Comput Mech (2008) 41:371–378

DOI 10.1007/s00466-007-0193-7

ORIGINAL PAPER

The role of continuity in residual-based variational multiscale

modeling of turbulence
I. Akkerman · Y. Bazilevs · V. M. Calo ·
T. J. R. Hughes · S. Hulshoff

Received: 10 May 2007 / Accepted: 16 May 2007 / Published online: 19 June 2007
© Springer-Verlag 2007

Abstract This paper examines the role of continuity of We also find that the effect of continuity is greater for higher
the basis in the computation of turbulent flows. We compare Reynolds number flows.
standard finite elements and non-uniform rational B-splines
(NURBS) discretizations that are employed in Isogeometric Keywords Incompressible flows · Finite elements ·
Analysis (Hughes et al. in Comput Methods Appl Mech Eng, NURBS · Navier–Stokes equations · Boundary layers ·
194:4135–4195, 2005). We make use of quadratic discreti- Turbulent channel flows · Residual-based turbulence
zations that are C 0 -continuous across element boundaries in modeling · Isogeometric Analysis · Continuity of
standard finite elements, and C 1 -continuous in the case of discretization · Variational multiscale formulation
NURBS. The variational multiscale residual-based method
(Bazilevs in Isogeometric analysis of turbulence and fluid-
structure interaction, PhD thesis, ICES, UT Austin, 2006;
Bazilevs et al. in Comput Methods Appl Mech Eng, sub- 1 Introduction
mitted, 2007; Calo in Residual-based multiscale turbulence
modeling: finite volume simulation of bypass transition. PhD Due to the complicated multiscale nature of turbulence,
thesis, Department of Civil and Environmental Engineering, numerical simulation of turbulent flows remains a signifi-
Stanford University, 2004; Hughes et al. in proceedings of cant computational challenge to this day. A community of
the XXI international congress of theoretical and applied researchers focused on understanding fundamental physical
mechanics (IUTAM), Kluwer, 2004; Scovazzi in Multiscale behavior of turbulence through numerical simulation typ-
methods in science and engineering, PhD thesis, Department ically makes use of very simple geometries and high-order
of Mechanical Engineering, Stanford Universty, 2004) spectral or compact finite difference methods (see, e.g.,
is employed as a turbulence modeling technique. We [32,34]). The underlying function spaces utilized in spectral
find that C 1 -continuous discretizations outperform their methods are of high continuity (C ∞ in the cases of Fourier
C 0 -continuous counterparts on a per-degree-of-freedom basis. series and global polynomials). On the other hand, turbulent
flows are also of great interest in general geometry industrial
applications. These are typically computed using finite vol-
ume and finite element methods, which employ low-order
I. Akkerman · S. Hulshoff approximation functions that are at most C 0 -continuous.
Department of Aerospace Engineering, Recently, Hughes et al. [25] introduced a new computa-
Delft University of Technology,
Kluyverweg 1, 2629HS Delft,
tional technique termed Isogeometric Analysis in an attempt
The Netherlands to generalize and improve on the finite element method in
the areas of geometry and solution representation, as well as
Y. Bazilevs (B) · V. M. Calo · T. J. R. Hughes mesh refinement. Isogeometric analysis is based on the geo-
Institute for Computational Engineering and Sciences,
metric primitives of computer graphics and computer-aided
The University of Texas at Austin, 201 East 24th Street,
1 University Station C0200, Austin, TX 78712, USA design systems. The first instantiation of isogeometric anal-
e-mail: [email protected] ysis made use of non-uniform rational B-splines (NURBS,

123
372 Comput Mech (2008) 41:371–378

see [16,36,37]). NURBS-based isogeometric analysis was on residual-based approaches [2,6,12,24], the conservative
successfully applied to solve problems of vascular fluid- form was employed.
structure interaction in [5]. Accuracy of the new methodol- The paper is organized as follows. In Sect. 2, we state
ogy for structural vibrations was shown to be superior to that the weak formulation of the incompressible Navier–Stokes
of standard finite elements in [17]. A mathematical theory of equations at the continuous level. In Sect. 3, we give the semi-
NURBS-based isogeometric analysis was put forth in [4]. A discrete, residual-based variational multiscale formulation
distinguishing feature of isogeometric analysis is so-called of the incompressible Navier–Stokes equations in advective
k-refinement, in which the order of the functions is increased form. In Sect. 4, we present numerical results for equilib-
together with their continuity. As a result, isogeometric anal- rium turbulent channel flows at friction-velocity Reynolds
ysis allows for higher-order and higher-continuity discreti- numbers 180 and 590. We compare C 0 - and C 1 -continuous
zations on complex geometries,1 and may be thought of as quadratic elements and assess accuracy on the basis of the
bridging the gap existing between the procedures employed number of degrees of freedom. High-fidelity, direct numer-
in the flow-physics and industrial-flow communities. ical simulation (DNS) results of [35] are used as reference
A variational multiscale, residual-based turbulence mod- solutions. In Sect. 5, we draw conclusions.
eling framework has recently emerged as a new concept in
large-eddy simulation (LES) of turbulent flows (see [2,6,12,
24,38]). In this methodology one obtains an exact equation 2 Incompressible Navier–Stokes equations
for the coarse, or resolved, scales and their explicit depen-
dence on the fine, or unresolved, scales, and the modeling We begin by considering a weak formulation of the incom-
task amounts to accurately representing the fine scales in the pressible Navier–Stokes equations. Let V denote the trial
coarse-scale equations. In this paper, inspired by the theory of solution and weighting function spaces, which
stabilized methods, we employ simple algebraic models for
are assumed to
be the same. We also assume u = 0 on
and  p(t) d = 0
the fine scales, which involve appropriately-scaled residuals for all t ∈ ]0, T [. The variational formulation is stated as fol-
of the underlying partial differential equations. For a sum- lows: Find U = {u, p} ∈ V such that ∀W = {w, q} ∈ V,
mary of the early literature on stabilized methods see Brooks
and Hughes [10]. Recent work on stabilized methods is pre- B(W , U) = (W , F) (1)
sented in [1,8,9,11,14,18–21,28,33,41–43].
Spectral studies of simple advective and diffusive where
model problems in [6] indicated better accuracy for the  
∂u
C 1 -continuous case, especially in the upper part of the spec- B(W , U) = w, − (∇w, u ⊗ u) + (q, ∇ · u)
∂t 
trum. NURBS-based isogeometric analysis, in conjunction  
with the multiscale approach, was applied to turbulent flow − (∇ · w, p) + ∇ s w, 2ν∇ s u  , (2)
computations in [2]. Preliminary results indicated better accu-
racy for higher-order and higher-continuity discretizations. It (W , F) = (w, f ) , (3)
was noted in [2] that when going from a C 0 -continuous lin-
ear to a C 1 -continuous quadratic B-spline basis an increase and
in solution accuracy was obtained. The question this raises
1 
is whether it was the order or continuity of the basis func- ∇s u = ∇u + ∇u T . (4)
2
tions that led to superior accuracy. It is the primary goal of
this paper to answer this question. It should be noted that f is the force (per unit mass), ν is the kinematic viscosity
good accuracy of B-spline discretizations for wall-bounded and p is the pressure divided by the density.
turbulent flows was also noted in [29–31,40]. Variational equations (1)–(3) imply weak satisfaction of
The variational multiscale formulation employed in this the linear momentum equations and incompressibility con-
study utilized the “advective form” of the convection term straint, namely
rather than the integrated by parts “conservative form”. The
∂u
advective form is often employed in finite element flow codes, +∇ ·(u ⊗ u) + ∇ p − ∇ · (2ν∇ s u) − f = 0 in , (5)
perhaps more so than the conservative form. In earlier works ∂t
∇ · u = 0 in . (6)

1
Note that one may use the incompressibility constraint to
It is important to note that the highest possible order of continuity of
the solution space in a NURBS-based isogeometric analysis is limited simplify the momentum equation as
to the continuity of the basis used in the definition of the geometrical
∂u
domain of interest. Pure k-refinement with maximal smoothness is only + u · ∇u + ∇ p − νu − f = 0 in . (7)
attainable in simple geometries. ∂t

123
Comput Mech (2008) 41:371–378 373

3 Multiscale residual-based formulation ∇ · (uh + u ) = 0, we compute:

of the incompressible Navier–Stokes equations  
employing the advective form − ∇w h , uh ⊗ uh − ∇w h , uh ⊗ u
   
  
We consider a direct-sum decomposition of V into “coarse- − ∇w , u ⊗ u
h h
− ∇w , u ⊗ uh
 
scale” and “fine-scale” subspaces, V h and V  , respectively, 

= − ∇w , u ⊗ (u + u )
h h h

  

V = Vh ⊕ V (8) − ∇w , u ⊗ u
h h
− ∇w h , u ⊗ u
 


= w , (u + u ) · ∇u
h h h

 
V h is assumed to be a finite-dimensional space, which will 
− ∇w , u ⊗ u
h h
− ∇w h , u ⊗ u (11)
be identified later with the space of functions with which we  
actually compute. One obtains a unique decomposition in
At this point we assume that  is partitioned into a set of
(8) with the aid of a linear projection operator P, that gives
subdomains, such as finite elements or NURBS elements,
U h = PU ∈ V h and U  = (I − P)U ∈ V  from a given
and on this partition we have a finite dimensional space of
U ∈ V (see [23] for details).
functions with local support that forms our approximation
By restricting the weighting space to V h in (1) and employ-
space for U h and W h . Let x = {xi }i=1
d , denote the coordi-
ing the direct-sum decomposition (8) for the solution space,
nates of element K in physical space, and let ξ = {ξi }i=1
d ,
we obtain the equation system for the large scales, namely:
 denote the coordinates of element K̂ in parametric space. Let
Find U h = {uh , p h } ∈ V h such that ∀W h = w h , q h ∈ V h ,
x = x(ξ ) : K̂ → K be a continuously differentiable map
with a continuously differentiable inverse.
   
B Wh, Uh + U = Wh, F . (9) We model the fine scales as in [6]:

U  ≈ −τ R(U h ), (12)
(9) indicates precisely the manner in which the large scales
depend on U  = {u , p  }. where τ is a 4 × 4 matrix (in three spatial dimensions) and
Combining (2) and (9), we obtain: Find U h ∈ V h , such R(U h ) is a 4 × 1 vector that collects momentum and conti-
that ∀W h ∈ V h , nuity residuals of the Navier–Stokes equations,

T
  R(U h ) = r TM (uh , p h ), rC (uh ) , (13)
∂ uh  
wh , − ∇w h , uh ⊗ uh + q h , ∇ · uh
∂t    in which
  
− ∇ · w h , p h + ∇ s w h , 2ν∇ s uh − w h , f  ∂ uh
   r M uh , p h = +uh ·∇uh + ∇ p h − νuh − f , (14)
    ∂t
∂ u rC (uh ) = ∇ · uh
+ w , h
− ∇w h , uh ⊗ u − ∇w h , u ⊗ uh (15)
∂t   
  We define τ as follows:
− ∇w h , u ⊗ u + q h , ∇ · u
 
  τ = diag(τ M , τ M , τ M , τC ), (16)
− ∇ · wh , p + ∇ s w h , 2ν∇ s u =0 (10)
 
where
 −1/2
For purposes of modeling the 4
 fine scales we make the simpli- τM = + uh · Guh + C I ν 2 G : G , (17)
∂ u t 2
fying assumption w , ∂t
h = 0. We note, however, that it

has been shown in [15] that it is beneficial to incorporate this

effect in modeling the fine scales. The term ∇ s w h , 2∇ s u  τC = (g · τ M g)−1 , (18)
may be omitted by selecting a projector that enforces the
orthogonality of the coarse and fine scales in the semi-norm with G a second rank metric tensor
induced by this term (see, e.g., [2]).
We turn our attention to the convective terms in (10). ∂ξ T ∂ξ
G= , (19)
Assuming incompressibility of the velocity field, namely, ∂x ∂x

123
374 Comput Mech (2008) 41:371–378
 
∂ξ
and g a vector obtained from the column sums of ∂ x , + ∇ s w h , 2ν∇ s uh − w h , f
 
 
g = {gi } + u · ∇w + ∇q , τ M r M
h h h
+ ∇ · w h , τC ∇ · uh
d     
∂ξ
gi = . (20) − ∇w , τ M r M ⊗ τ M r M
h
= 0. (21)
∂ x ji 
j=1

The definition of τ M in (17) is inspired by the theory of

stabilized methods for advection-diffusion-reaction systems
(see, e.g., Shakib et al. [39], Hughes and Mallet [22]). The 4 Numerical experiments for turbulent channel flow
definition of τC comes from the small-scale Shur comple-
ment operator for the pressure (see [2] for a details). In the 4.1 Problem setup
definition of τ M (17), C I is a positive constant, independent
of the mesh size, that derives from the element-wise inverse To examine the effects of continuity, we compute turbulent
estimate (see, e.g., [27]). channel flows at Reynolds numbers Reτ = 180 and Reτ =
Combining equations (10)–(12), we obtain our discrete 590, based on the friction velocity and the channel half width.
formulation: Find U h ∈ V h , such that ∀W h ∈ V h , To assess the accuracy of the calculations, we compare with
  the DNS of [35].
∂ uh  The problem setup is as follows. The computational
wh , + w h , (uh − τ M r M ) · ∇uh
∂t   domain is a rectangular box, and the flow is driven by a
 
+ q h , ∇ · uh − ∇ · w h , p h
 
25
0 2
C1 P2 32x33x32
1 C P 32x34x32
20 DNS

15
U+

10

0
0 1 10 100
0 1 2 3 4 5 6 y+
(a) C 0 -continuous quadratic elements (a) Mean stream-wise velocity

4.5
0 2
4 C1 P2 32x33x32
C P 32x34x32
3.5 DNS
3
2.5
u+

2
1.5
1
0.5
0
0 1 2 3 4 5 6 0
1 0 20 40 60 80 100 120 140 160 180
(b) C -continuous quadratic NURBS
y+
Fig. 1 Basis functions employed in homogeneous directions. Basis (b) Stream-wise velocity fluctuations
functions are shown in a univariate setting. Three-dimensional basis
functions are obtained by taking tensor products of one-dimensional Fig. 2 Turbulent channel flow at Reτ = 180 computed on a 323
basis functions. The meshes are stretched in the wall-normal direction element mesh. Comparison of C 0 - versus C 1 -continuous discretizations

123
Comput Mech (2008) 41:371–378 375

1.4 30
C01 P22 32x33x32 0 2
C P 32x33x32
1.2 C P 32x34x32 1 2
25 C P 32x34x32
DNS DNS
1
20
0.8
v+

U+
15
0.6

0.4 10

0.2 5

0 0
0 20 40 60 80 100 120 140 160 180
y+ 1 10 100
(a) Wall-normal velocity fluctuations y+
(a) Mean stream-wise velocity
1.4
C0 P2 32x33x32 6
1.2 C1 P2 32x34x32 0 2
DNS C P 32x33x32
5 1 2
1 C P 32x34x32
DNS
0.8 4
w+

0.6
u+
3
0.4
2
0.2
1
0
0 20 40 60 80 100 120 140 160 180
y+ 0
0 100 200 300 400 500
(b) Span-wise velocity fluctuations
y+
Fig. 3 Turbulent channel flow at Reτ = 180 computed on a 323 (b) Stream-wise velocity fluctuations
element mesh. Comparison of C 0 - versus C 1 -continuous discretizations
Fig. 4 Turbulent channel flow at Reτ = 590 computed on a 323
element mesh. Comparison of C 0 - versus C 1 -continuous discretizations

constant pressure gradient in the stream-wise direction. Peri-

odic boundary conditions are imposed in the stream-wise and
span-wise directions, commonly referred to as homogeneous hyperbolic function to cluster points near the wall. More-
directions. A no-slip boundary condition is applied at the over, in the definition of τ M (17) we set C I to 36.
walls. The no-slip boundary condition is enforced strongly, Numerical results for all cases are reported in the form of
that is, the discrete velocity is set to zero at the walls. An alter- statistics of the mean velocity and root-mean-square velocity
native approach is to enforce Dirichlet boundary conditions fluctuations. The statistics were computed by sampling the
weakly. This is accomplished by appropriately augmenting velocity field at the mesh knots and averaging the solution in
the semi-discrete equations (21) by terms that enforce the time as well as in the homogeneous directions. The meshes
no-slip condition weakly (see [3,7] for details). Although were chosen such that the number of degrees of freedom for
the latter approach was shown to be superior to the former, both quadratic discretizations are approximately the same.
we did not employ it in the computations presented herein. All computational results are presented in non-dimensional
We employ quadratic finite elements that are C 0 -continu- wall units.
ous, and quadratic NURBS elements that are C 1 -continuous
across element interfaces. The basis functions utilized in our 4.2 Turbulent channel flow at Reτ = 180
computations are shown in Fig. 1.
The semi-discrete equations (21) are advanced in time The domain size is 2π , 2, and 4/3π in the stream-wise, wall-
using the generalized-α method [13,26]. We use meshes that normal, and span-wise directions, respectively. The corre-
are stretched in the wall-normal direction according to a sponding DNS computation was carried out on a domain of

123
376 Comput Mech (2008) 41:371–378

1.4 30
0 2 0 2
C P 32x33x32 C P 64x65x64
1 2 1 2
1.2 C P 32x34x32 25 C P 64x66x64
DNS DNS
1 20
0.8

U+
v+

15
0.6
10
0.4
5
0.2
0
0 1 10 100
0 100 200 300 400 500 y+
y+ (a) Mean stream-wise velocity
(a) Wall-normal velocity fluctuations
6
0 2
1.8 C P 64x65x64
1 2
C
0
P
2
32x33x32 5 C P 64x66x64
1.6 1 2 DNS
C P 32x34x32
1.4 DNS 4
1.2
u+
3
1
w+

0.8 2

0.6
1
0.4
0.2 0
0 100 200 300 400 500
0 y+
0 100 200 300 400 500
y+
(b) Stream-wise velocity fluctuations

(b) Span-wise velocity fluctuations Fig. 6 Turbulent channel flow at Reτ = 590 computed on a 643 ele-
ment mesh. Comparison of C 0 - versus C 1 -continuous discretizations
Fig. 5 Turbulent channel flow at Reτ = 590 computed on a 323
element mesh. Comparison of C 0 - versus C 1 -continuous discretizations
4.3 Turbulent channel flow at Reτ = 590

the same size with 128 × 129 × 128 spectral functions in the We also performed turbulent channel flow computations for
stream-wise, wall-normal and span-wise direction, respec- Reτ = 590 to examine the effects of the continuity as the
tively. Reynolds number of the flow is increased. For this simu-
We carried out computations employing C 0 - and lation the domain size is 2π , 2, and π in the stream-wise,
C 1 -continuous quadratic discretizations keeping the num- wall-normal and span-wise directions, respectively. The cor-
ber of degrees of freedom nearly the same in both cases. For responding DNS used the same domain size with a resolution
the C 0 case we used a mesh of 163 elements, which gave of 384 × 257 × 384 spectral functions in the stream-wise,
32 × 33 × 32 basis functions in our discrete space, whereas wall-normal and span-wise directions.
for the C 1 case we employed a mesh of 323 elements, which As in the Reτ = 180 case, we present results compar-
led to a discrete space comprised of 32 × 34 × 32 basis ing quadratic discretizations that are C 0 - and C 1 -continuous
functions. (The open knot vector construction is responsible while keeping the number of degrees of freedom nearly the
for the extra basis function in the wall-normal direction; see same. Figures 4, 5 show the results obtained using a C 0 mesh
Hughes et al. [25].) of 32×33×32 basis functions and a C 1 mesh of 32×34×32
Figure 2 illustrates that the mean flow obtained with the basis functions. The difference in the number of degrees
C 1 -continuous discretization is slightly more accurate. How- of freedom is, as mentioned previously, attributable to the
ever, both discretizations show good agreement with the DNS open knot vector construction for the NURBS basis. These
result. On the other hand, the fluctuations are significantly meshes are considered coarse for this simulation, which is
better in the case of C 1 quadratics (see Figs. 2, 3). manifested by the fact that the mean stream-wise velocity is

123
Comput Mech (2008) 41:371–378 377

1.4 of the convection term, a popular choice in finite element flow

0 2
C P 64x65x64 codes. We compared turbulent channel flow results using
1.2 1 2
C P 64x66x64
DNS C 0 - and C 1 -continuous quadratic discretizations. Using a
1 C 1 -continuous quadratic basis yields more accurate mean
flow and fluctuating quantities than C 0 -continuous quadratic
0.8 basis functions. We conclude that smooth NURBS basis
v+

0.6
functions have advantages over C 0 -continuous finite
elements in turbulent flow calculations as anticipated in [6].
0.4
Acknowledgements This research was supported by Office of Naval
0.2 Research Contract N00014-03-0263, Dr. Luise Couchman, contract
monitor, and Sandia National Laboratories under contract number
0 114166. I. Akkerman was supported by the Dutch Technology Foun-
0 100 200 300 400 500 dation STW, applied science division of NWO, and the Technology
y+ Program of the Ministry of Economic Affairs. Y. Bazilevs was partially
supported by the J.T. Oden ICES Postdoctoral Fellowship at the Insti-
(a) Wall-normal velocity fluctuations tute for Computational Engineering and Sciences (ICES). This support
is gratefully acknowledged.
1.8
0 2
1.6 C P 64x65x64
1 2
C P 64x66x64 References
1.4 DNS
1.2 1. Akin JE, Tezduyar TE (2004) Calculation of the advective limit
of the SUPG stabilization parameter for linear and higher-order
1 elements. Comput Methods Appl Mech Eng 193:1909–1922
w+

2. Bazilevs Y (2006) Isogeometric analysis of turbulence and fluid-

0.8 structure interaction. PhD thesis, ICES, UT Austin
0.6 3. Bazilevs Y, Hughes TJR (2007) Weak imposition of Dirichlet
boundary conditions in fluid mechanics. Comput Fluids 36:12–26
0.4 4. Bazilevs Y, Beirao da Veiga L, Cottrell JA, Hughes TJR, Sangalli
0.2 G (2006) Isogeometric analysis: approximation, stability and error
estimates for h-refined meshes. Math Models Methods Appl Sci
0 16:1031–1090
0 100 200 300 400 500 5. Bazilevs Y, Calo VM, Zhang Y, Hughes TJR (2006) Isogeomet-
y+ ric fluid-structure interaction analysis with applications to arterial
blood flow. Comput Mech 38:310–322
(b) Span-wise velocity fluctuations 6. Bazilevs Y, Calo VM, Cottrell JA, Hughes TJR, Reali A, Scovazzi
G (2007) Variational multiscale residual-based turbulence mod-
Fig. 7 Turbulent channel flow at Reτ = 590 computed on a 643 ele-
eling for large eddy simulation of incompressible flows. Comput
ment mesh. Comparison of C 0 - versus C 1 -continuous discretizations
Methods Appl Mech Eng. Available as ICES Report 07-15, UT
Austin (submitted)
7. Bazilevs Y, Michler C, Calo VM, Hughes TJR (2007) Weak Dirich-
over-predicted by both discretizations, the C 1 discretization
let boundary conditions for wall-bounded turbulent flows. Comput
being more accurate (see Fig. 4). As in the Reτ = 180 case, Methods Appl Mech Eng (accepted for publication)
the velocity fluctuations are significantly more accurate for 8. Bischoff M, Bletzinger K-U (2004) Improving stability and accu-
C 1 quadratic NURBS, as shown in Figs. 4, 5. racy of Reissner-Mindlin plate finite elements via algebraic subgrid
scale stabilization. Comput Methods Appl Mech Eng 193:1491–
We performed the same comparison study on h-refined
1516
meshes, resulting in 64 × 65 × 64 functions for the C 0 basis 9. Bochev PB, Gunzburger MD, Shadid JN (2004) On inf-sup stabi-
and 64 × 66 × 64 functions for the C 1 basis. Results are lized finite element methods for transient problems. Comput Meth-
presented in Figs. 6, 7. Both the mean flow and fluctuations ods Appl Mech Eng 193:1471–1489
10. Brooks AN, Hughes TJR (1982) Streamline upwind/Petrov-
are more accurate in the case of C 1 NURBS, for which the
Galerkin formulations for convection dominated flows with par-
mean steam-wise velocity is very close to the DNS result. ticular emphasis on the incompressible Navier–Stokes equations.
From the above results one may infer that solution accuracy Comput Methods Appl Mech Eng 32:199–259
depends more critically on the continuity of the discretization 11. Burman E, Hansbo P (2004) Edge stabilization for Galerkin
approximations of convection-diffusion-reaction problems. Com-
for flows at higher Reynolds number.
put Methods Appl Mech Eng 193:1437–1453
12. Calo VM (2004) Residual-based multiscale turbulence modeling:
5 Conclusions finite volume simulation of bypass transistion. PhD thesis, Depart-
ment of Civil and Environmental Engineering, Stanford University
13. Chung J, Hulbert GM (1993) A time integration algorithm for
In this work, we presented a residual-based, variational multi- structural dynamics with improved numerical dissipation: the gen-
scale model of turbulence that is based on the advective form eralized-α method. J Appl Mech 60:371–375

123
378 Comput Mech (2008) 41:371–378

14. Codina R, Soto O (2004) Approximation of the incompressible Na- 27. Johnson C (1987) Numerical solution of partial differential equa-
vier–Stokes equations using orthogonal subscale stabilization and tions by the finite element method. Cambridge University Press,
pressure segregation on anisotropic finite element meshes. Comput Sweden
Methods Appl Mech Eng 193:1403–1419 28. Koobus B, Farhat C (2004) A variational multiscale method for the
15. Codina R, Principe J, Guasch O, Badia S (2007) Time depen- large eddy simulation of compressible turbulent flows on unstruc-
dent subscales in the stabilized finite element approximation of tured meshes—application to vortex shedding. Comput Methods
incompressible flow problems. Comput Methods Appl Mech Eng Appl Mech Eng 193:1367–1383
(accepted for publication) 29. Kravchenko AG, Moin P, Moser R (1996) Zonal embedded grids
16. Cohen E, Riesenfeld R, Elber G (2001) Geometric modeling with for numerical simulation of wall-bounded turbulent flows. J Com-
splines. An introduction. A K Peters Ltd, Wellesley put Phys 127:412–423
17. Cottrell JA, Reali A, Bazilevs Y, Hughes TJR (2006) Isogeometric 30. Kravchenko AG, Moin P, Shariff K (1999) B-spline method and
analysis of structural vibrations. Comput Methods Appl Mech Eng zonal grids for simulation of complex turbulent flows. J Comput
195:5257–5297 Phys 151:757–789
18. Coutinho ALGA, Diaz CM, Alvez JLD, Landau L, Loula AFD, 31. Kwok WY, Moser RD, Jiménez J (2001) A critical evaluation of
Malta SMC, Castro RGS, Garcia ELM (2004) Stabilized methods the resolution properties of B-spline and compact finite difference
and post-processing techniques for miscible displacements. Com- methods. J Comput Phys 174:510–551
put Methods Appl Mech Eng 193:1421–1436 32. Lele SK (1992) Compact finite difference schemes with spectral-
19. Gravemeier V, Wall WA, Ramm E (2004) A three-level finite like resolution. J Comput Phys 103:16–42
element method for the instationary incompressible Navier– 33. Masud A, Khurram RA (2004) A multiscale/stabilized finite ele-
Stokes equations. Comput Methods Appl Mech Eng 193:1323– ment method for the advection-diffusion equation. Comput Meth-
1366 ods Appl Mech Eng 193:1997–2018
20. Harari I (2004) Stability of semidiscrete formulations for parabolic 34. Moin P (2001) Fundamentals of engineering numerical analysis.
problems at small time steps. Comput Methods Appl Mech Eng Cambridge University Press, London
193:1491–1516 35. Moser R, Kim J, Mansour R (1999) DNS of turbulent channel flow
21. Hauke G, Valiño L (2004) Computing reactive flows with a field up to Re = 590. Phys Fluids 11:943–945
Monte Carlo formulation and multi-scale methods. Comput Meth- 36. Piegl L, Tiller W (1997) The NURBS book (Monographs in visual
ods Appl Mech Eng 193:1455–1470 communication), 2nd edn. Springer, New York
22. Hughes TJR, Mallet M (1986) A new finite element formulation 37. Rogers DF (2001) An introduction to NURBS with historical per-
for fluid dynamics: III. The generalized streamline operator for spective. Academic, San Diego
multidimensional advective-diffusive systems. Comput Methods 38. Scovazzi G (2004) Multiscale methods in science and engineer-
Appl Mech Eng 58:305–328 ing. PhD thesis, Department of Mechanical Engineering, Stanford
23. Hughes TJR, Sangalli G (2006) Variational multiscale analysis: the Universty. Available at http://www.cs.sandia.gov/~ gscovaz
fine-scale Green’s function, projection, optimization, localization, 39. Shakib F, Hughes TJR, Johan Z (1991) A new finite element for-
and stabilized methods. SIAM J Numer Anal (accepted). Available mulation for computational fluid dynamics: X. The compressible
as ICES Report 05-46, UT Austin Euler and Navier–Stokes equations. Comput Methods Appl Mech
24. Hughes TJR, Calo VM, Scovazzi G (2004) Variational and mul- Eng 89:141–219
tiscale methods in turbulence. In: Gutkowski W, Kowalewski TA 40. Shariff K, Moser RD (1998) Two-dimensional mesh embedding
(eds) Proceedings of the XXI international congress of theoretical for B-spline methods. J Comput Phys 145:471–488
and applied mechanics (IUTAM). Kluwer, Dordrecht 41. Tezduyar TE (2003) Computation of moving boundaries and inter-
25. Hughes TJR, Cottrell JA, Bazilevs Y (2005) Isogeometric analysis: faces and stabilization parameters. Int J Numer Methods Fluids
CAD, finite elements, NURBS, exact geometry, and mesh refine- 43:555–575
ment. Comput Methods Appl Mech Eng 194:4135–4195 42. Tezduyar TE, Sathe S (2004) Enhanced-discretization space-time
26. Jansen KE, Whiting CH, Hulbert GM (1999) A generalized-α technique (EDSTT). Comput Methods Appl Mech Eng 193:1385–
method for integrating the filtered Navier–Stokes equations with 1401
a stabilized finite element method. Comput Methods Appl Mech 43. Tezduyar TE, Sathe S (2006) Enhanced-discretization selective
Eng 190:305–319 stabilization procedure (EDSSP). Comput Mech 38:456–468

123