Sādhanā Vol. 39, Part 5, October 2014, pp. 1071–1094.


c Indian Academy of Sciences

An implicit non-staggered Cartesian grid method

for incompressible viscous flows in complex geometries

A K DE

Department of Mechanical Engineering, Indian Institute of Technology Guwahati,

Guwahati 781 039, India
e-mail: [email protected]

MS received 9 January 2012; revised 5 May 2014; accepted 14 May 2014

Abstract. A discrete forcing based Cartesian grid method is presented. The non-
staggered arrangement of velocity and pressure is considered. The pressure gradient
in localized discrete form is added separately with the velocity making them explicitly
coupled. The governing equation is time-integrated implicitly with both linearized and
non-linear forms are investigated. Both linear and bi-linear reconstruction techniques
are tested for extrapolation of velocity near a complex boundary. The present method
is tested for vortical flow in an inclined cavity, flow past circular and inclined square
cylinder. Both homogeneous and non-homogeneous Dirichlet forcing problems are
tested. The parallelized version of the method is applied to 2D-to-3D transitional flow
behind a single and multiple circular cylinders. The present numerical results compare
well with the previously documented results.

Keywords. Immersed boundary; non-staggered; implicit; viscous flow.

1. Introduction

One of the major issues of numerical computation of fluid flow is the ability of handling complex
geometries that appear in numerous situations ranging from practical flows to problems of fun-
damental importance. Though, a number of techniques, such as the coordinate transformation,
non-conforming chimera grid technique have been demonstrated in the structured grid domain,
each of them have their own issues that limit their direct application. It is obvious that if any
arbitrary domain is handled with Cartesian grids the computational efficiency, in terms of accu-
racy achievable and the cost of computation, is enhanced significantly. This motivation has led
to the idea of Cartesian grid method which is more popularly known as the immersed bound-
ary technique. In this method, a complex surface is kept immersed in non-conforming Cartesian
grids and the fluid flow equations are solved satisfying all the physical constraints.
The immersed boundary method can be broadly classified into two groups; the continuous
forcing approach and the direct forcing technique. Peskin (1972) first introduced the idea of

1071
1072 A K De

using Cartesian grids for flow simulation in heart valves where he modelled the force field on a
surface by a distributed function. Since then a number variants proposing different distribution
functions for elastic boundaries (Saiki & Biringen 1996; Lai & Peskin 2000; Zhu & Peskin 2003)
and rigid surfaces (Goldstein et al 1993; Angot et al 1999; Iaccarino & Verzicco 2003) have
appeared. Arthurs et al (1998), Bottino (1998) and McQueen & Peskin (1997) applied this tech-
nique in biological flows. Successful application of this technique can be found in a variety of
flows as wide as locomotion of aquatic animals (Fauci 1990), flow of suspensions (Fogelson &
Peskin 1988) and multi-phase flows (Unverdi & Tryggvason 1992). However, this method inher-
ently demands smoothing of the distributed function for rigid surfaces which is not suitable for
flows with small scale changes, especially at high Reynolds numbers.
In the second category, a number of discrete points are selected on the immersed surface
where the boundary conditions are forced to satisfy directly. Mohd-Yosuf (1997) and later Kim
et al (2001) introduced a forcing term in the Navier–Stokes equations which enforces the no-
slip boundary condition on a solid surface while keeping the fluid domain unchanged. As the
physical constraints are directly (and thus more accurately) forced on the boundaries, this tech-
nique is more suitable for flows that involve complex evolution. However, implementation of
the boundary forcing has a strong relationship (Mittal & Iaccarino 2005) with the solution tech-
nique, though a consistent boundary treatment does not restrict the time-step of the marching
algorithm.
In the discrete direct forcing approach, usually a number of closely connected fluid points are
extrapolated for the evaluation of the forcing term. Thus, stability and convergence behaviour
of the method principally depends on the extrapolation technique incorporated. Both linear as
well as bi-linear interpolation procedures were successfully employed (Kim et al 2001; Fadlun
et al 2000; Tseng & Ferziger 2003; Verzicco et al 2000; Ye et al 1999) for the boundary forc-
ing, though Majumdar et al (2001) and later Tseng & Ferziger (2003) found no appreciable
improvement by using higher-order interpolation technique in both laminar as well as turbulent
flows. Among several modified methods, Tau (1994), Tucker & Pan (2000) and Kirkpatrick et al
(2003) proposed the cut-cell method where the arbitrary intersected boundary cells are treated
with their true shapes. For more complicated shapes this technique requires generation of the
surface, determination of the arbitrary boundary cells and thus primarily it operates on a hybrid
mesh. Attempts have also been made with the complex boundaries immersed in curvilinear struc-
tured grids (Ge & Sotiropoulos 2007; Borazjani et al 2008), unstructured mesh (Xia et al 2009;
Denaro & Sarghini 2002) and overlapping chimera-type grids (Tai et al 2005). However, in such
forms the advantages of Cartesian grid computations are partly lost because of algorithmic com-
plexity, advanced data structure and fall in accuracy otherwise achievable. On the same note this
method has received wide acceptance because of its prospect of handling complex geometry
without bringing in new difficulties. For instance, applications including fluid structure inter-
action (Huang & Sung 2009; Gilmanov & Acharya 2008), large-eddy simulation of turbulent
flows (Li & Wang 2004; Tyagi & Acharya 2005; Balaras 2004) and two-phase flows (Yang &
Stern 2009; Berthelsen & Ytrehus 2007) have emphasized this aspect.
A major concern in flow calculations using non-staggered arrangement of variables is the weak
coupling between velocity and pressure which often leads to numerical wiggles. To alleviate this
issue, use of staggered grid is generally preferred. However, over the years, use of non-staggered
arrangement has found its application where an adequate amount of numerical dissipation (Rhie
& Chow 1983; der Wijngaart 1990; Armfield 1991; Barton & Kirby 2000) is added to damp
out the oscillations. The present paper treats this issue in a different way: the discrete pressure
gradient which is realized to drive the flow locally, is added with the velocity computed without
the effects of pressure and thereby explicitly achieving a strong coupling between them. The
 Non-staggered Cartesian grid method for incompressible viscous flows 1073

solution strategy lends its idea from the projection method, though it focusses on stable implicit
treatment of the variables — both in linearized and non-linear forms. The solution of the pressure
Poisson equation, the heaviest task in incompressible flows, is briefly reviewed in the light of
pre-conditioned conjugate gradient method when applied to the immersed boundary method.
The entire solution procedure is parallelized to facilitate its use in large three-dimensional flows
involving complex geometries. As the present work does not address the accuracy and resolution
properties of this method, the test cases are limited to moderate Reynolds number which includes
laminar and 2D-to-3D transitional flows.
The paper is organized as follows. Section 2 briefly describes the governing equations
followed by detailed discussion on the present projection method, implicit treatment of the vari-
ables, the immersed boundary technique, linear solvers and the vectorization method in section 3.
In section 4, results from a number of internal and external flow test cases involving complex
boundaries are reported that measure various aspects of the present technique.

2. Governing equations

Incompressible viscous flows are governed by the continuity and momentum equations, written
below in normalized form by using suitable length (L), velocity (U ) and time (L/U ) scales
∂ui
= 0 (1)
∂xi
∂ui ∂(ui uj ) ∂p 1 ∂ 2 ui
+ = − + , (2)
∂t ∂xj ∂xi Re ∂xj ∂xj
where Reynolds number is defined as Re = U L/ν.

3. Numerical details

Using the finite volume approximations, Eqs. (1–2) can be written in terms of approximate inte-
grals for a cell denoted by its center P , faces f and volume VP

Ffn+1 = 0 (3)
f

n+1
ui,P − uni,P  
VP + c1 Ffn+1 ui,f
n+1
+ c2 Ffn uni,f + c3 Ffn−1 ui,f
n−1
=
t
f

1   n+1  1   n+1 
− pf + pfn Sf,i + Fdf i + Fdf
n
i , (4)
2 2Re
f f

where c1 = c2 = 1/2, c3 = 0 and c1 = 0, c2 = 3/2, c3 = −1/2 correspond to the 2nd -order

implicit Crank–Nicolson (CN) and semi-implicit Adams–Bashforth–Crank–Nicolson (ABCN)
time integration schemes, respectively. Note that diffusion flux of a scalar φ is given as
    
∂ 2φ
dV = ∇ · ∇φdV = ∇φ · dS = (∇φ · S)f = Fdφi . (5)
V ∂xj ∂xj V S f f
1074 A K De

3.1 Velocity pressure coupling

As the pressure field is not known at the (n + 1)th time-level, first a provisional velocity field
(u∗i ) is predicted excluding the pressure term with mass flux taken as the latest available one

u∗i,P − uni,P  
VP + c1 Ffn+1,l u∗i,f + c2 Ffn uni,f + c3 Ffn−1 ui,f
n−1
=
t
f
1  ∗ 
Fdf i + Fdf
n
i (6)
2Re
f

The face velocity is then calculated by adding a pressure gradient with the provisional velocity
linearly interpolated at the face

uf = L(u∗P , u∗nb ) − t (∇p)f = u∗f − t (∇p)f . (7)

Note that L is the linear interpolation operator for the cell center (uP ) and the corresponding
neighboring velocity (unb ). As the face velocity carries the local pressure gradient, mass flux
can be calculated by

Ff = uf · Sf = u∗f · Sf − t (∇p)f · Sf = Ff∗ − t (∇p)f · Sf , (8)

where
Ff∗ = u∗f · Sf . (9)
Thus a strong coupling between the ensuing velocity field and the subsequent pressure field is
explicitly achieved which prevents spurious oscillations (Verma & Eswaran 1999) often associ-
ated with non-staggered arrangement of variables. A similar time-splitting form in staggered grid
was proposed by Kim & Moin (1985) while Kim & Choi (2000) treated the pressure with 2nd -
order accuracy in the unstructured-grid framework. The pressure equation can now be obtained
by inserting Eq. (8) into the discrete conservation law, Eq. (3)
 n+1,l+1 
Ff = [Ff∗ − t (∇p)f · Sf ] = 0
f f
 1  ∗
=⇒ (∇p)f · Sf = Ff . (10)
t
f f

The pressure field obtained from Eq. (10) projects the velocity field such that the discrete
mass conservation law (Eq. 3) is satisfied. As the mass flux can be viewed as the carrier of
momentum, it (along with the pressure field) can be used to solve the momentum equations
(Eq. 4) which now has become a generic convection–diffusion transport equation owing to its
linearity.

3.2 Implicit treatment of the nonlinearity

In the present method, the nonlinearity of the governing equations is realized through the predic-
tion of the provisional velocity field seen in Eq. (6). Two different approaches are attempted to
resolve it amicably. Though these two alternative routes are formulated keeping in mind a range
 Non-staggered Cartesian grid method for incompressible viscous flows 1075

of stiff problems which may arise, they can be used with equal ease. The provisional velocity
field, in the absence of a pressure field, is transported by the mass flux leading to the functionality

F n+1 = f (u∗i ).

3.2a Linearized form (LM): If the convective flux is linearized to separate out the available
mass flux and the provisional velocity field to be solved it takes the form

(Ff ui,f )∗ = F n+1,l u∗i,f ,

where l denotes the latest available mass flux and should be corrected subsequently based on
the converged pressure field using Eq. (8). Thus, Eqs. (6), (8) and (10) have to be iterated until
flux-convergence is achieved, i.e.,

Ffn+1,l −→ Ffn+1 .

Essentially this procedure uses flux iteration (denoted here by l) to resolve the nonlinear con-
vective fluxes. It should be noted here that such linearization is only first-order accurate in
progressive estimates, i.e., O(φ n+1,l+1 − φ n+1,l ). Thus, owing to highly localized acceleration
(or deceleration) fluxes may not converge fast rendering the method to be slow. This can be
avoided by using a smaller t which makes the whole time integration technique stable. How-
ever, in such a situation purpose of using the implicit method is partially defeated. This issue is
likely to occur in problems with strong shear, violent 3D transition or sudden changes in the flow.
Though this method consistently treats the governing set of equations implicitly, the pressure
equation which is likely to consume majority of the computational resources has to be solved a
number of times. The second alternative approach stems from this feature of the linearized form.

3.2b Non-linear form (NM): If the equation depicting the provisional velocity field (Eq. 6) is
not linearized, it takes the form
  
t  1 ∗ t 
u∗i,P + c1 Ff∗ u∗i,f − Fdf i = uni,P − c2 Ffn uni,f
VP 2Re VP
f f

1 n
+ c3 Ffn−1 ui,f
n−1
− Fdf i , (11)
2Re
which can be cast into a set of nonlinear discrete equations

Ni (u∗j ) = 0, (12)

where i and j refer to the equation and variable index. This system can be solved using Newton’s
method which is computationally cheap, given by the formula
 −1
∂Np k
u∗j k+1 = u∗j k −α Np (u∗ )k . (13)
∂u∗j

Though this technique is gifted with fast quadratic convergence, inversion of the Jacobian matrix
∂Ni /∂u∗j is not straight forward. Note that the structure of the Jacobian matrix is banded as
Eq. (12) contains only immediate (4 in 2D and 6 in 3D) neighbours. However, if this matrix is
1076 A K De

assumed to be diagonal which is an approximation to the original sparse banded one, the formula
simplifies to
Nj (u∗ )k
u∗j k+1 = u∗j k − α , (14)
aP
where aP is the diagonal entry for the Jacobian matrix, given by
⎡ ⎤

t ⎣ 1
aP = 1 + Ff βf − γf ⎦ , (15)
VP 2Re
f

where βf and γf are geometric parameters that appear in the approximation of surface integrals
of convective and diffusive fluxes, respectively. This approximation results in cheap and fast
convergence of Eq. (12). However, owing to the diagonal approximation this method is expected
to fail for finer meshes as the interaction between neighbouring nodes become enhanced. Thus,
to compensate for the “poor" initial guess, possible instability and considerable non-uniformity
near solid boundaries the updates may need to be under-relaxed.
The above procedure comprising Eqs. (6) (or 12), (9), (10) and (8) lead to a pressure field
(and thus the mass flux) that satisfies mass conservation law, but not necessarily a divergence-
free velocity field. This can be ensured by solving Eq. (4) which now has become linear with
respect to uin+1 . Note that the issue of nonlinearity arises only for the implicit (CN), and not for
the explicit (ABCN) time marching procedure.

3.3 Immersed boundary formulation

The present immersed boundary method is based on direct discrete forcing approach where
Cartesian grids are laid on a solid object and the boundary conditions are forced on a number
of discrete points ( s ) on the surface of the object. The computational cells are categorized
as the fluid ( f ) and solid cells based on the cell centers lying outside and inside the object,
respectively, as shown in figure 1 which is a simple illustration. Among the solid cells, those
having at least one fluid cell to as its neighbour are termed as the inner cell ( I ) while the
neighbouring fluid cells are hereafter referre to as the band cells ( B ). In direct forcing approach
the inner cells are treated as field function (g) of band cells that enforces required boundary
conditions (φb ) on the forcing points.

φ( I ) = g{φ( B )} such that φs = φb . (16)

Thus, during calculations, the inner cells that appear as neighbouring stencil for band cells, are
required to be prescribed so that these band cells “sense" the proximity of the wall.
A blown-up view of a portion of the immersed boundary is shown in figure 2 where P , N and
b are typical inner cell, its fluid neighbour (band cell) and the forcing point, respectively, while
Q is the mirror point obtained by constructing the surface normal n through P . The relevant
band points (N1 , N2 · · · ) can be extrapolated to calculate the unknown either directly on the
forcing points (b1 , b2 · · · ) or on the mirror points (Q1 , Q2 · · · ). Tseng & Ferziger (2003) used
the second alternative as they found unstable solutions for the cases when the band cells are
too close to the surface. However, as this approach uses φ(P1 ) = 2φb − φ(Q1 ), the boundary
forcing is achieved in a stencil (P1 Q1 ) twice as large as the original distance (P1 b1 ). In the
present work, two extrapolation procedures are employed with the boundary conditions forced
on the surface points. In doing so it has been observed that with progressive refinement if the
 Non-staggered Cartesian grid method for incompressible viscous flows 1077

forcing points ( Ωs )

x
x x

x
fluid cell x
( Ωf ) x x
x

x fluid neighbors
x
x ( ΩB)
x
x immersed boundary
x
x
solid cell
x
inner cell ( ΩI )
x
x x
x

Figure 1. Definition of different category of cells in connection to the immersed boundary.

fluid cell forcing point ∇ mirror point

+ band cell inner cell solid cell

N
+ N3 + 4
Q2 ∇
b2
N2 +
P2 P3

N1
+ P1
Q1∇ n b1

N 5+

Figure 2. Geometric construction of the immersed boundary.

1078 A K De

inner points that are very close to the surface are modified as the fluid points, the convergence
behaviour improves significantly. The criterion that has been consistently used here is an inner
point modifies to a fluid point if its distance from the immersed boundary is less than 5% of the
cell representative length, (xy)1/2 or (xyz)1/3 .

Linear/Bi-linear reconstruction: A linear or bi-linear (relevant extra Bi-linear terms under-

lined) polynomial approximation

φ = a + bx + cy + dxy in 2D and φ = a + bx + cy + ez in 3D (17)

can be used to extrapolate φ at the desired location ( I ). If an inner point has φ1 , φ2 , φ3 and φb
as its neighbouring band points and the forcing point, respectively, then prescription of Dirichlet
condition, φ = φb leads to a linear system Ax = B where
⎡ ⎤ ⎡ ⎤
1 x1 y1 x1 y1 1 x 1 y 1 z1
⎢ 1 x2 y2 x2 y2 ⎥ ⎢ 1 x 2 y 2 z2 ⎥
A=⎢ ⎥ ⎢
⎣ 1 x3 y3 x3 y3 ⎦ and ⎣ 1 x3 y3 z3 ⎦ ,
⎥ (18)
1 xb yb xb yb 1 x b y b zb

x = [a b c d]T and x = [a b c e]T , B = [φ1 φ2 φ3 φb ]T and B = [φ1 φ2 φ3 φb ]T for 2D and

3D, respectively. The homogeneous Neumann condition on the boundary given by ∂φ/∂n = 0
can be cast into the form

∇φ · n = 0, (19)

which yields (b+dyb )nx +(c+dxb )ny = 0 and bnx +cny +enz = 0 in 2D and 3D, respectively.
Thus, A and B of the above linear system change to
⎡ ⎤ ⎡ ⎤
1 x1 y1 x1 y1 1 x1 y1 z1
⎢1 x2 y2 x2 y2 ⎥ ⎢ 1 x2 y2 z2 ⎥
⎢ ⎥,⎢ ⎥ (20)
⎣1 x3 y3 x3 y3 ⎦ ⎣ 1 x3 y3 z3 ⎦
0 nx ny xb ny + yb nx 0 nx ny nz

and [φ1 φ2 φ3 0]T , [φ1 φ2 φ3 0]T , respectively.

In the present work, no attempt has been made to use theoretically higher order [> O(x 2 )]
accurate interpolation scheme. The primary reason being it was demonstrated by Majumdar et al
(2001) that use of quadratic (or higher order) polynomials does not improve the quality of
solution. Moreover, it can be observed that as the choice of neighbouring band cells directly
influences forcing of boundary conditions, convergence behaviour is likely to change if the
coefficient matrix A becomes near-singular. With grid refinement, the rows of A become sim-
ilar to each other leading to small determinant, and more so when the number of rows are
more than 4. Thus, use of higher order polynomial is not favourable to grid refinement which
is detrimental to the quality of solution. As the primary drawback of Cartesian grid methods
is its inability to achieve systematic refinement near rigid boundaries, it is better to use over-
all refined mesh than to use higher-order interpolation polynomial for the forcing of boundary
conditions.
 Non-staggered Cartesian grid method for incompressible viscous flows 1079

3.4 Linear solver

All set of linear equations arising during the time marching procedure can be cast into a form

Af φ = bf , (21)

where the linear operator Af corresponds to a specific stencil size decided by the convective
scheme (and less likely by the diffusion term) used. However, the grids being Cartesian, the
pressure equation (Eq. 10) results in a simple penta-diagonal (2D) or septa-diagonal (3D) matrix
structure. The reconstruction of a variable near the immersed boundary yields a linear relation
among an inner cell and a number of band cells.

Ab φ = bb . (22)

Note that the above system is same as the one described in section 3.3. Equation sets (21) and
(22) need to be solved simultaneously so that progressive changes in the inner cell values force
the boundary conditions on the desired locations on the immersed boundary.
In the present work, the stabilized version of the Bi-Conjugate Gradient method (Zhang 1997)
has been used for the velocity steps (Eqs. 4 and 6) while Stone’s Strongly Implicit Procedure
(SIP) (Stone 1968) has been used as a pre-conditioner (Ferziger & Perić 1996) for monotonic
and rapid convergence of the pressure Poisson equation (Eq. 10). A brief description of the SIP
factorization along with the diagonals is provided in Appendix A.

3.5 Parallelization strategy

The three-dimensional numerical code has been parallelized based on distributed memory allo-
cation. The computational domain is decomposed into a number of sub-domains and each of
them is assigned to a computing processor. At the interfaces of these sub-domains, communi-
cation of solution variables are performed in the MPI environment to synchronize computations
among the processors. Though computational load on each processor should be theoretically
same so as to achieve optimum performance, it is often compromised because of grid distribu-
tion and nature of the solution. If a set of inner cells and the corresponding band cells lie in a
single processor, extra communications do not arise. However, when they are distributed among

0.8
U
0.6

L
0.4

0.2
α
0
L
0 0.5 1 1.5
x
Figure 3. Geometric detail and the Cartesian mesh (every fourth grid line is shown) used for the inclined
cavity problem.
1080 A K De

x x
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.
1 0.2 1
α = 45 α = 30
o o

0.8 0.8
+ + + 0.1 ++ + +
+
+
0.6 + 0.6 +
+ +
++ +0 + +
y

v
y
+
0.4 0.4
+ +
-0.1
0.2 + + 0.2 +
++++

0 -0.2 0
-0.2 0.2 0.6 1 -0.2 0.2 0.6
u u

Figure 4. Mid-sectional velocities for the vortical flow in the inclined cavity. —◦—:u,linear; ——:u,bi-
linear; —∇—:v,linear; ——:v,bi-linear correspond to present data while •:u and +:v refer to Demirdžić
et al (1992).

different processors either the original sub-domains has to be restructured at the cost of compu-
tational imbalance or complex communication paths have to be determined. The present method
has resorted to the first alternative so that extra communication is avoided. Preliminary tests have
shown that the overall parallel code scales well with processors given the problem size is large
enough. Though BiCGSTAB algorithm can be parallelized it is the ILU-type of factorization
and its subsequent forward and backward sweeps, that accelerates the convergence, can only be
made partially parallel owing to its sequential data dependence.

4. Results

The present immersed boundary method has been applied to flows in reasonably complex geome-
tries covering both homogeneous and non-homogeneous Dirichlet boundary conditions for the
velocities. However, as the method does not intend to venture into the high Reynolds num-
ber regime, the problems are so chosen that its capability to handle laminar unsteady flows or
2D-to-3D transitional flows are appropriately judged.

4.1 Vortical flow in an inclined cavity

Steady vortical flow in an inclined cavity whose geometric details are shown in figure 3, is a
suitable test of boundary non-orthogonality. Simulations have been carried out for α = 30◦ and

α = 450 α = 300
Figure 5. Computed stream lines inside the inclined cavity.
Non-staggered Cartesian grid method for incompressible viscous flows 1081

0.07
0.055

0.04

+
0.025
+
error (ε)

+ 2
+
0.01 1

u, linear
v, linear
+ u, bilinear
v, bilinear
0.05 0.07 0.09 0.11
1/2
(ΔxΔy)
Figure 6. Test of spatial accuracy for flow past a circular cylinder.

0
10

-1
10
residual

-2
10

-3
10

SIP, LM
-4 BiCGSTAB, LM
10 PBiCGSTAB, LM
SIP, NM
BiCGSTAB, NM
PBiCGSTAB, NM
10-5 1 2 3
10 10 10
iterations
Figure 7. Behaviour of the linear solvers applied with the two linearization forms.
1082 A K De

0.5

y
0

-0.5

-1

(a) -1 0 1 2 3 4 5
x
4

2
y

-2

-4
(b) 0 5 10 15
x

Figure 8. Flow pattern behind the circular cylinder; (a) stream lines at Re = 40 and (b) instantaneous
vorticity contours (negative ωz in dashed line) at Re = 100.

45◦ at Re(U L/ν) = 100 with a uniform mesh of 160 × 80. The steady velocity profile along
mid-width and mid-height, shown in figure 4, compare well with Demirdžić et al (1992) where
computations were performed on a body fitted refined mesh. It can be seen that the improvement
in solution accuracy by using bi-linear extrapolation as the reconstruction scheme is marginal.
This observation is consistent with earlier studies on discrete forcing approach. In view of pos-
sible singularities (in the coefficient matrix of Eqs. (18) and (20) arising from specific grid
arrangement and marginal improvement in the solution accuracy it is advisable to use linear
extrapolation as the reconstruction scheme. The computed streamlines are shown in figure 5
where the secondary vortex at α = 30◦ is correctly resolved.

Table 1. Comparison of force coefficients obtained from the two non-linear formulations with previous
studies for flow past a circular cylinder.

Re = 40 Re = 100
Cd St < Cd > CL
Tseng & Ferziger (2003) 1.53 0.164 1.42 0.29
Kim et al (2001) 1.51 0.165 1.33 –
Present, LM,linear 1.57 0.171 1.455 0.311
Present, NM,linear 1.59 0.169 1.47 0.298
Present, LM,bi-linear 1.53 0.168 1.462 0.3
Present, NM,bi-linear 1.56 0.17 1.448 0.317
 Non-staggered Cartesian grid method for incompressible viscous flows 1083

q = π/2

q=2

Figure 9. Instantaneous vorticity field at q = π/2 and steady stream lines at q = 2 for flow past a rotating
cylinder.

4.2 Flow past a stationary/rotating circular cylinder

Flow past a circular cylinder is an excellent choice for simulation in complex geometries as it
offers a rich combination of geometric and physical complexities. In the present study, both the
stationary and rotating cases are examined with the proposed technique. The physical problem
involves unbounded flow (U∞ ) past a circular cylinder of diameter d placed at the origin. The

-4
0.6

0.45
CD -4.5

0.3
CL
cD

-5
0.15

0 CL
-5.5

-0.15

150 300 450 600 750

t
Figure 10. Average drag and lift coefficients showing attainment of a asymptotic steady state at q = 2.
1084 A K De

Table 2. Force coefficients for flow past a rotating circular cylinder.

q = π/2 q=2
< Cd > < CL > St Cd CL
Maruoka (2003) 0.84 −4.11 0.191 – –
Mittal & Kumar (2003) – – – 0.31 −5.29
Zhang et al (2008) 0.77 −3.98 0.191 0.39 −5.26
Present 0.90 −4.109 0.193 0.352 −5.258

upstream, downstream and the cross-stream extent of the computational domain are chosen as
5d, 25d and 16d, respectively, and Reynolds number is defined as Re = U∞ d/ν. The non-
uniform Cartesian grid used for the calculations is 250 × 140 with a refined square block of size
1.1d consisting of uniform grid (60 × 60) is wrapped around the cylinder to resolve the flow
near the cylinder and in the near wake. For the rotating (counterclockwise with angular speed ω)
cylinder case, the tip velocity is defined as q = dω/2U∞ . In all the subsequent test problems the
Orlanski boundary condition, given by ∂φ/∂t + Uav ∂φ/∂n = 0 has been used where φ, Uav , n
refer to velocity components, average stream-wise velocity at the exit and stream-wise direc-
tion, respectively. This condition being similar to the first-order linear wave equation transports
ensuing flow structure without any distortion.

0.5

d
y

-0.5 α
Lr
-1
-1 0 1 2 3 4 5 6
(a)
x

2
y

-2

0 5 10 15 20 25
(b) x

2
y

-2

0 5 10 15 20 25
(c)
x

Figure 11. Wake behind the inclined cylinder; (a) steady separation bubble at Re = 40, α = 22.7◦ ,
instantaneous vorticity field (negative ωz in dashed line) at Re = 100 for (b) α = 6◦ and (c) α = 45◦ .
 Non-staggered Cartesian grid method for incompressible viscous flows 1085

To test the actual order of accuracy of the numerical solution obtained from the two recon-
struction schemes, a domain (−2d ≤ x, y ≤ 2d) including the cylinder is chosen. Computations
have been performed on progressively coarser mesh with results of the finest mesh (250 × 140)
taken as the benchmark solution. The solution at a coarser mesh is interpolated to the finer mesh
for comparison and is shown in figure 6. The error () in the figure is defined by the l2 norm
of the difference between the benchmark and the interpolated solutions. Both the extrapolation
procedures yield a convergence behaviour  ∝ (xy)n/2 with n ≈ 1.9. The loss in accuracy
can be attributed to the combined effects of overall solution error and the extrapolation error at
the boundary. No appreciable difference among the two implicit formulations (LM and NM) is
observed. Figure 7 shows the relative convergence behaviour of the different linear solvers when
applied with the two implicit formulations. Rate of convergence improves significantly from
the SIP method to the PBiCGSTAB technique while BiCGSTAB lies between them with non-
monotonic convergence. However, it is observed that with the non-linear form, PBiCGSTAB
gives a smooth convergence excluding the possibility of occasional instabilities of the linear
form (LM). The steady state streamlines and instantaneous vorticity contours at Re = 40 and
100 are shown in figure 8. The average force coefficients and strouhal number is compared in

1.9
α = 29.7
o Present
Sohankar et al. (1998)
3
+ Yoon et al. (2010)

1.7 +
<Cd>

2
Lr/a

1.5 +
1 + +
Present + +
Yoon et al. (2010)
1.3
0 15 30 45 0 10 20 30 40
(a) Re (b) α
0.6 0.2

+ +
+
0.4 + +
0.16 +
+ +
C′L

St

+
+
+
+
0.2
+ +
0.12

0
0 10 20 30 40 0 10 20 30 40
(c) (d)
α α
Figure 12. Comparison of (a) re-attachment length at α = 29.7◦ , (b) average drag coefficient, (c) root
mean square of the lift coefficient and (d) Strouhal number at Re = 100.
1086 A K De

Table 3. Computational details for three-dimensional flow past a circular cylinder.

Re mesh band cells t

200 225 × 110 × 80 6320 0.01
300 241 × 120 × 100 9108 0.005

table 1 which reemphasizes the marginal improvement of solution accuracy due to the bi-linear
extrapolation scheme and near- identical prediction by the two implicit formulations.
The rotating cylinder case tests the ability of the method for non-homogeneous boundary
conditions as the boundary velocity is the tangential velocity of the cylinder that changes in the
azimuthal direction. All the computational features are kept same as of the stationary case and
simulations have been carried out for q = π/2 and 2 at Re = 200. In agreement to Zhang et al
(2008), the flow is found to be time-periodic at q = π/2 while long time simulation shows that
the wake oscillation decays slowly to reach the steady state at q = 2. Instantaneous vorticity
contours and steady state streamlines are shown in figure 9 while attainment of the asymptotic

Figure 13. Span-wise vorticity at (a) Re = 200 and (b) Re = 300 behind the circular cylinder; Mode A
and B instabilities are shown with the span-wise wavelength.
 Non-staggered Cartesian grid method for incompressible viscous flows 1087

steady state can be seen form the drag and lift coefficient signal, shown in figure 10. The average
force coefficients, shown in table 2, are close to the reported values.

4.3 Flow past a inclined cylinder

As an example of flow over a body with sharp corners, two-dimensional unbounded flow past a
square cylinder of side a placed at an angle α (see figure 11 (a)) with the incoming stream has
been tested. The cylinder is located at the origin making a blockage of width d = a(cos α +
sin α). Only the nonlinear formulation (NM) is used while all the geometric parameters are kept
same as the circular cylinder case. A non-uniform mesh of size 300 × 150 is used which has a
refined (60 × 60) block (|x, y| < 0.6d) containing the cylinder with t = 0.01U/d being used
for the time marching.
The steady separation bubble at α = 22.7◦ , Re = 40 is shown in figure 11(a) while vorticity
field of two selective cases from the unsteady periodic regime are shown in figures 11(b) and
(c) which correspond to α = 6◦ , 45◦ and Re = 100, respectively. The recirculation length (Lr ),

Figure 14. Stream-wise vorticity behind the circular cylinder; (a)Re = 200 and (b)Re = 300.
1088 A K De

Table 4. Summary of 2D-3D transitional flow behind a circular cylinder.

Re = 200 Re = 300
< Cd > St λz /d < Cd > St λz /d
Labbé & Wilson (2003) 1.318 0.195 ≈4 1.287 0.205 ≈1
Williamson (1988) – 0.196 – 0.204 –
Williamson (1996) – – ≈ 3.55 – – ≈ 1.05
Balaras (2004) – – – 1.27 0.21 –
Present 1.42 0.198 ≈ 3.1 1.283 0.211 ≈1

defined as the distance of the saddle point from the cylinder center (shown in figure 11(a)), as a
function of Re is shown for α = 22.7◦ in figure 12(a). Note that Lr is only shown for the steady
cases. The average drag coefficient (< Cd >), root mean square of the lift coefficient (CL  ) and
the vortex shedding frequency (St) are compared in figures 12(b)–(d) with Sohankar et al (1998)
and Yoon et al (2010) for the range 0 < α < 45◦ at Re = 100. The results agree well with
two previous works and the maximum difference is found at α ≈ 30◦ when the wake shows
maximum fluctuation evident from CL  . However, the difference is acceptable considering the
difference in resolution and interval of time integration among the present and the two previous
studies.

Figure 15. Geometric details for the flow past two staggered cylinders.
 Non-staggered Cartesian grid method for incompressible viscous flows 1089

4.4 Three-dimensional wake transition behind a circular cylinder

The present technique in its parallelized form has been employed to predict three-dimensional
wake transition behind a circular cylinder placed in uniform flow. The span-wise length of
the cylinder is taken as 2πd while other details are shown in table 3. The non-uniform mesh
described in section 3.3 is uniformly expanded in the span-wise direction. A coarse mesh simu-
lation is first conducted and the dynamically stationary result is interpolated to the chosen grid
(221 × 151 × 100) which serves as the initial condition for the results shown here. The paral-
lelized code is run on 8 processors for nearly fifty vortex shedding cycles. It has been observed
that the pre-conditioned-BiCGSTAB solver takes only a few iterations (< 10) to reduce the
errors of four orders if the interpolated result from a coarser grid is used as the initial condition.

Figure 16. Stream-wise vorticity showing onset of three-dimensionality; (a) Re = 200 and (b) Re =
300.
1090 A K De

In the dynamically stationary state, the numerical code takes about 40 minutes to advance one
vortex shedding cycle.
It has been observed that the two-dimensional wake shows signs of three-dimensionality at
Re ≈ 190. Around this Reynolds number a three-dimensional instability, known as the mode A
instability, occurs whose span-wise wavelength (λz /d) varies between 3d and 4d. At even higher
Reynolds number the transition of the vortex shedding mode becomes more complicated and at
Re ≈ 300 the wake exhibits a span-wise wavelength λz /d ≈ 1. This instability is known as the
mode B instability. While in mode A the span-wise vorticity appears in staggered arrangement in
the stream-wise direction, they rearrange to in-line configuration in mode B. Figures 13 and 14
show span-wise (ωx ) and stream-wise (ωz ) vorticity surfaces at Re = 200(a) and 300 (b). The
onset of the instabilities at Re = 200 and breaking of the von Karman vortex street at Re = 300
is evident form figures 14 (a) and (b), respectively. The staggered arrangement of the span-wise

Figure 17. Span-wise vorticity showing the mode of instability; (a) Re = 200 and (b) Re = 300.
 Non-staggered Cartesian grid method for incompressible viscous flows 1091

vorticity (of mode A) at Re = 200 is seen to change to in-line pattern (of mode B) at Re = 200
in figures 13 (a) and (b). A detailed view of the span-wise vorticity in the near wake (x = 2
plane) is shown in the inset of figure 13 which confirms the span-wise wavelength (λz /d) of the
flow instabilities. The average drag coefficient (< Cd >), the strouhal frequency (St) and λz is
compared with a few previous studies in table 4.

4.5 Three-dimensional wake behind two staggered circular cylinders

The final test case is the flow past two cylinders in staggered arrangement with a pitch ratio
p/d = 1 and angle α = 10◦ . The geometric details and the decomposition are shown in
figure 15. The Reynolds number is defined as Re = U∞ d/ν where U∞ and d are the incoming
flow velocity and diameter of the cylinders, respectively. Simulations have been performed for
Re = 200 and 300 with a non-uniform mesh 257 × 202 × 120 distributed among 32 computing
processors. Three-dimensional wake transition is clearly observed in figure 16 where periodic
nature of the stream-wise vorticity at Re = 200 breaks into irregular pattern at Re = 300. At
this range of Reynolds number, the wake shows sign of mode B instability (see figure 17) as in
both the cases span-wise vorticity exhibits a wavelength λ ≈ 1.3d. Though a regular staggered
arrangement of the span-wise vorticity is seen at Re = 200, complex chain like structure is seen
at Re = 300.

5. Conclusions

The objective of the present work was to develop a robust implicit Cartesian grid method based
on the discrete forcing approach. The arrangement of variables was chosen in non-staggered
form which, in contrast to the staggered form, is known to produce pressure wiggles owing to
the lack of velocity–pressure coupling. This issue has been addressed adequately as to how the
coupling can be retained. The solution is achieved in predictor–corrector steps where a provi-
sional velocity is predicted without considering the pressure and a localized pressure gradient
is applied with the velocity in the corrector step. The non linearity arising due to the implicit
treatment of the convective terms has been handled both by linearization as well as a complete
non linear system. Both linear and bi-linear extrapolation techniques have been tested for the
reconstruction of velocity and pressure on the immersed boundary. The Poisson equation for
the pressure has been solved using the pre-conditioned BiCGSTAB method whose convergence
behaviour is studied for different reconstruction schemes.
In agreement with the previous studies, marginal improvement in the quality of solution is
found by using bi-linear extrapolation while often convergence behaviour deteriorates owing to
the ensuing matrix structure. Thus, for three-dimensional computations, linear reconstruction
scheme is found to be the optimum choice. An accelerated monotonic convergence behaviour
is found when the BiCGSTAB technique is pre-conditioned by the SIP method. However, the
solver performance may vary for problems with highly curved boundaries as local refinement is
not straight forward for the present method. The two-dimensional test cases with both stationary
and moving boundaries show reasonable agreement with the previous experimental studies and
body-fitted numerical simulations. In three-dimensional simulations, parallelized calculations
yielded expected speed-up with the solution agreeing well with the literature.
1092 A K De

Acknowledgements

The present research was carried out through the funds available from the institute start-up
grant R&D/07/SG/ME/P/ARKD/1/2009-2010. The author thanks Prof. Vinayak Eswaran for his
invaluable inputs.

Appendix A

The discrete linear equations (A φ = b) arising from the governing equations can be cast into a
septa-diagonal form

aB φB + aS φS + aW φW + aP φP + aE φE + aN φN + aT φT = b, (23)

where aB = aT = 0 in the two-dimensional form. Diagonals of the lower (LW, LS, LP , LB)
and upper (U E, U N, U T ) triangular factors, obtained by the normal LU factorization procedure
along with the implicit relations proposed by Stone (1968), are given below
LSP = aS /(1 + γ (U TS + U ES ))
LBP = aB /(1 + γ (U EB + U NB ))
LWP = aW /(1 + γ (U NW + U TW ))
LPP = aP + γ (LWP U NW + LWP U TW + LSP U TS + LSP U ES
+LBP U EB + LBP U NB ) − LWP U EW − LSP U NS − LBP U TB
U EP = (aE − γ (LSP U ES + LBP U EB )/LPP
U TP = (aT − γ (LSP U TS + LWP U TW ))/LPP
U NP = (aN − γ (LWP U NW + LBP U NB ))LPP
with γ being a implicit factor and γ ≈ 0.9 gives the best convergence for a range of problems.
The above factors reduce the original linear system into a two-step substitution procedure

LU φ = b =⇒ LM = b and U φ = M, (24)

References

Angot P, Bruneau C H and Frabrie P 1999 A penalization method to take into account obstacles in viscous
flows. Numer. Math. 81: 497–520
Armfield S W 1991 Finite difference solutions of the Navier-Stokes equations on staggered and non-
staggered grids. Comput. Fluids 20: 1–17
Arthurs K M, Moore L C, Peskin C S, Pitman E B and Layton H E 1998 Modeling arteriolar flow and mass
transport using the immersed boundary method. J. Comput. Phys. 147: 402–440
Balaras E 2004 Modeling complex boundaries using an external force field on fixed Cartesian grids in
large-eddy simulations. Comput. Fluids 33(3): 375–404
Barton I E and Kirby R 2000 Finite difference scheme for the solution of fluid flow problems on non-
staggered grids. Int. J. Numer. Methods Fluids 33: 939–959
Berthelsen P A and Ytrehus T 2007 Stratified smooth two-phase flow using the immersed interface method.
Comput. Fluids 36(7): 1273–1289
Borazjani I, Ge L and Sotiropoulos F 2008 Curvilinear immersed boundary method for simulating fluid
structure interaction with complex 3d rigid bodies. J. Comput. Phys. 227(16): 7587–7620
 Non-staggered Cartesian grid method for incompressible viscous flows 1093

Bottino D C 1998 Modeling viscoelastic networks and cell deformation in the context of the immersed
boundary method. J. Comput. Phys. 147: 86
Demirdžić I, Lilek Ž and Perić M 1992 Fluid flow and heat transfer test problems for non-orthogonal grids:
benchmark solutions. Int. J. Numer. Meth. Fluids 15: 329–354
Denaro F M and Sarghini F 2002 2-d transmitral flows simulation by means of the immersed boundary
method on unstructured grids. Int. J. Numer. Methods Fluids 38: 1133–1158
der Wijngaart R J F V 1990 Composite grid techniques and adaptive mesh refinement in computational
fluid dynamics. Report CLaSSiC-90-97
Fadlun E A, Verzicco R, Orlandi P and Mohd-Yusof J 2000 Combined immersed-boundary finite-difference
methods for three-dimensional complex flow simulations. J. Comput. Phys. 161(1): 35–60
Fauci L J 1990 Interaction of oscillating filamentsa computational study. J. Comput. Phys. 86: 294
Ferziger J H and Perić M 1996 Computational methods for fluid dynamics. Springer 95–106
Fogelson A L and Peskin C S 1988 A fast numerical method for solving three-dimensional stokes equations
in the presence of suspended particles. J. Comput. Phys. 79: 50
Ge L and Sotiropoulos F 2007 A numerical method for solving the 3d unsteady incompressible Navier-
Stokes equations in curvilinear domains with complex immersed boundaries. J. Comput. Phys. 225(2):
1782–1809
Gilmanov A and Acharya S 2008 A hybrid immersed boundary and material point method for simulating
3d fluidstructure interaction problems. Int. J. Numer. Methods Fluids 56: 2151–2177
Goldstein D, Handler R and Sirovich L 1993 Modeling a no-slip flow boundary with an external force field.
J. Comput. Phys. 105: 354–366
Huang W-X and Sung H J 2009 An immersed boundary method for fluid-flexible structure interaction.
Comput. Methods Appl. Mech. Engrg. 198(33–36): 2650–2661
Iaccarino G and Verzicco R 2003 Immersed boundary technique for turbulent flow simulations. Appl. Mech.
Rev. 56: 331–347
Kim D and Choi J 2000 A second-order time-accurate finite volume method for unsteady incompressible
flow on hybrid unstructured grids. J. Comput. Phys. 162: 411–428
Kim J, Kim D and Choi H 2001 An immersed-boundary finite-volume method for simulations of flow in
complex geometries. J. Comput. Phys. 171(1): 132–150
Kim J and Moin P 1985 Application of a fractional-step method to incompressible Navier-Stokes equations.
J. Comput. Phys. 59: 308–323
Kirkpatrick M P, Armfield S W and Kent J H 2003 A representation of curved boundaries for the solution
of the Navier-Stokes equations on a staggered three-dimensional Cartesian grid. J. Comput. Phys. 184:
1–36
Labbé D F L and Wilson P A 2003 A numerical investigation of the effects of the spanwise length on the
3-d wake of a circular cylinder. J. Fluid Mech., 476: 303–334
Lai M-C and Peskin C S 2000 An immersed boundary method with formal second- order accuracy and
reduced numerical viscosity. J. Comput. Phys. 160: 705–719
Li C W and Wang L L 2004 An immersed boundary finite difference method for les of flow around bluff
shapes. Int. J. Numer. Methods Fluids 46: 85–107
Majumdar S, Iaccarino G and Durbin P 2001 Rans solvers with adaptive structured boundary non-
conforming grids. Annual Research Briefs 353–366
Maruoka A 2003 Finite element analysis for flow around a rotating body using chimera method. Int. J.
Comput. Fluid Dyn. 17: 289–297
McQueen D M and Peskin C S 1997 Shared-memory parallel vector implementation of the immersed
boundary method for the computation of blood flow in the beating mammalian heart. J. Supercomput.
11: 213
Mittal R and Iaccarino G 2005 Immersed boundary methods. Annu. Rev. Fluid Mech. 37: 239–261
Mittal S and Kumar B 2003 Flow past a rotating cylinder. J. Fluid Mech., 476: 303–334
Mohd-Yosuf J 1997 Combined immersed boundary/b-spline methods for simulation of flow in complex
geometries. Annual Research Briefs 317–328
1094 A K De

Peskin C S 1972 Flow patterns around heart valves: A numerical method. Journal of Computational Physics
10(2): 252–271
Rhie C M and Chow W L 1983 A numerical study of the turbulent flow past an isolated airfoil with trailing
edge separation. AIAA J. 21: 1525–1532
Saiki E M and Biringen S 1996 Numerical simulation of a cylinder in uniform flow: Application of a virtual
boundary method. J. Comput. Phys. 123: 450–465
Sohankar A, Norberg C and Davidson L 1998 Low-reynolds-number flow around a square cylinder at
incidence: Study of blockage, onset of vortex shedding and outlet boundary condition. Int. J. Numer.
Meth. Fluids 26: 39
Stone H L 1968 Iterative solution of implicit approximation of multidimensional partial differential
equations. SIAM J. Numer. Anal. 5: 530–558
Tai C, Zhao Y and Liew K 2005 Parallel computation of unsteady incompressible viscous flows around
moving rigid bodies using an immersed object method with overlapping grids. J. Comput. Phys. 207(1):
151–172
Tau E Y 1994 A 2nd-order projection method for the incompressible Navier- Stokes equations in arbitrary
domains. J. Comput. Phys. 115: 147–152
Tseng Y-H and Ferziger J H 2003 A ghost-cell immersed boundary method for flow in complex geometry.
J. Comput. Phys. 192(2): 593–623
Tucker P G and Pan Z 2000 A Cartesian cut cell method for incompressible viscous flow. Appl. Math.
Modell. 24: 591–606
Tyagi M and Acharya S 2005 Large eddy simulation of turbulent flows in complex and moving rigid
geometries using the immersed boundary method. Int. J. Numer. Methods Fluids 48: 691–722
Unverdi S O and Tryggvason G 1992 A front-tracking method for viscous, incompressible, multi-fluid
flows. J. Comput. Phys. 100: 250–37
Verma A K and Eswaran V 1999 An overlapping control volume method for Navier-Stokes equations on
nonstaggered grids. Int. J. Numer. Meth. Fluids 30: 279–308
Verzicco R, Iaccarino G, Fatica M and Orlandi P 2000 Flow in an impeller stirred tank using an immersed
boundary method. Annual Research Briefs 251–261
Williamson C 1988 The existence of two stages in the transition to three-dimensionality of a cylinder wake.
Phys. Fluids 31: 3165–3168
Williamson C 1996 Vortex dynamics in the cylinder wake. Annu. Rev. Fluid Mech. 28: 477–539
Xia G, Zhao Y and Yeo J 2009 Parallel unstructured multigrid simulation of 3d unsteady flows and fluid-
structure interaction in mechanical heart valve using immersed membrane method. Comput. Fluids
38(1): 71–79
Yang J and Stern F 2009 Sharp interface immersed-boundary/level-set method for wave-body interactions.
J. Comput. Phys. 228(17): 6590–6616
Ye T, Mittal R, Udaykumar H S and Shyy W 1999 An accurate Cartesian grid method for viscous
incompressible flows with complex immersed boundaries. J. Comput. Phys. 156: 209–240
Yoon D-H, Yang K-S and Choi C-B 2010 Flow past a square cylinder with an angle of incidence. Phys.
Fluids 22: 043603(1–11)
Zhang S-L 1997 Gpbi-cg: Generalized product-type methods based on bi-cg for solving nonsymmetric
linear systems. SIAM J. Sci. Comput. 18: 537–551
Zhang X, Ni S and He G 2008 A pressure-correction method and its applications on an unstructured chimera
grid. Comput. Fluids 37: 993–1010
Zhu L and Peskin C 2003 Interaction of two filaments in a flowing soap film. Phys. Fluids 15: 128–136