎪
⎪ 
⎪
⎪
⎪
⎪ h T · J−1 · J−T · n0 ds(ξ).
⎩ λ size(M)α |J|∇p
σ
On each face we use exactly the same approximations for the boundary fluxes as in the neighboring control volume.
Therefore the proposed scheme is conservative.
Note that, in general, we are not allowed to use different Gauss rules to integrate the geometry information in
neighboring control volumes as it would lead to a non-conservative scheme. This reasoning extends immediately to
the whole mesh, which means that the same quadrature rule must be used for all interior faces.

3.2.3. Preservation of exact CAD data

As discussed in Section 2.2, the discretization presented here is based on the NURBS parameterization of the
computational domain and thus involves no error with respect to the geometry already at the coarsest mesh level.
Subsequent refinement steps preserve the geometry.
 Ch. Heinrich et al. / Mathematics and Computers in Simulation 82 (2012) 1645–1666 1655

Table 1
Matrices and vectors occurring in the semi-discretized Navier–Stokes equations.
Matrix Size Description

M R2 · ncv ×2 · ncv Mass matrix

A R2 · ncv ×2 · ncv Diffusive term
C(U) R2 · ncv ×2 · ncv Non-linear convective term
BT R2 · ncv ×ncv Pressure term
B̃ Rncv ×2 · ncv Continuity equation
D Rncv ×ncv Stabilization term
F R2 · ncv Boundary conditions
G Rncv Boundary conditions

Compared to a state-of-the-art polygonal approximation of a curved boundary, the NURBS-based approach thus
admits a very smooth treatment of all surface terms. E.g., if the functions g and h in the boundary conditions (31)
are non-linear and given on the exact boundary ∂, the transfer to a polygonal mesh inevitably leads to a piecewise
approximation with less smoothness and an interpolation error.
One might argue that these errors in the standard method are negligible if the mesh is carefully chosen and fine
enough. In many applications, this will be indeed the case. But if the flow close to the boundary is of special interest
or the overall complexity does not allow to fully resolve localized geometry features, the NURBS-based approach
represents a promising alternative. In the following section, we will study the example of fluid–structure interaction to
illustrate this point.

3.2.4. Remarks on existence, uniqueness and error estimates

To summarize further results on the approximate solution, we restrict ourselves to Poisson’s equation − u = f with
homogeneous Dirichlet boundary conditions. For simplicity, we assume orthogonal grids, i.e., a NURBS mapping such
that g12 = 0 (cf. Section 2), F ∈ C2 (0 ), and that there exist constants c1 , c2 > 0 such that c1 < g11 , g22 , |J| < c2 .
Employing the scheme for the diffusive term defined in Section 3 and treating the source term f like the time
derivative term, one can show that there exists a unique numerical solution uh . Additionally, an error estimate
||e||L2 () ≤ Csize(M) (39)
can be given, where || · ||L2 () is a discrete error norm defined in (70) below, e = u − uh is the error, and where it is
assumed that the unique variational solution satisfies u ∈ C2 (). The constant C > 0 depends on u, F and 0 , which is
in contrast to a discretization in physical space where C only depends on u and , cf. [12]. Therefore, it gets clear that
the parameterization affects the error and should be carefully chosen [32].
For brevity, we skip the corresponding proofs and remark that the approach taken in [12] can be extended by a
geometry function and adapted to our scheme.

3.3. Semi-discretized system

At the end of this section, we summarize the equations resulting from the discretization of (12) and (13) according
to the NURBS-based FVM and after incorporating the boundary conditions. Overall, the semi-discretized system can
be written as
M · U̇ + A · U + C(U) + BT · P = F, (40)
B̃ · U + D · P = G. (41)
Let ncv denote the number of control volumes in the planar case. The vectors of the unknowns for velocity and pressure
then read U = (u11 , u21 , . . . , u1ncv , u2ncv )T ∈ R2 · ncv and P = (p1 , . . . , pncv )T ∈ Rncv . Table 1 contains a short description
of all matrices and vectors in (40) and (41).
Not surprisingly, the structure of the semi-discretized equations is the same as in the classical FVM, i.e., for a
discretization in the physical domain.
1656 Ch. Heinrich et al. / Mathematics and Computers in Simulation 82 (2012) 1645–1666

We shortly comment on time discretization schemes for (40) and (41). A robust 2nd order and A-stable method is
the BDF-2 scheme. It reads in case of an ordinary differential equation ẏ = f (y)

3 (n+1) 1
y − 2y(n) + y(n−1) = tf (y(n+1) ), (42)
2 2

with time stepsize t and approximation y(n) in the nth time step. Applying this scheme to the semi-discretized
Navier–Stokes equations (40) and (41), we get

⎛ ⎞ ⎛ ⎞      
1 3 1
0 ⎠ ⎝ U(n+1) − 2U(n) + U(n−1) ⎠ C(U(n+1) ) A BT U(n+1)
⎝ tM · 2 2 + + ·
0 0 0 0 B̃ D P(n+1)
 
F(t (n+1) )
= . (43)
G(t (n+1) )

This is a non-linear system of equations that has to be solved in each time step. A canonical solution procedure is
Newton’s method
(n+1) (n+1) (n+1)
J(Xi ) Xi = −R(Xi ), (44)

(n+1) (n+1) (n+1)

Xi+1 = Xi + Xi , (45)

(n+1) (n+1) (n+1) T

where Xi = (Ui Pi ) at time step (n + 1) and iteration step i. The Jacobian reads
⎛ ⎞
  3 (n+1)
M + C (Ui ) + A BT ⎠
=⎝2 t
(n+1)
J Xi , (46)
B̃ D

(n+1)
and R(Xi ) represents the residual.
In this context, the stabilization procedure (30) shows a favorable property since the submatrix D is non-singular
due to the included boundary conditions. Therefore the index of the present differential algebraic equation reduces
from two to one and thus we can avoid the order reduction phenomenon in the pressure, cf. [33]. The BDF-2 method
is furthermore easily equipped with a step control algorithm and implemented as a predictor-corrector scheme.

4. Fluid–structure interaction

In fluid–structure interaction (FSI), the surrounding geometry of the fluid flow is determined by solving a structural
mechanics problem, and the mutual coupling between both systems demands particular attention. While the FVM is
still the method of choice in fluid dynamics, the finite element method (FEM) prevails in structural mechanics. The
coupling of both approaches along a common FSI interface requires usually so-called transfer operators that map
solution data given with respect to the fluid mesh to a representation suitable for the structural mesh, and vice versa.
Even more, due to the arbitrary Lagrangian Eulerian (ALE) description [10], an algorithmic fluid mesh comes into
play that accounts for the moving FSI interface.
We will demonstrate here that the NURBS-based FVM matches perfectly with a NURBS-based FEM, which is
also called isogeometric FEM or isogeometric analysis. In this way, it is possible both to achieve a gap-free and
non-overlapping interface and to easily transfer solution data between the different fields. This holds even in the case
of non-matching grids, which is a well-known problem in classical methods.
We again focus on the planar case in the following.
 Ch. Heinrich et al. / Mathematics and Computers in Simulation 82 (2012) 1645–1666 1657

4.1. Elasticity problem and moving mesh

The deformation of an elastic body is typically modeled in terms of the displacement field d(x, t) ∈ R2 . It satisfies
the balance equation

ρs d̈ = divσ(d) + f in s (⊂ R2 ) × [0, T ], (47)

with σ(d) denoting the stress tensor, ρs the mass density, and f the volume load.
We skip over the discussion of appropriate boundary and initial conditions as this will be made more precise below
when stating the FSI problem. To simplify the presentation, we also do not distinguish between the small displacement
and the large displacement cases, which differ in the definition of the strain and stress tensors.
The discretization of solids by means of isogeometric analysis is discussed, e.g., in [1,2,30,3], and for recent work
on shells we refer to [21]. In our context, it is important to realize that the numerical solution dh of (47) by means of
a NURBS-based Galerkin projection leads to a discrete displacement field that is a linear combination of the NURBS
Rs,ij that parameterize s (or a refinement thereof),


d(x, t) → dh (x, t) = Rs,ij (Fs (x)−1 )qij (t), (48)
i,j

with geometry function Fs of the structural domain and displacement coefficients qij ∈ R2 .
In FSI, the structure deforms according to the load exerted by the fluid, and this in turn requires adapting the fluid
mesh to the new shape. A standard way to compute the moving fluid mesh is to solve the Laplace equation

ug = 0 in f ⊂ R2 (49)

for the fluid mesh displacement ug : f → R2 , where f is the fluid domain in physical space. Zero Dirichlet boundary
conditions are set where the mesh remains fixed and combined with a prescribed coupling with the displacement of
the structure at the common interface, see below.
Even in advanced FVM, the mesh equation (49) is usually solved by means of finite elements [20]. We adopt this
viewpoint here and employ isogeometric analysis in order to obtain a new global fluid geometry function in terms
of the already available NURBS basis functions. We impose the Dirichlet boundary conditions strongly, i.e., they are
directly built into the discrete solution space as opposed to weakly imposed Dirichlet boundary conditions. In this
way it is possible to achieve a gap-free and non-overlapping interface in Section 4.4. Analogously to the displacement
problem above, the numerical solution of the mesh equation is a linear combination of the same NURBS that were
used to define the geometry. As it turns out, this approach leads to a particularly simple update procedure for the fluid
domain.
Let the solution of the stationary moving mesh problem be


Rij (Ff (x)−1 )wij
g
uh (x) = (50)
i,j

with coefficients wij ∈ R2 . The updated fluid geometry function F̃f is then computed from the old geometry function
Ff and the solution of the moving mesh equation (50) via

g
  
F̃f = Ff + uh = Rij cij + Rij wij = Rij (cij + wij ), (51)
i,j i,j i,j

which is performed by just adding the solution coefficients wij to the old control points cij .
1658 Ch. Heinrich et al. / Mathematics and Computers in Simulation 82 (2012) 1645–1666

Fig. 5. Parametric domains and physical domains in the undeformed and the deformed state of the FSI problem.

4.2. The fluid–structure interaction problem

In the following we restrict ourselves to solving a stationary FSI problem and introduce a short notation of all three
computational fields involved, see also Fig. 5:
⎧
⎪ f = Ff (0,f ),
⎪
⎪
⎨ ug = 0 in  ,
f
Mesh(ug ) : (52)
⎪
⎪ ug = 0 on ∂ \ if ,
⎪
⎩
f
˜ f = F̃f (0,f ),

⎧
⎪
⎪
1
−ν u + (u · ∇) u + ∇p = 0 in  ˜ f,
⎪
⎪
⎪
⎪ ρ
⎪
⎪
⎨ ∇ · u = 0 in  ˜ f,
Fluid(u, p) : (53)
⎪
⎪ u = uin on  ˜ ,
in
⎪
⎪
⎪
⎪ p = 0 on  ˜ out ,
⎪
⎪
⎩
u = 0 on ∂ ˜ f \ ( ˜ in ∪ 
˜ out ∪ 
˜ if ),
⎧
⎪
⎪ s = Fs (0,s ),
⎪
⎪
⎪
⎨ −div σ(d) = f in s ,
⎪
Structure(d) : d = 0 on fix,1 ∪ fix,2 , (54)
⎪
⎪
⎪
⎪ ␴ · n = 0 on  . free
⎪
⎪
⎩

˜ if , respectively. These coupling

This multiphysics system still lacks boundary conditions at the interface if or 
conditions read
u=0 ˜ if ,
on  (55)
u =d
g
on  ,if
(56)
 Ch. Heinrich et al. / Mathematics and Computers in Simulation 82 (2012) 1645–1666 1659

(pn) ◦ F̃f ◦ Ff −1 + σ · n = 0 on if , (57)

where viscous forces of the fluid have been neglected. By means of the coupling conditions we can define the interface
operators
⎧
⎪ Mesh(ug )
⎪
⎪
⎨ ug = d| if
if on 
Sf : d|if → find(u, p, ug ) : → pn (58)
⎪
⎪ Fluid(u, p)
⎪
⎩
u = 0 on  ˜ if

and

Structure(d)
Ss : pn → find d : → d|if , (59)
σ · n = − (pn) ◦ F̃f ◦ Ff −1 on if

where d|if represents the restriction of d to the interface. In the end we obtain the non-linear FSI equation
d|if = Ss (Sf (d|if )). (60)

4.3. FSI algorithm

After discretization, Eq. (60) can be solved by fixed-point iteration, cf. [22]. To keep the notation simple, we do not
mark here discretized quantities by an extra index h and write the iteration in the form
(i+1)
λc = Ss (Sf (λ(i)
c )), (61)
where λc is the corresponding vector of solution coefficients of d|if .
Depending on the ratio of the mass densities of fluid and structure, a relaxation is employed in order to accelerate
the convergence,
(i+1)
λ(i+1)
c = ω(i) λc + (1 − ω(i) )λ(i)
c . (62)
The relaxation parameter ω(i) can be set fixed or dynamic. An established choice is the Aitken adaptive relaxation
parameter
(i) (i+1) (i)
(λc − λ(i−1) )T (λc − λ(i)
c − λc + λ c
(i−1) )
ω(i) = −ω(i−1) c
(i+1) (i)
(63)
|λc − λ(i)
c − λc + λc
(i−1) |2

because it is easy to compute and efficient in practice [22]. Summing up, the Dirichlet-Neumann partitioned FSI
coupling algorithm reads:

1 Move the fluid mesh according to λ(i) c , solve the fluid equations, cf. (58),
2 Transfer of the pressure forces at the interface,
(i+1)
3 Solve the structural problem (cf. (59)) → λc ,
(i+1) (i+1)
4 If (1/ length(λc )) · |λc − λ(i)
c | <  → end,
5 Relaxation step according to (62) with Aitken relaxation (63),
6 i = i + 1 and go to 1.

4.4. Preservation of an exact and matching interface

Usually, the coupling of an FVM and an FEM solver in an FSI coupling algorithm suffers from the occurrence of
gaps or overlaps at the interface in the case of non-matching grids. One of the great advantages of our isogeometric
approach is the fact that we can overcome this problem in a straightforward way.
1660 Ch. Heinrich et al. / Mathematics and Computers in Simulation 82 (2012) 1645–1666

Fig. 6. FSI meshes (structure in black, fluid in gray) with one, two and three h-refinement steps in the fluid domain while preserving an exact
interface.

Fig. 7. FSI meshes (structure in black, fluid in gray) with one, two and three h-refinement steps in the fluid domain after the mesh has been moved
while preserving an exact interface.

We assume that the interface is oriented in ξ direction like in Fig. 5. As we always use open knot vectors, the
geometry functions Fs and Ff reduce to
s −1
n f −1
n
Fifs (ξ) = Ri,ps (ξ)cs,i and Fiff (ξ) = Ri,pf (ξ)cf,i (64)
i=0 i=0

at the interface, with ns NURBS basis functions Ri,ps (ξ) of degree ps for the structural domain and nf NURBS basis
functions Ri,pf (ξ) of degree pf for the fluid domain. The corresponding control points are cs,i and cf,i . The NURBS
bases define the spaces span{Ri,ps }i∈{0,...,ns −1} and span{Ri,pf }i∈{0,...,nf −1} .
Our aim is to establish a condition to guarantee an exact interface without gaps and overlaps even in the case of
non-matching grids. A naive approach would be to require that Fifs = Fiff . If we look a bit more into the details in the
coupling algorithm, however, we recognize that this is not sufficient. The key point is to adapt the fluid mesh exactly to
the solution of the structural equations. As the corresponding solution coefficients in (48) can be arbitrary, it becomes
clear that the condition to ensure an exact interface is

span{Rpi s }i∈{0,...,ns −1} ⊆ span{Rpi f }i∈{0,...,nf −1} . (65)

A simple possibility to ensure condition (65) is to start with a matching interface in the sense that in both fields the
knots, the degree and the weights are the same. Thereafter, several uniform h-refinement steps are performed in the fluid
domain as there mostly a higher resolution is required. By construction, this refinement preserves an exact interface
(cf. Fig. 6) and satisfies (65).
An additional advantage of this approach are the possible savings in the moving mesh algorithm. Here, it is sufficient
to use the first matching coarse grid of the fluid domain throughout the computation. Thus, (49) is always solved on
the coarse starting grid, and only in the update (51) the grid data are lifted to the fine grid through the new geometry
function F̃f , cf. Fig. 7. In the end, the benefit of this isogeometric coupling algorithm is twofold.

5. Numerical examples

In this section, we present numerical test cases that illustrate the FVM on NURBS geometries.
 Ch. Heinrich et al. / Mathematics and Computers in Simulation 82 (2012) 1645–1666 1661

Fig. 8. NURBS geometry with two quarters of exact circles as boundary parts.

Table 2
Control points cij and weights wij .

i ci0 ci1 ci2 wi0 wi1 wi2

0 (0, 0.5) (0, 0.75) (0, 1) 1√ 1

√ 1
√
1 (0.5, 0.5) (0.75, 0.75) (1, 1) 2/2 2/2 2/2
2 (0.5, 0) (0.75, 0) (1, 0) 1 1 1

5.1. Incompressible Navier–Stokes

In order to show the convergence behavior of the method described in Section 3, we examine the analytical solution
u1 (x, y) = sin(πx), u2 (x, y) = −π cos(πx)y and p(x, y) = sin(πxy) (66)
in the domain depicted in Fig. 8, which is a NURBS surface defined by the knot vectors
ξ = (0, 0, 0, 1, 1, 1) and η = (0, 0, 0, 1, 1, 1). (67)
The control points and weights are given in Table 2.
The curved boundaries are quarters of exact circles with radii 0.5 and 1.
As one can see, the velocity solution is divergence-free and the boundary conditions have to be set according to
the analytical solution (66). The corresponding source term can be computed by inserting the given analytical solution
into the momentum equation (12)
⎛ ⎞
νπ 2 sin(πx) + sin(πx) cos(πx)π + 1 cos(πxy)πy
1 ⎜ ρ ⎟
−ν u + (u · ∇)u + ∇p = ⎜ ⎝
⎟.
⎠ (68)
ρ 1
−νπ cos(πx)y + π y + cos(πxy)πx
3 2
ρ
For the simulation we use the parameters ρ = 1 and ν = 0.03. Figs. 9 and 10 show the convergence behavior of each
variable u1 , u2 and p. We distinguish between an integration of the geometry information with one Gauss point (dashed
lines) and with five Gauss points (solid lines) and compare the corresponding L2 error norm

 nCV 

||g − gh ||L2 (),cont =  |g − gh |2 · |J| dξ, (69)
i=1 i0

where g denotes the analytical solution of the physical magnitude and gh its numerical approximation. A comparison
based on the discrete L2 error norm

 nCV

||g − gh ||L2 (),disc =  |i0 | · (g(F(Pi )) − gh (F(Pi )))2 |J(Pi )|, (70)
i=1
1662 Ch. Heinrich et al. / Mathematics and Computers in Simulation 82 (2012) 1645–1666

Discrete and continuous L2 error for the u1 velocity Discrete and continuous L2 error for the u2 velocity
0 0
10 10

−1 −1
10 10
L error

L2 error
−2 −2
10 10
2

−3 −3
10 10
1 Gp cont 1 Gp cont
5 Gp cont 5 Gp cont
1 Gp disc 1 Gp disc
−4 5 Gp disc −4 5 Gp disc
10 10
−2 −1 −2 −1
10 10 10 10
mesh size mesh size

Fig. 9. Discrete and continuous L2 error for velocity components u1 (left) and u2 (right), integration with one and with five Gauss points.

Discrete and continuous L2 error for the pressure

0
10

−1
10
L2 error

−2
10

1 Gp cont
5 Gp cont
−3 1 Gp disc
10
5 Gp disc
−2 −1
10 10
mesh size

Fig. 10. Discrete and continuous L2 error of p for an integration with one and with five Gauss points.

which is just the evaluation of (69) with the midpoint rule at the centers of gravity in parametric space Pi , is also
included in Figs. 9 and 10.
Observe that for all fields a significantly lower error is obtained for coarse meshes if the five point Gauss rule is
applied instead of the one-point rule. For finer meshes the difference tends to zero. The same issue is illustrated by
computing the GCL. For the coarsest grid with mesh size 0.25 in parametric space, the error is 7.87 × 10−3 in each cell
with one-point quadrature as opposed to an error of 3.34 × 10−10 for five-point quadrature. The finer the grid becomes
the smaller the error of the GCL gets and consequently only small differences in the numerical solution of the fluid
problem have to be expected. We stress that the additional effort for the five-point rule is small with respect to the
overall computational costs and gets smaller the finer the mesh becomes (about 1% for this example and the finest
mesh). Furthermore, the additional evaluations of geometric quantities can be done efficiently in a pre-processing step.
This example also shows the convergence rates two in the discrete and one in the continuous norm, respectively.

5.2. Flow past a cylinder

The second test example is the flow past a cylinder, parameterized as a single patch with knot vectors

ξ = (0, 0, 0, 0.5, 1, 1.5, 2, 2.5, 3, 3.5, 4, 4, 4) and η = (0, 0, 0, 1, 1, 1). (71)

The control points and the weights are shown in Table 3.

 Ch. Heinrich et al. / Mathematics and Computers in Simulation 82 (2012) 1645–1666 1663

Table 3
Control points cij and weights wij of flow past a cylinder.

i ci0 ci1 ci2 wi0 wi1 wi2

0 (− 0.04, 0) (− 0.025, 0) (− 0.01, 0)

√ 1 1 1 √
1 (− 0.04, 0.04) (− 0.025, 0.0075) √ ( 2 − 1)/100)
(−0.01, 1 1 (1 + 1/√2)/2
2 (− 0.04, 0.04) (− 0.0075, 0.025) √− 2)/100, 0.01)
((1 1 1 (1 + 1/√2)/2
3 (0.12, 0.04) (0.0075, 0.025) (( 2 −√ 1)/100, 0.01) 1 1 (1 + 1/√2)/2
4 (0.12, 0.04) (0.025, 0.0075) (0.01, ( 2 −√1)/100 1 1 (1 + 1/√2)/2
5 (0.12, − 0.04) (0.025, − 0.0075) (0.01,
√ (1 − 2)/100) 1 1 (1 + 1/√2)/2
6 (0.12, − 0.04) (0.0075, − 0.025) (( 2 −√1)/100, −0.01) 1 1 (1 + 1/√2)/2
7 (− 0.04, − 0.04) (− 0.0075, − 0.025) ((1 − 2)/100, √ −0.01) 1 1 (1 + 1/√2)/2
8 (− 0.04, − 0.04) (− 0.025, − 0.0075) (−0.01, (1 − 2)/100) 1 1 (1 + 1/ 2)/2
9 (− 0.04, 0) (− 0.025, 0) (− 0.01, 0) 1 1 1

Fig. 11. Coarsest grid of ‘Flow past a cylinder’ with boundary conditions (left) and mesh after three uniform h-refinement steps (right).

Fig. 12. Pressure solution with contour lines (left) and u1 velocity in x-direction (right).

Fig. 11 displays the geometry and the coarsest grid defined by the NURBS surface on the left and the mesh after
three uniform h-refinement steps on the right. We impose a parabolic inlet profile at the left end with a maximum
velocity of 0.1. At the right end an outlet condition is set and a ‘no-slip’ condition on all other boundaries. The material
parameters are ν = 2 ×10−5 and ρ = 1, which leads to a Reynolds number of about 100.
In Fig. 12 on the left the pressure solution is shown and on the right the solution of the velocity in x-direction. We
compare our result of the latter solution with a reference solution along the white line in Fig. 12 on the right from
x = 0.01 to x = 0.08, where y = 0 is fixed. Fig. 13 depicts the result achieved with our method and 98,304 degrees of
freedom and the result computed with COMSOL [7] and linear finite elements with 311,976 degrees of freedom and
GLS stabilization. We can observe that the solution of the FVM simulation is in good agreement with the solution of
the COMSOL package.

5.3. FSI example

Finally, we apply the FSI coupling algorithm of Section 4 to the problem shown in Fig. 14. The fluid domain is
defined by two exact circles of radii 0.05 and 0.1, and the outer wall is modeled as a linear elasticity problem.
1664 Ch. Heinrich et al. / Mathematics and Computers in Simulation 82 (2012) 1645–1666

Comparison of u velocity
1
0.02

0.015 NURBS FVM

COMSOL
0.01

0.005

u1 velocity
0

−0.005

−0.01

−0.015

−0.02

−0.025
0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08
x−coordinate

Fig. 13. Comparison of the u1 velocity with a reference solution, computed with COMSOL, at the position of the white line in Fig. 12.

Fig. 14. Geometry of FSI problem: on the fluid domain (light gray) a parabolic inlet profile is imposed on the left and an outlet condition at the
bottom. The structure (dark gray) is fixed at both ends.

We impose a parabolic inlet profile with a maximum velocity of 0.01 on the left and an outlet condition (i.e. p = 0)
at the bottom. The kinematic viscosity is ν = 10−6 and the mass density ρ = 1000, which leads to a Reynolds number of
500. The structure (dark gray) is fixed at both ends and deforms according to the pressure load exerted by the fluid at
the common interface. Its thickness of 0.001 is exaggerated in Fig. 14 for purposes of demonstration. Young’s modulus
of the structure is set to 5000 and Poisson’s ratio is 0.25.
Fig. 15 depicts the total velocity of the fluid and the total displacement of the structure in the converged state using
66,576 degrees of freedom. In order to validate our results we simulated the same problem by means of COMSOL,
employing the same physical models for fluid and structure and 203,149 degrees of freedom. We compare the pressure
distribution at the interface in Fig. 16 on the left and the magnitude of the displacement of the interface on the right.
Both images demonstrate that our results are in good agreement with those achieved by COMSOL.

Fig. 15. Total velocity of the fluid and total displacement of the structure, converged state.
 Ch. Heinrich et al. / Mathematics and Computers in Simulation 82 (2012) 1645–1666 1665

pressure distribution at the interface −3

x 10 displacement at the interface
0.035 3
Isogeometric
0.03 NURBS FVM COMSOL
2.5
COMSOL

0.025

total displacement
2
pressure

0.02
1.5
0.015

1
0.01

0.5
0.005

0 0
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
arc−length arc−length

Fig. 16. Comparison of pressure (left) and magnitude of displacement (right) at the interface between our solution and the result by COMSOL.

6. Conclusions

In this paper we have introduced a finite volume method on geometries parameterized by NURBS and discussed
the discretization for the incompressible Navier–Stokes equations. As the computational grid is inherently defined
by the knot vectors of the NURBS parameterization, a mesh generation step can be omitted, which is an attractive
feature. Furthermore, this approach is able to preserve geometrical and physical properties, in particular the geometry
of free-form surfaces as well as the key feature of the FVM, the conservativity.
In addition, we have combined the FVM on NURBS geometries with a structural solver based on isogeometric finite
elements in a partitioned fluid–structure interaction coupling algorithm. While classical methods typically suffer from
the occurrence of gaps and overlaps at the interface if different mesh resolutions are used in the single fields, this problem
can be overcome in the isogeometric setting. A sufficient condition that guarantees a gap-free and non-overlapping
interface even in the case of non-matching grids has been presented.
More complex geometries stemming from real-life problems often cannot be parameterized by just a single patch.
Thus, the extension to the multiple patch case, which additionally allows for local refinement to some degree, definitely
deserves attention and will be addressed in our future work. We will furthermore investigate the influence of the
parameterization on the convergence behavior of the numerical error. First results indicate that this influence can be
significant.

Acknowledgments

The authors were supported by the 7th Framework Programme of the European Union, Project SCP8-218536
“EXCITING”. Furthermore, the support by the TUM Graduate School is gratefully acknowledged.

References

[1] M. Aigner, C. Heinrich, B. Jüttler, E. Pilgerstorfer, B. Simeon, A.-V. Vuong, Swept volume parameterization for isogeometric analysis, in: E.
Hancock, R. Martin (Eds.), The Mathematics of Surfaces (MoS XIII 2009), Springer, 2009.
[2] Y. Bazilevs, V. Calo, T. Hughes, Y. Zhang, Isogeometric fluid–structure interaction: theory, algorithms and computations, Computational
Mechanics 43 (2008) 3–37.
[3] Y. Bazilevs, V. Calo, Y. Zhang, T. Hughes, Isogeometric fluid–structure interaction analysis with applications to arterial blood flow, Computa-
tional Mechanics 38 (2006) 310–322.
[4] M. Borden, M. Scott, J. Evans, T. Hughes, Isogeometric finite element data structures based on Bézier extraction of NURBS, International
Journal for Numerical Methods in Engineering 87 (2011) 15–47.
[5] F. Brezzi, J. Pitkäranta, On the stabilization of finite element approximations of the Stokes equations, in: W. Hackbusch (Ed.), Efficient Solutions
of Elliptic Systems, Vieweg, Braunschweig, 1984, pp. 11–19.
1666 Ch. Heinrich et al. / Mathematics and Computers in Simulation 82 (2012) 1645–1666

[6] A. Buffa, C. de Falco, G. Sangalli, Isogeometric analysis: stable elements for the 2D stokes equation, International Journal for Numerical
Methods in Fluids 65 (2011) 1407–1422.
[7] COMSOL Multiphysics User’s Guide, version 3.5a, COMSOL AB., 2009.
[8] J.A. Cottrell, T.J. Hughes, Y. Bazilevs, Isogeometric Analysis, Toward Integration of CAD and FEA, Wiley, 2009.
[9] P. Davis, P. Rabinowitz, Methods of Numerical Integration, Academic Press, New York, 1975.
[10] J. Donea, A. Huerta, J.-P. Ponthot, A. Rodríguez-Ferran, Encyclopedia of Computational Mechanics, chapter Arbitrary Langrangian–Eulerian
Methods, Wiley, 2004.
[11] J. Droniou, R. Eymard, A mixed finite volume scheme for anisotropic diffusion problems on any grid, Numerische Mathematik 105 (2006)
35–71.
[12] R. Eymard, T. Gallouët, R. Herbin, Finite volume methods, in Handbook of Numerical Analysis, vol. VII, North-Holland, 2000, pp. 713–1020.
[13] R. Eymard, R. Herbin, J. Latché, Convergence Analysis of a Colocated Finite Volume Scheme for the Incompressible Navier–Stokes Equations
on General 2D or 3D Meshes, SIAM Journal on Numerical Analysis 45 (2007) 1–36.
[14] G. Farin, Curves and Surfaces for CAGD: A Practical Guide, 5th edn., Morgan Kaufmann, 2002.
[15] M. Feistauer, J. Felcman, I. Straškraba, Mathematical and Computational Methods for Compressible Flow, Oxford University Press, USA,
2003.
[16] J. Ferziger, M. Perić, Computational Methods for Fluid Dynamics, 3rd edn., Springer, 2002.
[17] C. Fletcher, Computational Techniques for Fluid Dynamics, vol. 2, 2nd edn., Springer, 1997.
[18] T. Hughes, J. Cottrell, Y. Bazilevs, Isogeometric analysis: CAD, finite elements, NURBS, exact geometry and mesh refinement, Computer
Methods in Applied Mechanics and Engineering 194 (2005) 4135–4195.
[19] T. Hughes, A. Reali, G. Sangalli, Efficient quadrature for NURBS-based isogeometric analysis, Computer Methods in Applied Mechanics and
Engineering 199 (2010) 301–313.
[20] H. Jasak, Z. Tukovic, Automatic mesh motion for the unstructured finite volume method, Transactions of FAMENA 30 (2007).
[21] J. Kiendl, K.-U. Bletzinger, J. Linhard, R. Wüchner, Isogeometric shell analysis with Kirchhoff-Love elements, Computer Methods in Applied
Mechanics and Engineering 198 (2009) 3902–3914.
[22] U. Küttler, W.A. Wall, Fixed-point fluid–structure interaction solvers with dynamic relaxation, Computational Mechanics 43 (2008) 61–72.
[23] P.P. Lamby, Parametric Multi-block Grid Generation and Application to Adaptive Flow Simulations, Ph.D. thesis, RWTH Aachen, 2007.
[24] B.Q. Li, Discontinuous Finite Elements in Fluid Dynamics and Heat Transfer, Springer Verlag, 2006.
[25] V.D. Liseikin, Grid Generation Methods, Springer, 1999.
[26] L. Piegl, W. Tiller, The NURBS Book, 2nd edn., Springer, 1996.
[27] P. Thomas, C. Lombard, Geometric conservation law and its application to flow computations on moving grids, AIAA 17 (1979) 1030–1037.
[28] A.-V. Vuong, C. Giannelli, B. Jüttler, B. Simeon, A hierarchical approach to local refinement in isogeometric analysis, Computer Methods in
Applied Mechanics and Engineering 200 (2011) 3554–3567.
[29] A.-V. Vuong, C. Heinrich, B. Simeon, ISOGAT: A 2D tutorial MATLAB code for Isogeometric Analysis, Computer Aided Geometric Design
27 (2010) 644–655.
[30] A.-V. Vuong, B. Simeon, On isogeometric analysis and its usage for stress calculation, in: H.G. Bock, X.P. Hoang, R. Rannacher, J.P. Schlöder
(Eds.), Modeling, Simulation and Optimization of Complex Processes, Springer, Berlin Heidelberg, 2012, pp. 305–314.
[31] W.A. Wall, Fluid–Struktur-Interaktion mit stabilisierten Finiten Elementen, Ph.D. thesis, Universität Stuttgart, 1999.
[32] G. Xu, B. Mourrain, R. Duvigneau, A. Galligo, Parameterization of computational domain in isogeometric analysis: methods and comparison,
Computer Methods in Applied Mechanics and Engineering 200 (2011) 2021–2031.
[33] Z. Zheng, L. Petzold, Runge–Kutta–Chebyshev projection method, Journal of Computational Physics 219 (2006) 976–991.