You are on page 1of 15

See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/256918344

A finite volume formulation for transient


convection and diffusion equations with
unstructured distorted grids and its
applications...

Article in Computer Methods in Applied Mechanics and Engineering January 2013


DOI: 10.1016/j.cma.2012.09.016

CITATIONS READS

4 90

2 authors, including:

Shicheng Xue
University of Sydney
41 PUBLICATIONS 614 CITATIONS

SEE PROFILE

All content following this page was uploaded by Shicheng Xue on 01 December 2015.

The user has requested enhancement of the downloaded file.


Comput. Methods Appl. Mech. Engrg. 253 (2013) 146159

Contents lists available at SciVerse ScienceDirect

Comput. Methods Appl. Mech. Engrg.


journal homepage: www.elsevier.com/locate/cma

A finite volume formulation for transient convection and diffusion equations


with unstructured distorted grids and its applications in fluid flow simulations
with a collocated variable arrangement
S.-C. Xue , G.W. Barton
School of Chemical and Biomolecular Engineering, The University of Sydney, NSW 2006, Australia

a r t i c l e i n f o a b s t r a c t

Article history: This paper describes a finite volume method for simulating transport processes governed by convection
Received 19 July 2012 diffusion type equations. The formulation is based on a cell-centred, unstructured grid. With an edge-
Received in revised form 24 September 2012 based data structure, discretisation is independent of control volume (or cell) shape. By using a surface
Accepted 27 September 2012
vector decomposition at the midpoint of the interface between cells, along with a deferred-correction
Available online 13 October 2012
approach, any cross-diffusion due to grid skewness can be readily accounted for when discretising the
diffusive flux. For modelling fluid flow processes, a collocated arrangement of variables is employed so
Keywords:
that a single coefficient matrix applies for the momentum equations of each velocity component. To avoid
Convective and diffusive fluxes
Finite volume formulation
checkerboard oscillation (arising from pressurevelocity decoupling) occurring under the collocated
Multi-dimensional flow variable arrangement when a pressure-based solution algorithm is employed, a novel pseudo-flux inter-
SIMPLER algorithm polation method is proposed for unstructured grids, ensuring that the solution is both under-relaxation
Unstructured grid factor and time-step (for transient calculation) independent. The methodology can be formulated within
a framework whereby either a coupled or a decoupled solution algorithm can be employed. The features
and advantages of the method are demonstrated by solution of the NavierStokes equations for two
benchmark flow problems.
! 2012 Elsevier B.V. All rights reserved.

1. Introduction ume methods (FVMs) [3], based on the conservation of mass,


momentum and energy over an arbitrarily shaped cell (or control
A wide range of transport processes in diverse fields can be volume), provide an attractive alternative to FEMs. Depending on
mathematically described by a generic convectiondiffusion equa- where dependent variables are stored, how the control volume
tion that requires numerical solution in all but the simplest cases. In (over which the conservation laws are applied) is defined, what
many computational scenarios, the solution domain can be very kind of data structure is used for convectivediffusive flux discret-
complex. Thus the ability to efficiently handle complex geometry isation, and what kind of solution strategy is implemented, a variety
can be a major consideration in developing suitable numerical sim- of FVMs with different features have been developed.
ulation methods. In this regard, finite element methods (FEMs) [1] In a vertex-centred formulation [49], dependent variables are
are often the preferred option. However, to ensure numerical con- stored (and solved for) at each vertex, and a control volume for
vergence and solution accuracy, there are a range of restrictions the vertex is constructed using the median dual of the cell grid
on FEM grid generation, such as using a single element shape (for associated with the vertex, as shown in Fig. 1. Depending on the
example, either triangular or a quadrilateral for two-dimensional data structure used, there are two different ways of discretising
cases) with conditions placed on the angles within an element the convectivediffusive fluxes across the control volume faces.
[2]. Considerable effort has been devoted to developing FEM mesh With an element-based structure [46], the local variation of a
generation algorithms that meet these requirements. However, in dependent variable inside a control volume is described by a piece-
some modelling situations, such as flow problems involving free wise polynomial function. This formulation is similar to the FEM
surfaces, an initial smooth mesh may be severely distorted during approach where the solution within each element is described by
the solution process. Therefore, numerical methods based on local shape functions. To make a FVM free of shape and internal
unstructured meshes, free of any shape or angle restrictions, are angle restrictions, and to allow hybrid meshes to be used in the for-
highly desirable. In this regard, the physically meaningful finite vol- mulation, any use of shape functions should be avoided, meaning
that the data structure used should be edge-based [7]. Certainly
Corresponding author. Tel.: +61 2 93518176; fax: +61 2 93527060. the use of such a data structure within a vertex-centred formulation
E-mail address: shicheng.xue@sydney.edu.au (S.-C. Xue). is computationally more efficient in terms of CPU time and memory

0045-7825/$ - see front matter ! 2012 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.cma.2012.09.016
S.-C. Xue, G.W. Barton / Comput. Methods Appl. Mech. Engrg. 253 (2013) 146159 147

Nomenclature

A area K convective coefficient


A area vector (or coefficient matrix) h volume fraction
b source term for a scalar equation q fluid density
b source term vector for a vector equation s stress tensor
B source term vector for a linear system X volume
d distance from a cell centre to a vertex x under-relaxation factor
ds distance from cell centroid to the midpoint of a face
^
d unit vector along a connecting line Subscripts
D vector along a connecting line bd boundary face designator
F flux f-f0 from point f to point f0
F flux vector f midpoint of a face
n
^ unit vector normal to a surface f0 intersection point of a connecting line with a face
H height U dependent variable
I unit tensor i coordinate index
m mass flow rate L left-hand side
max(x, y) maximum value in x and y max maximum value
min(x, y) minimum value in x and y n normal direction
N total number of control volumes (or cells) in the solu- NB neighbouring cell designator
tion domain nd node (or vertex)
NB neighbouring cell centroid P cell designator
p isotropic pressure R right-hand side
P cell centroid t tangential direction
R residual
Re Reynolds number Superscripts
r distance vector bd total number of boundary faces for a control volume (or
S volumetric source term for a scalar equation cell)
S volumetric source term vector for a vector equation 0 previous time value
T tangential vector to a face # intermediate value
^t unit vector tangential to a surface previous iteration
t time c convection
U speed d diffusion
V velocity vector FO first-order upwind differencing scheme
V velocity component HO higher-order scheme
x, y Cartesian coordinates k cell index
n normal direction
Greek symbols NC total number of neighbouring control volumes (or cells)
a, c scaling factors nd node or vertex
DM mass residual nf total number of faces bounding a control volume (or
d difference cell)
U dependent variable SO second-order linear interpolation
W position-dependent variable T transpose operation
C diffusive coefficient V volumetric
j blending factor
l fluid viscosity

usage [8], although there are challenges: the control volumes con-
structed around the vertices are usually larger than the cells, while
often flux evaluation for each edge is simplified using an approach control volume (cell)
based on the normal vector to the edge corresponding to the cen- cell centroid
troid dual. As a result, the local truncation error is usually larger
[9] than that for a cell-centred formulation [11] where the cells Cell B
themselves, formed by the grid-generation process, are chosen as
the control volumes and dependent variables are solved for (and
stored) at the cells geometric centre. In some situations, there is
the additional drawback that a vertex can become a geometrical Cell A
singularity, where mathematically the values of some field vari-
ables are not uniquely defined. midpoint
With a cell-centred formulation [1014], geometrical singulari-
ties can be avoided, while construction of dual control volumes is
vertex or node connecting line
unnecessary. In addition, if the mesh is orthogonal, such an ap-
proach is computationally efficient because low-order finite differ- Fig. 1. Schematic representation of an unstructured, hybrid mesh for a 2D
ence approximations along the straight line connecting adjacent discretisation.
148 S.-C. Xue, G.W. Barton / Comput. Methods Appl. Mech. Engrg. 253 (2013) 146159

cell centroids can be used for discretisation of the convective-diffu- is referred to as an interface. Thus, the total number of neighbour-
sive fluxes across each cell face without losing accuracy. For an ing cells NC for a given cell P is equal to the total number of faces nf
unstructured, non-orthogonal mesh, however, the connecting line bounding (or defining) the cell. A boundary cell has at least one of
between two cell centroids generally neither passes through the its faces (referred to as a boundary face) coincident with the solu-
midpoint of the face nor is normal to the face itself, as is the case tion domain boundary.
for an orthogonal mesh. How to accurately discretise the gradient The generic transport equation for a dependent variable U can
normal to a face becomes critically important in situations where be cast as follows:
the mesh is (or becomes during the solution process) highly dis-
@
torted. Resolution of this issue is one of the central aspects of the KU r $ KV U r $ CrU SU 1
@t
present study.
A related issue, arising in a pressure-based solution algorithm where the coefficients K and C have different meanings for differ-
for simulating processes involving fluid flow, is the so-called ent dependent variable U which can be a tensor, vector or a scalar.
checkerboard problem [3]. Here non-physical oscillatory behav- The source term SU lumps together all terms that cannot be accom-
iour can arise due to inappropriate decoupling between the pres- modated within the convection and diffusion terms, and is specific
sure and velocity fields. The traditional resolution of this to a particular U.
problem is to use a staggered grid arrangement. However, using The discretised form of Eq. (1) over a cell is obtained by inte-
different grids for different variables makes the method difficult grating it over the time interval Dt and the corresponding cell vol-
to extend to unstructured grids, particularly for three-dimensional ume X. The use of Gausss divergence theorem turns the volume
(3D) calculations [14]. integrals for the convection and diffusion terms into surface
Since the early 1980s, the use of non-staggered grids has be- integrals:
come increasingly popular in finite volume based approaches, lar- I I I
@
gely due to the momentum interpolation method (MIM) proposed KUdX KVU & CrU $ dA SU dX 2
@t X A X
by Rhie and Chow [15]. However, in the original MIM, under-relax-
ation factors in the discretised momentum equations were not Conversion of this integral equation into a linear algebraic equa-
considered. Independently, both Majumdar [16] and Miller and tion is achieved by invoking the midpoint rule. Accordingly, Eq. (2)
Schmidt [17] showed that converged solutions based on this can be expressed as:
MIM were under-relaxation factor dependent. Each proposed a
KP XP ! " Xnf h i
similar remedy, employing explicit under-relaxation in the inter- UP & U0P mf Uf & Cf rU $ n
^ f A f SP 3
polated convective velocities, although they only considered stea- Dt f
dy-state flow situations (see [18]). For unsteady-state problems, it H
where SP X SU dX SU P XP is the volumetric source term, while
was later found ([19,20]) that these interpolated velocities could be
subscripts P and f are the cell and face indicators, respectively. Thus,
influenced by the size of the time step used. Attempts have been
made to develop a time-step independent MIM (for example, mf KV $ n
^ Af 4
[21]), although all proposed methods have been based on an
orthogonal structured mesh system. In this study, we propose a is the mass flow rate across the face, and Uf is the value of the
modified MIM that is both under-relaxation factor and time-step dependent variable at the midpoint of the face. Here, Af n ^ f Af is
independent for non-orthogonal, unstructured grids of arbitrary a surface vector with n ^ f being the outward unit normal, while Af
cell shape. is the area of the face considered. The summation is over all of
The remainder of this paper is structured as follows. In Section 2, the faces (nf of them) bounding the cell, which can be an interface
the discretisation methodology for a generic convection shared by a neighbouring cell NB, or a boundary face bd if the face is
diffusion equation is discussed with special attention being paid coincided with the boundary of the solution domain.
to accommodating any cross-diffusion due to grid skewness. The fi- Due to its simplicity and unconditional (i.e. free of restrictions
nal discretised form of the equation, together with its solution, are on time-step size) stability for transient calculations, the first-
discussed in Section 3. In Section 4, the proposed formulation is ex- order backward Euler implicit formula is adopted for the temporal
tended to the NavierStokes equations for fluid flow simulations, approximation, although the method is of course not restricted to
while the issue of velocity and pressure coupling, as well as the this particular choice.
treatment of boundary conditions, are discussed in the following The final algebraic form of the discretised equation that links
section. In Section 6, the SIMPLER solution algorithm is imple- the value of the dependent variable at the cell centroid (i.e. UP)
mented, along with the derivation of an appropriate pressure cor- to the values at all neighbouring cell centroids (i.e. UNB with
rection equation. The overall methodology is validated in Section 1 6 NB 6 NC) will depend on how the convected values at the cell
7 by solving two benchmark flow problems from the literature, faces and the gradients normal to the cell faces are interpolated in
while concluding remarks are presented in Section 8. terms of UP and UNB. For ease of presentation, in this paper details
are only given for 2D formulation. However, extension to any 3D
case is straightforward, thanks to the formulation being cell shape
2. The generic transport equation and its discretisation independent.

The computational domain is discretised into non-overlapping 2.1. Convected values at a cell face
control volumes or cells (N in total). These cells are arbitrarily
shaped polyhedra (in 3D) or polygons (in 2D), as shown in Fig. 1 The estimation of the convected value of a dependent variable
where the unstructured mesh for a 2D discretisation is schemati- at a cell face (i.e. Uf) can be crucial for numerical accuracy and sta-
cally depicted. Cell vertices are referred to as nodes. Each cell with bility. Key issues are how to avoid false diffusion due to the use of
a volume X has an associated centroid denoted by P where depen- low-order interpolation schemes and the numerical instability of
dent variable U is stored and denoted as UP. The boundaries form- the resultant discretised system due to the use of unbounded high
ing a cell are called faces (or edges, in the 2D case) with an resolution (HR) schemes.
associated area vector A. An interior cell has no face that is part In the Appendix, for a general position-dependent quantity W,
of the domain boundaries, while a face shared by two adjacent cells the interpolation methods used to reconstruct its value at the mid-
S.-C. Xue, G.W. Barton / Comput. Methods Appl. Mech. Engrg. 253 (2013) 146159 149

point of the interface shared by two adjacent cells are discussed in That is, dUHO
f USO
f & Uf
FO
with the value of the blending factor
the framework of unstructured (and possibly highly distorted) j being set to 1, except when solution oscillations require some
grids along with the methods for gradient vector evaluation. Obvi- measure of damping, then a lower value (typically around 0.9) is
ously, none of the reconstruction methods can be directly applied used. As first-order upwinding is the only scheme which can en-
for the convected value for the reasons mentioned above. sure boundedness, with other schemes potentially becoming un-
From a physical viewpoint, convection represents the transport bounded [29], the deferred-correction technique of Khosla and
of information from upstream to downstream, and if the overall Rubin [30] is employed. That is, the correction term is evaluated
transport process is dominated by convection, one would expect using values from the previous iteration (as denoted by the over-
that the upstream value would exert a major influence on the value bars in Eq. (7)), and is lumped into the source term of the discre-
calculated at a cell face. Any robust numerical scheme should re- tised equation as a deferred-correction term. In this way, the most
flect this feature, and thus an appropriate profile for the convected desirable feature (i.e. the resultant system matrix being diagonally
value should have an upwind-biased form: dominant) inherent in first-order upwinding is maintained, while
high accuracy is achieved at a far lower computational cost than
Uf UFO HO
f jdUf 06j61 5
employing higher-order techniques involving information from
where cells far upstream. It should be noted that this scheme is employed
# because of its simplicity in implementation (i.e. being both cell
UP mf P 0 shape independent and easily extended to 3D cases), low computa-
UFO
f 6
UNB mf < 0 tional cost (e.g. free of monotonicity considerations), and good
numerical accuracy. The validity of the current methodology is,
represents the first-order upwind differencing scheme (henceforth
of course, not restricted to this choice alone.
FO), while dUHOf represents a higher-order (henceforth HO) correc-
Convected values across the boundary faces must either be
tion term (when j > 0) to the first-order scheme (j = 0).
specified (at an inflow boundary) or evaluated in the same way
To date, HR schemes of varying complexity have been devel-
as for an interior face (at an outflow), except for boundaries where
oped within the framework of a structured mesh grid [2227].
no flow crosses the faces (i.e. mf = 0), such as at a wall or symmet-
An upwind-biased high-order reconstruction of cell face values
rical plane.
from upstream and downstream cell averages, along with a
monotonicity criterion, are commonly used in all such schemes.
2.2. Gradients normal to the cell faces
Within an unstructured mesh framework, involving cells bounded
by an arbitrary number of faces, however, the use of a large sten-
With an orthogonal mesh, the connecting line for two adjacent
cil (i.e. involving cells far upstream) is problematic. Attempts
cells is normal to the face and passes through the face midpoint.
have certainly been made to extend these HR schemes to unstruc-
Thus, it is straightforward to evaluate the gradients along the line
tured meshes, for example, by assuming a virtual point far up-
normal to the face and passing through the face midpoint in terms
stream that is symmetrically located at the other end of the
of the values at the centroids of two adjacent cells sharing the face.
line connecting upstream and downstream cell centroids [28].
In an unstructured and non-orthogonal (possibly highly distorted)
Such a scheme is, however, hard to justify when applied to a
mesh, however, as considered in this study, the connecting line
highly skewed mesh, as the connecting line (i.e. the line joining
does not necessarily pass through the midpoint f of the face, nor
the cell centroid P and a neighbouring cell centroid NB as shown
is normal to the face itself, as shown in Fig. 2. In this case, in order
in Fig. 2) is not necessarily a streamline, nor does it necessarily
to accommodate possible mesh skewness, a vector decomposition
pass through the midpoint of the interface common to the two
is performed along a virtual line that passes through the midpoint
neighbouring cells, or in extreme cases may not even intersect
of the face and is parallel to the connecting line. This virtual line is
the interface (see the line connecting cells A and B in Fig. 1, for
created by moving (in a parallel fashion) the connecting line via the
example).
distance vector directed from the crossing point to the midpoint of
In the present work, instead of making an exhaustive evaluation
the face, that is, rf &f 0 rf 0 & rf . As shown in Fig. 2, values at the two
of alternative higher-order schemes or developing yet another
projected points (from cell centroids P and NB, respectively) P0 and
approximate scheme, a blend of the first-order upwind differenc-
NB0 can be obtained using a second-order linear interpolation as gi-
ing scheme and second-order linear interpolation (see Eq. (A.4)
ven in Eq. (A.4) from the values at the corresponding cell centroids.
in the Appendix) is employed:
! " The unit normal of this virtual line can be decomposed as:
Uf UFO SO FO
f j Uf & Uf 7 ^ c ^t
^ f Df Tf af d
n f f f 8

Here Df af d ^ is a vector along the virtual line with d ^ being


f f
the unit vector of the connecting line, while af is a scaling factor.
L Tf cf ^tf is a vector tangential to the face with ^tf being the unit tan-
gential vector while cf is a second scaling factor dependent on af as
n T
follows:
NB

P D cf &af ^tf $ d^ f 9
f NB
Thus, the only requirement is to determine af . A restriction here
f
is that af should be unity on an orthogonal mesh, so that the dis-

P cretisation reverts to classical stencils. Several options are available
R [31,32] but in this work,
1
af ^
10
n
^f $ d f
Fig. 2. Vector decomposition along a virtual line passing through the midpoint f of
the interface between two adjacent cells with their centroids located at points P and is used as its value is always greater than or equal to unity, thus
NB. enhancing the diagonal dominance of the coefficient matrix of the
150 S.-C. Xue, G.W. Barton / Comput. Methods Appl. Mech. Engrg. 253 (2013) 146159

discretised equations, especially in the cases when the mesh is se- Cbd Abd
abd abd max&mbd ; 0 17
verely skewed. dsbd
With vector decomposition along this virtual line, the gradient
The coefficient for the temporal term takes the form:
along the normal direction at the midpoint of the face can be cast
as: KP XP
a0P 18
8a af
Dt
< dsf UNB & UP ds rUNB & rUP $ rf &f 0 cf rUf $ ^tf for an interior face Pnf
rU $ n
^ f f f
In the last term, DM P f mf is the net mass residual for cell P
: abd Ubd & UP c rUbd $ ^tbd for a boundary face
ds bd bd which is zero if mass conservation is satisfied in the flow field.
11 However, this term can be non-zero during the course of iteration
in simulations involving fluid flow before the velocity field is
where ds is the distance from cell centroid P to its neighbouring cell
converged.
centroid NB (for an interior face), or to the midpoint of a boundary
The term bP in Eq. (14) is the source term which takes the form:
face.
In Eq. (11), all terms other than the first are higher-order correc- bP SP abd Ubd F P maxDMP ; 0 & DMP (U)P 19
tions accounting for cross-diffusion due to skewness of the mesh.
where the superscript means values from the previous iteration
The gradient vector rUf can be estimated using the values of the P
level. The term F P nf c d
f F f F f includes all deferred higher-order
dependent variable at the two vertices L and R forming the face
correction terms arising from the discretisation of the convective
(see Fig. 2), so that correction for any non-orthogonality (i.e. the
and diffusive fluxes, which are:
connecting line not being normal to the face) can be simplified
as follows: F cf &jmf dUHO
f 20

cf rUf $ ^tf cf UL & UR =Af 12 8 CA


>
> a f f rUNB & rUP $ rf &f 0
< f dsf
For a boundary face, since the connecting line always passes F df Cf c rUf $ ^tf Af for an interior face 21
>
> f
through the midpoint of the face, we only need to correct for :
Cbd cbd rUbd $ ^tbd Abd for a boundary face
cross-diffusion due to the non-orthogonality of the face itself. Thus,
for a Dirichlet boundary condition, the gradient vector rUbd on the Here, a top bar means that the corresponding item is evaluated
boundary face can be directly evaluated using the boundary nodal using values from the previous iteration.
values imposed. For a Neumann type boundary condition, where By performing discretisation over all cells within the solution
rU $ n^ bd is specified, Eq. (11) can be used to update the boundary domain, a nominally linear system AU U BU is generated. Here
value Ubd as follows: AU is a diagonally dominant coefficient matrix that is always spare
"
-bd
# due to the irregular connectivity for unstructured grids, while BU is
dsbd cbd nfX the source term vector. Because of the possible dependence of the
Ubd UP rU $ n
^ bd & Uf Af $ ^tbd 13
abd XP f coefficients and the source terms on the dependent variables
themselves, the system is only nominally linear, and an iterative
where nf-bd means that the boundary face itself is excluded from method is needed to handle the inherent nonlinearity. To prevent
the summation. divergence when solving highly nonlinear problems, under-relaxa-
tion in a manner described in [3] is always necessary. Accordingly,
3. Discretised equation and solution methodology Eq. (14) is rewritten as:
NC-bd
X
Using reconstructed convected values (Eq. (5)) and gradients aP UP aNB UNB bP a0P U0P 1 & xaP U)P 22
(Eq. (11)) to evaluate the convective and diffusive fluxes across NB
each cell face in Eq. (3), and with the higher-order correction terms
deferred (i.e. evaluated using values from the previous iteration) where x is the under-relaxation factor. Note here that the coeffi-
and lumped into the source term, for each cell, the resulting discre- cient aP has been redefined as aP =x.
tised equation can be cast into the form: For a scalar variable, with the mass flow rate mf assumed
NC-bd
known, this system can be readily solved by employing a linear sol-
X
aP UP aNB UNB a0P U0P bP 14 ver [33]. In this study, LU decomposition with pivoting is em-
NB ployed. For steady-state calculations, convergence is assumed
when the maximum volume-weighted, square-root residual of
where the summation is over all neighbouring cells (denoted as NB) the discretised equation for U, defined as:
with which cell P shares an interface (denoted as f) but excluding v

u 2 #2 ,
any cells that are outside the boundary (as indicated by NC-bd). uX N NC-bd N
u X X
The coefficient aNB linking the values at cell P and its neighbouring RU t Xk 4aP UP & aNB UNB & bP & 1 & xaP U)P Xk
cell NB takes the form: k1 NB k1

C f Af 23
aNB af max&mf ; 0 15
dsf
becomes smaller than a terminating value, typically 10 . &6

The central coefficient takes the form: For problems where the velocity field is unknown, the Navier
NC-bd
X Stokes equations need to be solved, either in a coupled or decou-
aP aNB abd a0P maxDM P ; 0 16 pled manner, as will be addressed in the following section.
NB

in which the coefficient abd will be zero if all faces bounding the 4. Discretisation of the NavierStokes equations
cell are interfaces, otherwise it links the values at a boundary cell
centroid P and the midpoint of a boundary face bd and takes the The NavierStokes equations that govern incompressible fluid
form: flow in the absence of any body force are given by:
S.-C. Xue, G.W. Barton / Comput. Methods Appl. Mech. Engrg. 253 (2013) 146159 151

r $ qV 0 24 nf
X
SP sf & lrVf ( $ n
^ f Af 33
@ f
qV r $ qVV r $ &pI s 25 Pnf
@t and FP Fcf Fdf . Thus bP can be cast as:
f
which represent mass and momentum conservation, respectively. nf
X
Here the total traction force is decomposed into two parts: the bP FSf abd Vbd maxDM P ; 0 & DM P (V)P 34
hydrostatic pressure force (with I being the unit tensor), and the dy- f
namic force with s being the extra stress tensor which is related to
with
kinematic quantities by an appropriate constitutive equation. For a
Newtonian fluid, this stress tensor is proportional to the deforma- FSf sf & lrVf ( $ n
^ f Af Fcf Fdf 35
tion rate, that is:
For an interior cell, FSf can be simplified as:
s lrV rVT 26
FSf T
lrV f $ n
^ f Af Fcf Fdf 36
As seen, both mass and momentum conservations can be recast
For cells along the boundaries that are not coincident with the
in the form of a generic transport (i.e. convectiondiffusion) equa-
Cartesian coordinates, however, it is more convenient if the local
tion, with each having a different dependent variable. For the mass
tangential-normal coordinate system is employed for the traction
conservation equation, we have U 1; SU 0. While for the
force term, that is:
momentum equation, we have:
FSbd F sbd & lrVbd $ n
^ bd Abd Fcbd Fdbd 37
@
qV r $ qVV r $ lrV SV & rp 27
@t where the traction force vector is expressed as:

where the pressure force term has been separated from the source Fsbd sbd $ n
^ bd Abd snn n
^ snt^tbd Abd 38
term SV which thus takes the form:
In this form, stress components normal and tangential to the
SV r $ s & lrV 28 boundary can either be directly replaced by a natural traction con-
dition (if it is imposed) or evaluated in terms of the velocity vectors
Thus, the discretised forms of these equations can be written
at the midpoint of the boundary face and the centroid of the
out as:
boundary cell to which the face belongs. For example, along a
nf
X nf
X no-slip wall boundary with the tangential velocity Vt,bd being
mf qV $ n
^ Af 0 29 either zero (i.e. a solid wall) or constant (i.e. a moving wall), the
f f normal stress vanishes for a Newtonian fluid. Accordingly, the trac-
tion force vector reduces to Fsbd snt^tbd Abd which after some
NC&bd
X manipulation can be expressed in terms of the velocity vector Vp
aP VP aNB VNB bP a0P V0P 1 & xaP V)P & XP rpP 30 at the boundary cell centroid and the imposed boundary velocity
NB
Vt,bd:
where the coefficients have forms analogous to Eqs. (15)(18) for a
lbd Abd
scalar variable (with parameters K and C replaced by q and l, Fsbd abd Vt;bd & VP $ ^tbd ^tbd ; 39
dsbd
respectively), with the pressure force term evaluated separately
via the divergence theorem (see Eq. (A.5)) as: Similarly, along a symmetrical or slip wall boundary, the shear
! (i.e. tangential) stress is zero, and the traction force vector reduces
nf
X
1 to Fsbd snn n
^ bd Abd which can be cast in terms of the velocity vec-
rpP pf Af 31
XP f tor at the boundary cell centroid:

where the pressure pf at the midpoint of a face is linearly interpo- lbd Abd
Fsbd &2abd VP $ n
^ bd n
^ bd 40
lated from values at two adjacent cell centroids (see Eq. (A.4)) for an dsbd
interior face. If the face is on a boundary, then depending on the
boundary conditions, the pressure at the face is either specified,
5. Coupling between velocity and pressure
interpolated or extrapolated from values at the centroids of cells
adjacent to the boundary.
Pressure is the driving force for the kinematical field, and only a
As pressure is also an unknown variable, the determination of
correct pressure field can ensure that mass conservation is realised.
the kinematic field requires an equation describing pressure varia-
To date, the majority of solution algorithms developed [3,3437]
tion. This can be a major issue in modelling fluid flow processes
have been based on a pressure-correction procedure whereby an
and will be discussed in more detail in the next section.
equation for pressure is obtained by enforcing the continuity con-
Discretisation of the momentum equations is independent of
straint, with the momentum equations retaining their role in
the velocity vector components, and thus one form of discretised
determining the velocity field.
equation can be applied for all governing equations if a collocated
As shown in the discretised continuity relation Eq. (29), to eval-
arrangement (i.e. all unknowns are solved and stored at the cell
uate the mass flux across a face, the convective velocity Vf at the
centroid) is used, which is highly desirable in terms of solution effi-
midpoint of the face is needed. However, with a collocated
ciency and storage requirements. However, care has to be taken in
arrangement, what is obtained from the discretised momentum
evaluating the source vector bP as addressed below.
equation is the velocity vector at a cell centroid. From Eq. (30),
Explicitly, bP takes a form as given by Eq. (19) but with a scalar
we have:
replaced by a vector:
bP SP abd Vbd FP maxDMP ; 0 & DMP (V)P 32 aP VP HP V a0P V0P 1 & xaP V)P & XP rpP 41

where the volumetric source term now becomes: where


152 S.-C. Xue, G.W. Barton / Comput. Methods Appl. Mech. Engrg. 253 (2013) 146159

NC-bd
X where, when interpolating Hf V, both forms of the pseudo-flux vector
HP V aNB VNB bP 42 at the cell centroid as shown in Eqs. (42) and (43) can be used if a segre-
NB
gated solution algorithm is employed. With a coupled solution strategy,
however, Eq. (43) should be used so that the coupling between VP and pP
is a pseudo-flux vector at the cell centroid, which can also be ex-
is explicitly expressed in the discretised momentum equations.
pressed as:

HP V aP VP & a0P V0P & 1 & xaP V)P XP rpP 43 5.1. Pressure equation

The velocity vector at the cell centroid is then expressed as: Coupling between the velocity and pressure fields is obtained
^ P A0 V0 1 & xV) & GP rp
VP V 44 by enforcing the continuity constraint (see Eq. (29)) via the recon-
P P P P
structed convective velocity at the midpoint of face f (see Eq. (46)),
that is:
where GP XP =aP , A0P a0P =aP , and V
^ P HP V=aP is a pseudo-veloc-
ity vector. nf
X nf h
X i
mf ^ Af A0f qV0 $ n
^ $n
qV ^ Af 1 & xqV) $ n
^ Af & qGrp $ n
^ Af 0
Using linear interpolation of the velocities at adjacent cell cen-
f f
troids to obtain the convective velocities at the midpoint of the
51
face shared by two adjacent cells may lead to the generation of a
checkerboard pattern in the pressure field. As previously noted, or alternatively:
this can be avoided via the widely used MIM approach [15], where nf h
X i Xnf
the equation for Vf is reconstructed by averaging the discretised K f rp $ n
^ f m
ef 52
momentum equations for the velocity vectors at two adjacent cell f f
centroids as that:
with K qGA and
^ f A0 V0 1 & xV) & Gf rp
Vf V 45
f f f f
~f m
m ^ f A0f m0f 1 & xm)f 53
Here the pseudo-velocity vector and all coefficients are ob-
where m ^ Af ; m0f qV0 $ n
^ $n
^ f qV ^ Af ; m)f qV) $ n
^ Af are the
tained by averaging the corresponding values at two adjacent cell
centroids, but with rpf (i.e. the pressure gradient at the midpoint pseudo-mass flux across face f, the mass flux at the previous time
of face f) used in the pressure force term. Unfortunately, it has be step, and the mass flux at the previous iteration level, respectively.
shown [1617] that the resultant Vf is both under-relaxation factor Now, with the vector decomposition along the virtual line passing
and time-step dependent. To remedy these problems in the origi- through the midpoint of the face, the pressure gradient along the
nal MIM approach, in the present work, we reconstruct the equa- normal direction can be discretised in the same way as show in
tion for the convective velocity at the midpoint of a face by Eq. (11). Accordingly, the resultant pressure equation takes the
mimicking the form of the discretised momentum equation for a same form as for a scalar variable (see Eq. (14)) in the absence of
cell centroid as: any convective or temporal terms, that is:

af Vf Hf V a0f V0f 1 & xaf V)f & Xf rpf 46 NC-bd


X p
aP pP aNB pNB bP 54
Here the pressure gradient rpf at the midpoint of face f is used NB

in the pressure force term, while all coefficients (as well as the
Here coefficients have the same form as given by Eqs. (15)(17)
pseudo-flux across the face) are obtained using the volume-
for a scalar variable with parameters K 0 and C qG, but with-
weighted average (see Eq. (A.1) for details) from the corresponding
out any convective or temporal terms (that is, max&mf ; 0 0 and
values for two adjacent cell centroids. That is,
a0P 0).
af hNB aP hP aP NB ; a0f hNB a0P hP a0P NB ; The source term for the pressure equation can be cast as:
Xf hNB XP hP XNB 47 nf
X nf
X
p
bP & m ~ f abd p)bd F df 55
and f f

Hf V hNB HP V hP HNB V 48 where F df has the same form as in Eq. (21) with C qG and U p.
Accordingly, the convective velocity at face f obtained takes the The final form of the discretised pressure equation depends on
same form as Eq. (45) but with a different pseudo-velocity vector the solution algorithm used. If a coupled solution method is used,
then V ^ f in the source term needs to be expressed in terms of VP .
and coefficients. Now, we have:
This can be done using a pseudo-flux interpolation in which
^ f Hf V=af ;
V A0f a0f =af ; Gf Xf =af 49 HP V in Eq. (43) is used so that an equation in terms of pP and
VP is obtained. When this is solved together with the discretised
Our approach is different from the original MIM in that, instead momentum equation (Eq. (30)), the dynamic and kinematic fields
of interpolating the pseudo-velocity vectors, we interpolate the are obtained simultaneously. With a decoupled solution method,
pseudo-fluxes at two adjacent cell centroids to reconstruct Vf . In such as SIMPLE and its variants, both forms of the pseudo-flux vec-
this way, the resultant convective velocity vector is both under- tor at the cell centroid as shown in Eqs. (42) and (43) can be used in
relaxation factor and time-step independent, which can be easily evaluating V ^ f . In this solution option, we do not need to store the
verified by setting V0f V)f Vf (i.e. when the solution has con- velocity vector at the cell centroids from the previous iteration le-
verged) in Eq. (45) with Eq. (49) being used for V ^ f ; A0 and Gf .
f vel if Eq. (42) is used.
The convergent convective velocity vector takes the form:
bHf V & Xf rpf c 5.2. Issue with the pressure value along a boundary
Vf D h!P " i h!P " iE
NC-bd NC-bd
hNB aNB maxDMP ;0 hP aNB maxDMP ;0
NB
P
NB
NB From the previous analysis, it can be seen that for a boundary
50 cell in which one or more bounding faces are coincident with the
S.-C. Xue, G.W. Barton / Comput. Methods Appl. Mech. Engrg. 253 (2013) 146159 153

boundary of the solution domain, the mass flux across the bound- qbd DP Abd
abd abd 58
ary face, if it is specified, is directly used to enforce the continuity dsbd
constraint for the boundary cell, which leads to abd 0. Thus, no
information about the boundary pressure value is needed in solv- 6. Solution algorithm for fluid flow scenarios
ing the pressure equation. Therefore, if the mass fluxes across all
boundaries of the solution domain are specified in some way, then For fluid flow problems, either a coupled or a decoupled ap-
the resultant system matrix for the pressure equation will become proach can be used to solve the discretised momentum equations
singular. In this case, how to ensure global mass conservation be- and the pressure equation to obtain the kinematic and dynamic
comes a critical issue. How this matter can be resolved within fields. In this work, a decoupled solution method based on the SIM-
our modelling framework will be addressed in a follow-up paper. PLER algorithm [3] is employed. Three steps are involved in each
Once the pressure equation is solved, pressure values within the iteration, namely: (i) pressure field estimation, (ii) intermediate
domain are obtained. However, values along boundaries where velocity prediction, and (iii) velocity correction. This procedure is
mass fluxes are specified still have to be determined, not only for repeated until the maximum residual of the discretised governing
general post-processing, but also for evaluating the pressure gradi- equations falls below a nominal value (here 10&6).
ents (via the divergence theorem) in the momentum equations for The intermediate velocity vector denoted as V# (in Step 2) is ob-
any cell with one or more bounding faces coincident with the tained by solving the momentum equation (see Eq. (46)) in which
boundary. the estimated pressure p# (obtained in Step 1) is used, that is:
For those cases where no flow is crossing the boundary (that is, ^# 0 0
V#
f Vf Af Vf 1 & xVf & Gf rpf
) #
59
mbd 0), such as a solid wall or a symmetrical boundary, then from
Eq. (52), we have rp $ n ^ bd 0. That is, a Neumann boundary con- With an estimated pressure field, however, the resultant veloc-
dition is obtained for the pressure equation, and the boundary ity field is not expected to satisfy the continuity constraint. So it is
pressure value can be obtained by interpolation as shown in Eq. assumed that an improved velocity field that satisfies continuity
(13). can be obtained by solving the momentum equation in which an
Where flow is crossing a boundary, such as inflow and outflow improved pressure is used:
in flow problems involving open boundaries, the boundary pres-
^ f A0 V0 1 & xV) & Gf rp
Vf V 60
sure values have to be linearly extrapolated from values at the f f f f
boundary cell P and its nearest interior cell NP, as shown in
Thus, using Eqs. (59) and (60) leads to:
Fig. 3. Here P0 and NP0 are projections of P and NP onto the line nor-
Vf & V # ^ ^#
f V f & Vf & Gf rpf
mal and pass through the midpoint of the boundary face with the 0
61
normal being n ^ bd . The nearest interior cell NP for a boundary cell P
is determined as follows. From the midpoint of the boundary face, where p0f pf & p#
f is the pressure corrector.
draw a line normal to the face, and determine the projected dis- Pnf P
By requiring f mf nf f qV $ nAf 0, and dropping the term
^
tance H rNP & rbd $ ^tbd between the cell centre and the normal ^f & V
V ^ in Eq. (61), we have a semi-discretised pressure correc-
#
f
line for all neighbouring cells. The one with the minimum pro- tion equation that has a similar form to the semi-discretised pres-
jected distance is treated as the nearest interior cell. The boundary sure equation (see Eq. (52)), and so the same discretisation
pressure value is then obtained by linear extrapolation as: procedure can be employed. The resultant pressure correction
rbd & rP $ n
^ bd equation can be expressed as:
pbd pP0 p 0 & pP0 56
^ bd NP
rNP & rP $ n NC-bd
X p0
aP p0P aNB p0NB bP 62
where pP0 pP rpP $ rP0 & rP and pNP0 pNP rpNP $ rNP0 & rNP . NB
However, for the special case where there is a fully developed
where the coefficients are the same as those for the pressure equa-
outflow boundary, we have V $ n ^ bd 0 and the reference pressure
tion, but with a different source term:
value pbd can be specified along the outflow. With the Neumann
boundary condition for the momentum equations, the velocity vec- nf
X nf
X
p0
tor at the midpoint of the boundary face can be interpolated using bP & m#f abd p0
bd F df 63
f f
Eq. (13). Accordingly, the mass flux across such a boundary face
can be expressed as: As the pressure corrector tends to zero over the entire domain
qbd DP Abd upon solution convergence, all terms containing the gradient of
mbd qbd V ^ bd Abd A0P m0bd 1 & xm)bd & abd
^P $ n
the pressure corrector can be dropped, that is, F df + 0. Once p0 is ob-
dsbd
tained by solving Eq. (62), the velocity vector at a cell centroid and
* pbd & pP & qbd DP cbd rpbd $ ^tbd Abd 57
the mass flux across a face are corrected (Step 3) as follows:
By using this mass flux across the boundary face in the discre- nf
1X 0
tised mass conservation equation, the coefficient abd for the pres- VP V#
P & p Af 64
sure equation can be obtained as: aP f f

mf m#
f & aNB pNB & pP
0 0
65

As can be seen, the source term for the pressure correction


Pnf #
bd equation is basically the mass imbalance (i.e. DM # P f mf ) in
NP P the intermediate velocity field. As this is approaching zero, no fur-
ther pressure correction is needed. Note that in this context, we de-
NP P n bd fine the residual for the pressure correction equation as:
v
,
u N N
uX 2
X
Rm t Xk DM# k Xk 66
k k
Fig. 3. Boundary pressure extrapolation.
154 S.-C. Xue, G.W. Barton / Comput. Methods Appl. Mech. Engrg. 253 (2013) 146159

7. Methodology validation using benchmark problems 0.2 1

Key features of our proposed finite volume formulation are now 0.1 0.8
demonstrated by solving two well-researched benchmark prob-
lems from the fluid flow literature. 0.6
0
Mesh1
7.1. Lid-driven square cavity Mesh2 0.4
Mesh3

Vy

Vx
-0.1
Laminar flow of an incompressible fluid within a square cavity Benchmark 0.2
with a moving lid has long served as a standard benchmarking -0.2
problem for new solution algorithms and their coding because of 0
its inherent geometric simplicity yet strongly elliptical character.
-0.3 -0.2
As this problem and its solution are well documented in the liter-
ature (for example, [38]), a detailed description will not be repro-
duced here. -0.4 -0.4
0 0.25 0.5 0.75 1
To demonstrate our proposed methodology in accommodating x, y
an arbitrary hybrid mesh structure, as well as handling severe dis-
tortion in a skewed mesh, calculations were performed for three Fig. 5. Velocity profiles along the horizontal and vertical centre lines for Re = 100.
different mesh topologies, each having a similar cell density as
shown in Fig. 4. Mesh 1 had uniform quadrilateral (36 * 36) cells
with 1225 cells in total; Mesh 2 had triangular cells with 1298 cells 0.4 1
in total; Mesh 3 was a hybrid of distorted quadrilateral and triangle 36x36 (1225 cells)
81x81 (6400 cells)
cells with 1338 cells in total. Our calculation results are compared triangle mesh (6304 cells) 0.8
with the standard benchmark results of Ghia et al. [38] based on a 0.2 Benchmark

second-order differencing scheme and a uniform Cartesian (i.e. 0.6


structured) mesh of 129 * 129 (containing 16384 cells which is
0
over an order of magnitude more than our Meshes 13). 0.4
Vy

Vx
To test the accuracy of our proposed method, Mesh 1 was used
for a Reynolds number Re qUH=l 100. An under-relaxation 0.2
-0.2
factor x = 0.95 was used for the momentum equations only. The
velocity profiles along the horizontal and vertical centre lines are 0
compared with the benchmark solutions in Fig. 5, and as can be -0.4
seen, excellent agreement is achieved despite our much coarser -0.2
mesh.
To test the ability of our method to handle different mesh -0.6 -0.4
0 0.2 0.4 0.6 0.8 1
structures and mesh distortion, calculations were repeated with x, y
Mesh 2 (x = 0.9) and Mesh 3 (x = 0.75). Velocity profiles along
Fig. 6. Velocity profiles along the horizontal and vertical centre lines for Re = 400.

(a) Mesh 1 (b) Mesh 2


1 1 the horizontal and vertical centre lines in these cases are also com-
pared with the published benchmark solution in Fig. 5. Our method
handles distorted meshes very well indeed, with no divergence
being encountered as was experienced by McBride et al. [39]
who used a cell-centred collocated method for their distorted,
0.5 0.5 non-orthogonal meshes, together with a stronger under-relaxation
y

factor (x = 0.7). From Fig. 4(d), where the contour of the x-compo-
nent of the velocity predicted with Mesh 3 is shown, it can be seen
that our solutions are smooth and cell-shape independent, with no
0 0 localised abrupt changes being observed at locations where there
0 0.5 1 0 0.5 1
x x are dramatic changes in cell shape. However, a somewhat stronger
under-relaxation factor was required for the momentum equations
(c) Mesh 3 (d) Velocity Contour with increasing mesh distortion and non-orthogonality. As a result,
1 1
the total number of iterations required for convergence is in-
creased from 27 for M1, to 73 for M2 and 202 for M3, the increases
being mainly due to mesh distortion. While, with a similar cell
density, the CPU time per iteration seems to be largely dependent
0.5 0.5 on the orthogonality in the mesh structure. With M1 (orthogonal
y
y

mesh), for example, it only takes 1.36 s per iteration (on an Intel
QX9300 processor with 2.53 GHz). With the M2 mesh, which con-
tains many non-orthogonal cells, the CPU time per iteration in-
0 0 creases to 1.63 s. For the M3 case, although the mesh in some
0 0.5 1 0 0.5 1 areas is quite distorted, it is orthogonal in other areas. Accordingly,
x x
the CPU time per iteration is 1.52 s, which is intermediate between
Fig. 4. Mesh structures used and the contour of the x-component of velocity as the other two cases. Therefore, the overall CPU time for a particular
predicted with Mesh 3. simulation is not solely depending on mesh density. It is also
S.-C. Xue, G.W. Barton / Comput. Methods Appl. Mech. Engrg. 253 (2013) 146159 155

0 0.06

0.05
-0.05 0.7
0.8
0.04
-0.1 0.95

(Vx)m

(Vy)m
0.03

-0.15
0.02

-0.2
0.01

-0.25 0
100 200 300
Iteration

Fig. 7. Evolution of velocity components against iteration number for different under-relaxation factors and Re = 100.

100

Mass imbalance
10
-2

Momentum equation for Vx

10
-4 Momentum equation for Vy

= 0.7
10-6

10-8
Residuals

= 0.8
10
-10

Fig. 9. Schematic view of the backward-facing step flow case study.


10
-12

10-14 = 0.95
0.8 and 0.95). Evolution of the velocity components Vx and Vy at
10-16 a central monitoring point (i.e. x = y = 0.5) with iteration number
for these under-relaxation factors are presented in Fig. 7. Each
10
-18
velocity component converges to the same final value, although
100 200 300
as expected more iterations are required for the stronger under-
Iteration
relaxation factor. Fig. 8 shows the corresponding evolutions of
Fig. 8. Evolution of residuals for the momentum equations and the global mass the residuals for the momentum equations and the mass imbal-
imbalance with iteration number for Re = 100. ance with the number of iteration. With the nominated convergent
criterion (i.e. the maximum residual of all governing equations is
less than 10&6), the iterations required for the three under-relaxa-
strongly related to the degree of mesh distortion and its non- tion factors are 27, 109 and 183, respectively.
orthogonality.
With Mesh 1, calculations were also carried out for Re 400. As 7.2. Laminar flow over a backward-facing step
shown in Fig. 6, the local minimum and maximum velocities are
well resolved compared to results reported by Lee and Tzong [40] Laminar flow over a backward-facing step (shown schemati-
who used an 81 * 81 Cartesian mesh. If the same (6400 cells) uni- cally in Fig. 9) has both a simple geometry yet rich flow behaviour,
form mesh is used with our formulation (and an under-relaxation manifested by flow separation, flow reattachment and multiple
factor of 0.95), the results obtained are in excellent agreement with recirculation zones. Flow behaviour in the channel depends on
benchmark solutions obtained with a much finer mesh (129 * 129 the Reynolds number Re as well as geometrical parameters such
cells) after 110 iterations. If a triangular mesh is used at a similar as the expansion ratio of the step height h to the channel height
cell density, our results again agree well with reported benchmark H. Thus, this problem has long been used as a benchmark to test
solutions, although additional iterations (,400) are required as a the accuracy of proposed numerical methods.
somewhat stronger (0.85) under-relaxation factor was employed. A typical expansion ratio is 2. Experimentally [41], it has been
To check the under-relaxation factor dependence of our ap- shown that flow is basically two-dimensional (2D) with respect
proach, calculations for Re 100 with Mesh 1 were performed to the plane of symmetry for Re below approximately 400, where
with three different under-relaxation factors for velocity (x = 0.7, Re is based on the average inflow velocity and the channel height
156 S.-C. Xue, G.W. Barton / Comput. Methods Appl. Mech. Engrg. 253 (2013) 146159

12 heights), and a very fine mesh containing some 4251 * 101 points,
*
+
x Erturk [45] was able to obtain numerical solutions up to Re = 3000.
Rouizi (2009) +
Although the inlet and exit boundary conditions certainly impact
+ Erturk (2008) *
x
+ the ultimate stability of the solution obtained, most numerical
10 * Zhang (1994)
simulations have demonstrated the main features of this laminar
x Rogers (1991) +
Present flow: for Re 6 800, downstream of the step entrance there is a main
recirculation region whose length increases with Re, while in the
*
x
+ range 400 6 Re 6 800, a second recirculation zone appears that is
8
attached to the upper wall of the flow channel. X1, X2 and X3 (see
25
X1/h

Fig. 9 for definitions) are the (Re dependent) lengths that define
+
20 + these various flow characteristics.
6
X3/h + In this work, however, we are not seeking to investigate the
15 +
+ instability features of the flow, but rather to demonstrate the accu-
*
x
+
10+ X2/h
racy of our proposed procedure when compared to experimental
+ + +
+ + data and other numerical solutions found in the literature. Thus,
4 5 an expansion ratio of 2 is used, while the Re range was limited to
0
lie between 100 and 800. A fully developed parabolic profile, that
+
x 400 500 600 700 800
is, V x y 24y0:5 & y and V y y 0 for 0 6 y 6 0:5, is imposed
Re
at the inlet step boundary. The length of the computational domain
2
0 200 400 600 800 1000 is 30h, as this is assumed long enough to allow the flow to become
Re fully developed at the outflow boundary for the Re range chosen. A
151 * 61 uniform quadrilateral mesh was used, with simulations
Fig. 10. Reattachment length X1/h as a function of Reynolds number (Re). Locations
performed at five different Re values (100, 200, 400, 600 and 800)
of the detachment (X2/h) and reattachment (X3/h) points against Re are shown in the
inset. with x in the range 0.85 to 0.9.
Fig. 10 presents the evolution of the dimensionless lengths X1/h,
X2/h and X3/h with respect to Re. Our results are compared with
H. It is widely accepted that flow features are stable and steady up published values from [4548]. Contours of the axial velocity Vx
to around Re = 800, although beyond Re = 400, 3D behaviour has at Re = 200, 400 and 800 are displayed in Fig. 11. Certainly, the
been observed. With 2D numerical simulations, some authors main features of the flow are correctly captured, although there
(e.g. [42,43]) have concluded that it is possible to obtain a steady are some slight differences, especially in X1/h (the normalised pri-
solution up to Re = 800, while other authors (e.g. [44]), have gone mary reattachment length) with increasing values of Re.
further and stated that the flow remains stable up to Re = 1500. Re- In Table 1, the predicted X1/h values at Re = 800 from different
cently, using a computational domain in which the outflow bound- studies have been listed along with the corresponding inlet and
ary was located far downstream (i.e. at 300 step heights), an inlet outlet boundary conditions as well as details of the meshes used.
boundary located in a channel some way upstream (i.e. at 20 step These reported X1/h values range from 11.48 [46] to 12.18 [42]

Re = 200

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

Re = 400

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

Re = 800
0.5

0
y

-0.5
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
x

Fig. 11. Contours of the axial velocity Vx profile for three different Re values.
S.-C. Xue, G.W. Barton / Comput. Methods Appl. Mech. Engrg. 253 (2013) 146159 157

Table 1
Numerically predicted X1/h values for Re = 800.

Authors Inlet channel Downstream channel Mesh used X1/h


a
Gartling [42] None 60h 400 * 20 12.2
Darbandi and Schneide [50] None 60h 320 * 80a 12.1
Kim and Moin [49] None 30h 101 * 101 11.9
Erturk [45] 20h 300h 4251 * 101a 11.8
Zang et al. [47] None 30h 130 * 66 11.8
Rouizi et al. [48] 4h 30h 144247 nodes 11.8
Rogers and Kwak [46] 2h 30h 100 * 53 11.5
Present study None 30h 151 * 61 11.7
a
Cases where the mesh is not uniformly distributed along the channel.

1.2 edge-based structure for the 2D case), a surface vector decomposi-


tion along a virtual line passing through the midpoint of the face
together with a deferred-correction approach, any cross-diffusion
1 arising from skewness of a grid composed of arbitrary shapes is
well accounted for. As a result, restrictions associated with grid
shape are removed, and a hybrid arrangement of arbitrary cell
0.8
shapes forming extremely skewed grids can be accommodated in
the computational domain discretisation.
0.6 The formulation is extended to problems involving fluid flow.
With a collocated arrangement of variables, discretisation of the
Vx

momentum equations is component-independent. To avoid the


0.4
checker-board pressure problem, a novel pseudo-flux interpola-
tion method that ensures numerical solutions are both
0.2 under-relaxation factor and time-step (for transient calculations)
x/h = 14 independent is proposed for non-orthogonal unstructured meshes.
With the proposed formulation, both coupled and decoupled
0 x/h = 30 solution algorithms can be employed for determination of the
kinematic and dynamic fields. In this work, a decoupled solution
-0.2 procedure based on the SIMPLER algorithm is developed.
-0.5 -0.25 0 0.25 0.5 The key features of the proposed method (including cell-shape
y and under-relaxation factor independence) were numerically
investigated by solving the NavierStokes equations that define
Fig. 12. Axial velocity (Vx) profile across the channel at x/h = 14 and 30 for Re = 800
two benchmarking problems from the literature. Our results show
(lines) compared with the mesh independent predictions (symbols) given in [42].
excellent agreement with published results but using far coarser
meshes, while our proposed method handles distorted meshes
exceedingly well as demonstrated by the smoothness of the solu-
which bracket the value obtained using our methodology. Obvi-
tion at locations where there are dramatic changes in cell shape.
ously, any difference cannot be attributed to different outflow loca-
tions and meshes used because mesh refinement analysis has been
Acknowledgements
done in all of those simulations. As this laminar flow is convection-
dominated, the observed differences could be related to the
The authors wish to thank the Australia Research Council for
schemes used in discretising the convection flux in the momentum
financially supporting this work.
equations. For example, with a skewed hybrid upstream finite dif-
ferencing scheme for this discretisation, Heckman et al. [51] pre-
dicted X1/h = 13.0 for an expansion ratio 2 at Re = 458 with a Appendix A. Spatial interpolation and gradient reconstruction
relatively coarse mesh (42 * 42), but X1/h < 11 if a simple upwind-
ing scheme was used. Indeed, if we use this simple upwinding In the finite volume formulation, for any scalar variable, a piece-
scheme for convection flux discretisation in our approach, then wise constant distribution is assumed over each control volume.
with the same mesh at Re = 800, we obtain the result that X1/ With unstructured (and possibly highly distorted) grids, for a posi-
h = 10.5, while at Re = 400 the second recirculation zone essentially tion-dependent quantity W (which can be a dependent variable or
disappears. a coefficient that is not necessarily continuous, or continuous but
Despite slight differences in the reattachment length, however, with an unknown gradient vector), the value at the midpoint of
calculated velocity profiles are essentially indistinguishable, with the interface shared by two adjacent cells P and NB can be approx-
Fig. 12 comparing our predicted x-component velocity profiles imated using a volume-weighted average of the values at the two
across the channel at two axial positions with mesh independent cell centroids, that is:
predictions reported in an earlier study [42]. Wf hNB WP hp WNB A:1

8. Conclusions where

XNB
A cell-centred finite volume formulation, based on non-orthog- hNB ; hP 1 & hNB A:2
XP XNB
onal unstructured meshes, has been proposed for the general
transport equation. Special attention is paid to the discretisation are the volume fractions for cells NB and P, respectively. This
of the diffusive flux. With a face-based data structure (or an approximation only gives first-order accuracy.
158 S.-C. Xue, G.W. Barton / Comput. Methods Appl. Mech. Engrg. 253 (2013) 146159

For a continuous variable W with a known gradient vector, [7] T.J. Barth, D.C. Jespersen, The design and application of upwind schemes on
unstructured meshes, in: 27th AIAA Aerospace Science Meeting, Reno, NV,
however, to ensure continuity, its value at a position r inside a cell
AIAA paper 89-0366, 1989.
is evaluated using linear interpolation with respect to the value at [8] Y. Zhao, B. Zhang, A high-order characteristics upwind FV method for
the cell centroid. For cell P, for example, Incompressible flow and heat transfer simulation on unstructured grids,
Comput. Methods Appl. Mech. Engrg. 190 (2000) 733756.
Wr WP rWP $ r & rP A:3 [9] H. Choi, D. Lee, J. Maeng, A node-centered pressure-based method for all speed
flows on unstructured grids, Numer. Heat Transfer 44 (2003) 165185.
where rP is the position vector for cell centroid P, while rWP is the [10] J. Ferziger, M. Peric, Computational Methods for Fluid Dynamics, third ed.,
gradient vector at this point. The value at the midpoint of the inter- Springer, Berlin, 2001.
[11] L. Davidson, A pressure correction method for unstructured meshes with
face shared by two adjacent cells P and NB is then reconstructed arbitrary control volumes, Int. J. Numer. Methods Fluids 22 (1996) 265281.
using a second-order symmetric formula: [12] G. Manzini, A. Russo, A finite volume method for advectiondiffusion problems
in convection dominated regimes, Comput. Methods Appl. Mech. Engrg. 197
WP rWP $ rf & rP ( WNB rWNB $ rf & rNB ( (2008) 12421261.
Wf A:4
2 [13] R. Rossi, Direct numerical simulation of scalar transport using unstructured
finite-volume schemes, J. Comput. Phys. 228 (2009) 16391657.
This gradient vector rWP is extensively used for interpolations, [14] S.-C. Xue, N. Phan-Thien, R.I. Tanner, Numerical study of secondary flows of
and how it is estimated can have a significant impact on solution viscoelastic in straight pipes by an implicit finite volume method, J. Non-
Newtonian Fluid Mech. 59 (1995) 191213.
accuracy. With unstructured grids, depending on mesh structure, [15] C.M. Rhie, W.L. Chow, A numerical study of the turbulent flow past an isolated
either the Green-Gauss theorem [7,52] or a least-squares approach airfoil with trailing edge separation, AIAA J. 21 (1983) 15251535.
[53] can be used to estimate rWP . For an uniform mesh structure, [16] S. Majumdar, Role of underrelaxation in momentum interpolation for
calculation of flow with nonstaggered grids, Numer. Heat Transfer 13 (1988)
the least-squares method is more accurate being exact for linear 125132.
profiles. However, for a non-uniform mesh structure in which high [17] T.F. Miller, F.W. Schmidt, Use of a pressure-weighted interpolation method for
aspect ratios are used in some cells, such as those along a boundary the solution of incompressible NavierStokes equations on a non-staggered
grid system, Numer. Heat Transfer 14 (1988) 213233.
layer, application of the Green-Gauss theorem is preferred, so that:
[18] J. Papageorgakopoulos, G. Arampatzis, D. Assimacopoulos, N.C. Markatos,
nf Enhancement of the momentum interpolation method on non-staggered grids,
1 X Int. J. Numer. Methods Fluids 33 (2000) 122.
rWP Wf A A:5
[19] F.S. Lien, M.A. Leschziner, A general non-orthogonal collocated finite volume
XP f
algorithm for turbulent flow at all speeds incorporating second-moment
turbulence-transport closure, Part 1: Computational implementation, Comput.
where the summation is over all faces nf bounding cell P, and Wf (i.e. Methods Appl. Mech. Engrg. 114 (1994) 123148.
the value at the midpoint of a face) is approximated firstly using the [20] S.K. Choi, Note on the use of momentum interpolation method for unsteady
simple interpolation Eq. (A.1). With an approximation to rWP ob- flows, Numer. Heat Transfer A 36 (1999) 545550.
[21] B. Yu, Y. Kawaguchi, W.Q. Tao, H. Ozoe, Checkerboard pressure predictions due
tained, Wf is then reconstructed using a second-order symmetric to the underrelaxation factor and time step size for a nonstaggered grid with
formula (A.4). The final value of the gradient vector at point P is then momentum interpolation method, Numer. Heat Transfer B 41 (2002) 8594.
obtained using Eq. (A.5) but now with the reconstructed Wf . Note [22] G.D. Raithby, Skew upstream differencing schemes for problems involving
fluid flow, Comput. Methods Appl. Mech. Engrg. 9 (1976) 153164.
that the reconstructed Wf may not be bounded in a region with a
[23] B.P. Leonard, A stable and accurate convective modelling procedure based on
steep gradient, a situation likely to slow overall convergence. To quadratic upstream interpolation, Comput. Methods Appl. Mech. Engrg. 19
avoid this, any reconstructed value should be in the following range: (1979) 5998.
[24] J.P. Borris, D.L. Brook, Flux-corrected transport I: SHASTA, a fluid transport
minfWNB g 6 Wf 6 maxfWNB g A:6 algorithm that works, J. Comput. Phys. 11 (1973) 3869.
[25] B. Van Leer, Towards the ultimate conservative difference scheme V. A second-
and thus we take the reconstructed value to be as follows: order sequel to Godunovs method, J. Comput. Phys. 32 (1979) 101136.
[26] A. Harten, High resolution schemes for hyperbolic conservation laws, J.
Wf maxminmaxfWNB g; Wf ; minfWNB g A:7 Comput. Phys. 49 (1983) 357393.
[27] M.S. Darwish, F. Moukalled, TVD schemes for unstructured grids, Int. J. Heat
where fWNB g is a set consisting of the values at the cell centroid Mass Transfer 46 (2003) 599611.
being considered and its immediate neighbours. [28] Y.-Y. Tsui, T.-C. Wu, Use of characteristic-based flux limiters in a pressure-
based unstructured-grid algorithm incorporating high-resolution schemes,
The value of a dependent variable U at a cell vertex (or node) is Numer. Heat Transfer B 55 (2009) 1439.
evaluated using a distance-weighted average of the values at the [29] S.-C. Xue, N. Phan-Thien, R.I. Tanner, Upwinding with deferred correction
centroids of the cells associated with that vertex: (UPDC): an effective implementation of higher-order convection schemes for
implicit finite volume, J. Non-Newtonian Fluid Mech. 108 (2002) 124.
P
NC &1 [30] P.K. Khosla, S.G. Rubin, A diagonally dominant second-order accurate explicit
k Uk dk
Und P NC &1
A:8 scheme, Comput. Fluids 2 (1974) 207209.
k dk [31] P. Traore, Y.M. Ahopo, C. Louste, A robust and efficient finite volume scheme
for the discretization of diffusive flux on extremely skewed meshes in complex
where nd is the node identifier, Uk is the value at the kth cell cen- geometries, J. Comput. Phys. 228 (2009) 51485159.
troid, and dk is the distance from this centroid to the vertex where [32] P.J. Zeart, G.D. Raithby, An integrated space-time finite volume method for
moving-boundary problems, Numer. Heat Transfer B 34 (1998) 257270.
the value is to be estimated. The summation is over all cells (NC of [33] H.A. Van der Vorst, Bi-CGSTAB: a fast and smoothly converging variant of Bi-
them) associated with the vertex in question. CG for the solution of nonsymmetric linear systems, SIAM J. Sci. Stat. Comput.
13 (1992) 631644.
[34] R.I. Issa, Solution of implicitly discretised fluid flow equations by operator-
References splitting, J. Comput. Phys. 62 (1985) 4065.
[35] S. Acharya, B.R. Baliga, K. Karki, J.Y. Murthy, C. Prakash, S.P. Vanka, Pressure-
[1] O.C. Zienkiewicz, K. Morgan, Finite Element and Approximation, Wiley, 1983. based finite-volume methods in computational fluid dynamics, ASME J. Heat
[2] I. Babuska, A.K. Aziz, Angle condition in finite element method, SIAM J. Numer. Transfer 129 (2007) 407424.
Anal. 13 (1976) 214226. [36] M.S. Darwish, I. Sraj, F. Moukalled, A coupled finite volume solver for the
[3] S.V. Patankar, Numerical Heat Transfer and Fluid Flow, second ed., solution of incompressible flows on unstructured grids, J. Comput. Phys. 228
Hemisphere, NY, 1996. (2010) 180201.
[4] B.R. Baliga, S.V. Patankar, A control-volume finite element method for two- [37] Z.J. Chen, A.J. Przekwas, A coupled pressure-based computational method for
dimensional incompressible fluid flow and heat transfer, Numer. Heat Transfer incompressible/compressible flows, J. Comput. Phys. 229 (2010) 91509165.
6 (1983) 245261. [38] U. Ghia, K.N. Ghia, C.T. Shin, High-Re solutions for incompressible flow using
[5] D. McBride, M. Cross, N. Croft, A coupled finite volume method for the solution the NavierStokes equations and a multigrid method, J. Comput Phys. 48
of flow processes on complex geometries, Int. J. Numer. Methods Fluids 53 (1982) 387411.
(2007) 81104. [39] D. McBride, N. Croft, M. Cross, Finite volume method for the solution of flow on
[6] D. Hilhorst, M. Vohralik, A posteriori error estimates for combined finite distorted meshes, Int. J. Numer. Methods Heat Fluid Flow 17 (2007) 213239.
volumefinite element discretizations of reactive transport equations on [40] S.L. Lee, R.Y. Tzong, Artificial pressure for pressure-linked equation, Int. J. Heat
nonmatching grids, Comput. Methods Appl. Mech. Engrg. 200 (2011) 597613. Mass Transfer 35 (1992) 27052716.
S.-C. Xue, G.W. Barton / Comput. Methods Appl. Mech. Engrg. 253 (2013) 146159 159

[41] B.F. Armally, F. Durst, J.C.F. Pereira, B. Schonung, Experimental and theoretical [47] Y. Zang, R.L. Street, J.R. Koseff, A non-staggered grid, fractional step method for
investigation of backward-facing step flow, J. Fluid Mech. 127 (1983) 473496. time-dependent incompressible NavierStokes equations in curvilinear
[42] D.K. Gartling, A test problem for outflow boundary conditions flow coordinates, J. Comput. Phys. 114 (1994) 1833.
over a backward-facing step, Int. J. Numer. Methods Fluids 11 (1990) [48] Y. Rouizi, Y. Favennec, J. Ventura, D. Petit, Numerical model reduction of 2-D
953967. steady incompressible laminar flows. Application on the flow over a
[43] P.M. Gresho, D.K. Gartling, J.R. Torczynski, K.A. Cliffe, K.H. Winters, T.J. Garrat, backward-facing step, J. Comput. Phys. 228 (2009) 22392255.
A. Spence, J.W. Goodrich, Is the steady viscous incompressible two [49] J. Kim, P. Moin, Application of a fractional-step method to incompressible
dimensional flow over a backward facing step at Re = 800 stable?, Int J. NavierStokes equations, J. Comput. Phys. 59 (1985) 308323.
Numer. Methods Fluids 17 (1993) 501541. [50] M. Darbandi, G.E. Schneider, Application of an all-speed flow algorithm to heat
[44] D. Barkley, M.G.M. Gomes, R.D. Henderson, Three-dimensional instability in transfer problems, Numer. Heat Transfer A 35 (2010) 695715.
flow over a backward-facing step, J. Fluid Mech. 473 (2002) 167190. [51] L.P. Hackman, G.D. Raithby, A.B. Strong, Numerical predictions of flows over
[45] E. Erturk, Numerical solutions of 2-D steady incompressible flow over a backward facing steps, Int. J. Numer. Methods Fluids 4 (1984) 711724.
backward-facing step, Part I: high Reynolds number solutions, Comput. Fluids [52] S.R. Mathur, J.Y. Murthy, A pressure-based method for unstructured meshes,
37 (6) (2008) 633655. Numer. Heat Transfer B 31 (1997) 195215.
[46] S.E. Rogers, D. Kwak, An upwind differencing scheme for the incompressible [53] T.J. Barth, A 3-D Upwind Euler Solver for unstructured meshes, in: AIAA 10th
NavierStokes equations, Appl. Numer. Math. 8 (1991) 4364. Computational Fluid Dynamics Conf., Honolulu, HI, 1991 pp. 228238.

View publication stats

You might also like