You are on page 1of 115

Helsinki University of Technology Laboratory of Applied Thermodynamics

A Study of the Pressure Correction Approach in the Colocated Grid Arrangement Ari Miettinen1 Helsinki University of Technology, Espoo, Finland Report No. 110

Approved by Timo Siikonen

1997 Otaniemi ISBN 951223887X ISSN 12378372

1 Research

Scientist, Laboratory of Applied Thermodynamics

Abstract
A pressure correction approach in a colocated grid arrangement is studied. The SIMPLE algorithm has been implemented into a two-dimensional Navier-Stokes solver for incompressible ows. A multigrid Poisson solver for the pressure correction equation is utilized. It has been tested alone and as a part of the NavierStokes solver. In the former case, the improvement in the convergence rate is remarkable. As a part of a carefully performed Navier-Stokes computation, the multigrid Poisson solver is able to decrease the total CPU time to one third. Cell-face velocity formulas for the continuity equation are also studied. The solution method is tested using several ow cases including a lid-driven cavity and a buoyance-driven ow. The Rhie & Chow interpolation method and its simpli ed and limited versions are compared with each other. The traditional Rhie & Chow interpolation method had problems if the ow eld contained a high pressure gradient, which can be overcome by using the two latter versions. For buoyancy-driven ows, some authors have proposed a body force term in the cellface velocity formula, but the buoyancy-driven test case Ra ! 109 indicated that this is not necessary.

Contents
Nomenclature 1 Introduction 2 Calculation of Incompressible Flows 5 9 16

2.1 Governing Equations . . . . . . . . . . . . . . . . . . . . . . . . . 16 2.2 Numerical Calculation of the Equation System . . . . . . . . . . . 17 2.3 Convection-Di usion Solver . . . . . . . . . . . . . . . . . . . . . 17 2.3.1 Discretization . . . . . . . . . . . . . . . . . . . . . . . . . 19 2.3.2 Extension to Non-orthogonal Grids . . . . . . . . . . . . . 22 2.3.3 Time Integration . . . . . . . . . . . . . . . . . . . . . . . 26 2.3.4 Boundary Conditions . . . . . . . . . . . . . . . . . . . . . 29 2.3.5 Flow Chart of the Convection-Di usion Solver . . . . . . . 31

3 Treatment of Pressure in Incompressible Navier-Stokes Equations 32


3.1 Velocity-Pressure Coupling Using Pressure Correction Approach in Colocated Grid Arrangement . . . . . . . . . . . . . . . . . . . 33 3.2 Calculation of Cell-Face Velocity in Colocated Grid . . . . . . . . 38 3.2.1 A Short Discussion about the Cell-Face Velocity Interpolation 42 3.3 Solution Sequence . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

4 Pressure Correction Equation

46

4.1 Fine-Grid Problem . . . . . . . . . . . . . . . . . . . . . . . . . . 47 4.2 Multigrid Operators . . . . . . . . . . . . . . . . . . . . . . . . . 48 2

4.2.1 Grid Transfer . . . . . . . . . . . . . . . . . . . . . . . . . 48 4.2.2 Coarse-Grid Problem and Restriction Operator . . . . . . 49 4.2.3 Prolongation Operator . . . . . . . . . . . . . . . . . . . . 50 4.2.4 Accuracy of the Operators . . . . . . . . . . . . . . . . . . 54 4.3 Multigrid Technique . . . . . . . . . . . . . . . . . . . . . . . . . 55 4.3.1 The Solution Sequence of the FAS Method . . . . . . . . . 55 4.3.2 The Solution Sequence of the MG Method . . . . . . . . . 57

5 Test Calculations

60

5.1 Poisson Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . 60 5.2 Lid-driven Cavity Flow . . . . . . . . . . . . . . . . . . . . . . . . 69 5.2.1 Re = 400 . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 5.2.2 Re = 10 000 . . . . . . . . . . . . . . . . . . . . . . . . . . 79 5.3 Two-dimensional Inviscid Channel Flow . . . . . . . . . . . . . . 84 5.4 Two-dimensional Inviscid Flow in Diverging Channel . . . . . . . 96 5.5 Buoyancy-Driven Cavity Flow . . . . . . . . . . . . . . . . . . . . 102

6 Conclusions

107

Nomenclature
Symbols

a b C cp CFL DAMPLI F ~ g ~ i ~ j k Lx Lxx Ly Lyy MAX ;  MIN ;  m m _ m _ out mP mR ~ n0 Nu p p p0 P : p0 ! p0 ~ r R : p0 ! p0 Ra ReL S T T1 t ~ t0

cell surface area coe cient matrix of pressure correction vector coe cient in momentum equation body force term coe cient of damping term of cell-face velocity speci c heat at constant pressure Courant number local limiter of relative damping term ux gravitational acceleration vector Cartesian unit vector in -direction coe cient of convection term of damping of cell-face velocity convection operator in -direction di usion operator in -direction convection operator in -direction di usion operator in -direction maximum operator minimum operator order of partial di erential equation divided by two mass ux mass ux out of cell accuracy of prolongation operator accuracy of restriction operator exterior unit normal vector Nusselt number pressure guessed pressure or pressure from previous iteration cycle pressure correction prolongation operator location vector restriction operator Rayleigh number Reynolds number, ref source term temperature ambient temperature time unit tangential vector

x Cartesian unit vector in y -direction i i j

U L=

u u u0 U v v v0 ~ v V vn vt x X y Y
p v

"     


velocity component in x-direction u-velocity from momentum equation, where p has been used velocity correction dimensionless u-velocity velocity component in y -direction v -velocity from momentum equation, where p has been used velocity correction velocity vector, ~ v = u~ i + v~ j volume of cell or dimensionless V-velocity normal velocity tangential velocity Cartesian coordinate axis dimensionless coordinate, = x=L Cartesian coordinate axis dimensionless coordinate, = y=L a general di usion coe cient or thermal di usivity relaxation factor of pressure relaxation factor of velocity volumetric thermal expansion coe cient surface boundary layer thickness angle between vector ~ g and y -axis convergence error convergence criteria for mass balance residual parameter in MUSCL scheme thermal conductivity dynamic viscosity kinematic viscosity normalized second di erence a general variable  1;1 ; :::; i;j ; :::; imax;jmax  density volume

Superscripts
c d expl: impl: n num
convection di usion explicit implicit time step numerical auxiliar quantity guessed or previous value, or value based on p correction 5

# 

Subscripts
0 initial value cold characteristic value eastern cell denotes to location at cell or domain face hot spatial indices of ner grid

c char e f ace h i; j I; J in max min n ne nw nb out p ref s se sw w W

spatial indices of coarser grid inlet maximum minimum northern cell north-eastern cell north-western cell neighbouring cells outlet present cell reference southern cell south-eastern cell south-western cell western cell wall

Abbreviations
ACM CD CFD CPU DCA FA FAS FMG FOU GCA GS HYBRID LGS MG MUSCL MWIM arti cial compressibility approach convection-di computational usion uid dynamics

central processing unit discretization coarse-grid approximation nite analytic full approximation storage full multigrid rst-order upwind Galerkin coarse-grid approximation Gauss-Seidel Spalding's convection scheme line-Gauss-Seidel multigrid monotone upstream-centred schemes for conservation laws momentum weighted interpolation method

PCM PISO PPEM PWIM QUICK SIMP1 SIMP2 SIMPLE SIMPLEC SIMPLER SL SOR

pressure correction method a predictor-corrector PCM Poisson pressure equation method pressure weighted interpolation method quadratic-upstream interpolation for convective kinematics rst simpli cation of damping term second simpli cation of damping term semi-implicit method for pressure-linked equations SIMPLE consistent SIMPLE revised single-level successive overrelaxation

Introduction

A wide range of physical phenomena depends remarkably on uid ow. Therefore, there has been much interest in solving ow elds. But the complexity of the complete ow equations, proposed by Navier in 1822 and Stokes in 1845 1 , became the problem. Closed or serial-form solutions, always for the simpli ed equations, had to be devised case by case. Finally, computers o ered a generic approach to solving the ow elds, but even computers could not solve all the problems. The mystery of turbulence seems to remain to be solved during the third millennium. Since early days, the compressible ow phenomenena have been the most potential area of the CFD Computational Fluid Dynamics simulations. The applications, especially in aerodynamics and turbine engine design, have caused a great deal of attention to be focused on the development of methods for the numerical solution of the compressible ow equations. And little by little, the CFD approach has spread to other elds of engineering. The solution methods developed for compressible ows cannot be adopted for incompressible ows. From the mathematical point of view, the incompressible and compressible ow equations are not similar 2 . The major di erence between them is their mathematical character. The compressible ow equations are hyberbolic, which means that they have real characteristics on which the signals travel at nite propagation speeds. By contrast, incompressible equations have a mixed parabolic-elliptic character. If methods for compressible ow are to be used to compute incompressible ow, the character of the equations will need to be modi ed. Or alternatively, methods suitable for equations having a mixed parabolic-elliptic character need to used. The di erence in characters can be traced by the lack of a time derivative term in the incompressible continuity equation. In the arti cial compressibility approach ACM, rst proposed by Chorin 3 in 1967, a time derivative of pressure is added into the continuity equation. At convergence, the time derivative is zero and the solution satis es the incompressible continuity equation. The method is suitable for steady-state ows. A number of modi ed versions has been presented. Also, modi cations suitable for time-dependent calculations have been proposed. In another approach, a solution method suitable for equations of a mixed parabolicelliptic character has been constructed. At least streamfunction-vorticity, fractional step, Poisson pressure equation PPEM and pressure correction methods PCM can be referred. The main di erence between the methods is the treatment of pressure. The streamfunction-vorticity method has been used mainly to solve two-dimensional incompressible ows, e.g. 4 and 5 . The principal idea of the method is 8

to take the curl, variables

pressure disappears, because

u, v

and

r  , of the momentum equation of the vector form, when the r  rp 0. The original equations for unknown
p
are replaced by equations for

and

from which the new

unknowns are easily solved. The streamfunction is used to calculate the velocities

u and v .

The pressure can be calculated afterwards by the aid of streamfunction. ows is di cult, and the and

The method is e ective in two-dimensional cases, but its bene ts are lost in threedimensional space. Its extension to three-dimensional x , the number of unknown variables is now increased from four u,

y , z , !x , !y

and

!z ,

v, w

p

to six

i.e. the vorticity and the streamfunction become

three-component vectors in three dimensions 2 . Regardless of the dimensions of ow, the streamfunction-vorticity method has also problems with boundary uid properties and compressibility 2 . In recent years, the conditions, variable

method has become less popular. The rst multidimensional pressure correction method was developed by Harlow ow and used

and Welch 6 in 1965. They were interested in time-dependent an ine ective way solving the steady-state of the pressure

explicit time integration. Soon the explicit time integration proved itself to be ows. An implicit pressure correction The correction method was proposed by Patankar and Spalding 7 in 1972.

eld using the continuity equation was the principal idea in both

proposals. The method proposed by Patankar and Spalding became very popular

Pressure-Linked Equations.

and is better known as the SIMPLE algorithm 8  emi- mplicit

Method

for

Many improvements to increase the convergence

rate have been presented, and the version proposed by Doormaal and Raithby 9 is nowadays popular like SIMPLE. Doormaal and Raithby's version is better known as the SIMPLEC algorithm where C refers to the word consistent. The pressure correction approach is the most popular way of solving incompressible ows. The SIMPLE algorithm has been used by Melaaen 10 , 11 , Kobayashi and Pereira 12 , Peric and Kessler 13 , Peric 14 , Lapworth 15 , Jessee and Fiveland 16 , Choi et al. 17 , Jeng and Liou 18 , Rhie and Chow 19 , and Mathur and Murthy 20 . SIMPLEC algorithm has been adopted by Thiart 21 , 22 , Tamamidis et al. 23 , 24 , Rodi et al. 25 , Johansson and Davidson 26 , and Miller and Schmidt 27 . Most of them have used both under-relaxed velocities and pressure, which should have been avoided when using SIMPLEC. Furthermore, SIMPLE-like algorithms can be found in articles by Date 28 , 29 , and Lien and Leschziner 30 . The performance of the PISO, SIMPLER and SIMPLEC pressure correction algorithms for the treatment of the pressure-velocity coupling in steady- ow problems was examined by Jang et al. 31 . They found that, for problems in which the momentum equation was not coupled to a scalar variable, the PISO algorithm was superior, but when the scalar variable, e.g. temperature in buoyancy-driven ows, was strongly coupled to the momentum equation, SIMPLER and SIMPLEC exhibited better behaviour, and reasonable solutions with PISO algorithm were obtained only for small time-steps. They did not observe clear superiority of 9

SIMPLER over SIMPLEC or vice versa. Their study was executed in staggered grid. Tamamidis et al. 24 compared SIMPLEC pressure correction and arti cial compressibility methods in a three-dimensional colocated grid arrangement. The benchmark evaluation of the two methods was performed through calculations of laminar ows in 90o -curved ducts of square and circular cross-sections. In general, the control-volume-based PCM compared slightly more favorably with measurements and other well-resolved computations than the nite-di erencebased ACM. However, both methods produced grid-independent results at the same grid resolution. The ACM needed fewer iterations because it solved the continuity and momentum equations in a fully coupled manner, but it required 16 times more memory. The ACM method requires the use of only one adjustable numerical parameter to control convergence. However, the method's stability and convergence are sensitive to this parameter, and its selection requires an experienced user. The PCM typically requires two under-relaxation factors, which can be selected quite easily, thus making the method more user-friendly. Williams 32 proposed a method in which arti cial compressibility and pressure correction approaches were combined. He combined the continuity equations of both methods leading to the diagonally dominant continuity equation for pressure like quantity. The test computations of lid-driven cavity ow Re = 100 and 400 showed, that the CPU time spent in pressure solver could be diminished even to one-tenth part compared with the SIMPLE algorithm. Thus, Williams was able to decrease the total CPU-time about 45. The Poisson solver used by Williams was the LU-type equation solver. The bene ts of Williams' solver should be gained by using an e ective solver for the pressure-like variable. Abdallah 33 , 34 and Sotiropoulos et al. 35 , 36 , 37 used PPEM for incompressible ows. The idea was to substitute the momentum equations into the continuity equation, from which the new pressure was solved. The algorithm was similar to the method proposed by Harlow and Welsh 6 . Aksoy and Chen 38 compared the preformance of SIMPLEC and PPEM in the lid-driven cavity ow of Re = 100 and 1 000. They concluded, that PPEM converged  1=3 faster than SIMPLEC, but the continuity equation was poorly satis ed. Sotiropoulos and Abdallah 35 developed a method in which the damping term of the continuity equation was separated from the other terms so that it could be controlled case by case. Thus, the continuity constraint became better satis ed. The fractional step concept has similarities with PPEM. It is more a generic approach than a particular method 2 . An application of a fractional step method to incompressible Navier-Stokes equations has been presented by Kim and Moin 39 . The method can be packed into the next steps. First, the momentum equations are solved without the pressure. During the second step, the velocities of the previous step are used to calculate the pressure from the continuity equation. During the third step, the new pressure eld is used to correct the velocities of 10

the

rst step. This three-step description presents the main idea of the method.

In recent years, the method has become more popular. Pressure or pressure correction in PPEM, fractional step method or pressure correction method is solved from a Poisson equation. A noticeable portion of computational time can be spent on solving these equations, e.g. 32 . Therefore, an e ective Poisson solver is needed. Some single-grid methods, e.g. conjugate gradient method and successive over-relaxation SOR technique 40 , are quite powerful, but the multigrid technique 41 seems to be the most e ective, e.g. 2 , and recommended, 25 . Besides the pressure correction approach, Harlow and Welsh 6 introduced the staggered grid arrangement. The staggering was also utilized in the original versions of SIMPLE 7 , SIMPLER 8 and SIMPLEC 9 . It was explained to strengthen the velocity-pressure coupling to prevent the oscillation of the solution, but its disadvantage soon became clear. The treatment of the complex geometries became very clumsy. This was the main reason for studies to develop an oscillation-free method for the colocated grid arrangement. The list of merits of colocated grid can easily be continued, 13 : the convection contribution to the coe cients in the discretized equations is the same for all variables, cartesian velocity components can be used in conjunction with non-orthogonal coordinates, there are fewer constraints on the numerical grid because there is no need to evaluate the so-called curvature terms, and multigrid techniques are much easier applied to the colocated arrangement. Rhie and Chow 19 proposed a method of preventing the pressure oscillations in the colocated grid arrangement. The principal idea was to add a damping term to the central di erence stencil of the cell-face velocity. In practice, the damping term was a di erence of pressure gradient stencils at the cell-face, and thus the formula began to be called the pressure-weighted interpolation method PWIM along with the Rhie & Chow interpolation. The damping term can be derived from the momentum equation of the "staggered face cell", which has been directly used when developing the so-called momentumweighted interpolation method MWIM. In fact, MWIM can be treated as a half-done Rhie & Chow interpolation scheme. And nally, both methods utilize the staggering idea indirectly, 13 . PWIM has been used by Melaaen 10 , 11 , Acharya and Moukalled 42 , Lapworth 15 , Wu and Rath 43 , Gu 44 , Tamamidis et al. 23 , 24 , Mathur and Murthy 20 , Lien and Leschiziner 30 , and Miller and Schimdt 27 . MWIM has been utilized by Thiart 21 , 22 , Kobayashi and Pereira 12 , Parameswaran et al. 45 , Peric and Kessler 13 , Majumdar 46 , Jessee and Fiveland 16 , Choi et al. 17 , Zhang et al. 47 , Peric 14 and Rodi et al. 25 . Aksoy and Chen 38 compared PWIM and MWIM in conjunction with the pressure correction approach. They concluded that the di erences between the per11

formances of the methods were minor in lid-driven cavity

ow tests of

Re = 100

and 1 000. Based on the same origin of the formulas, the conclusion could have been expected.

Peric et al.

13

compared staggered and colocated grid arrangements in twoows. Lid-driven square cavity

dimensional laminar step

Re = 100, and circular pipe with a sudden contraction were chosen.

Re = 100, backward-facing
The

number of iterations and the total computing times were almost identical for the rst and the second case. In the third case, the colocated version required fewer iterations

29. The typical di erences between the solutions obtained with the

staggered and colocated versions were about 0 01 of the appropriate reference values. Di erences of the order of 0 1 occurred only in rather small regions. 13 recommended the colocated grid

Also, due to its other merits, Peric et al. arrangement.

Some modi cations to the Rhie & Chow interpolation have been proposed. The earliest proposal found during this study was presented by Han 48 . He substi-

tuted the dissipation term created by PWIM directly into the continuity equation. He also multiplied the dissipation term by a factor to be xed case by case. The

idea was to control the amount of damping to keep simultaneously the convergence rate acceptable and the solution oscillation-free.

Johansson and Davidson 26 proposed a Rhie & Chow interpolation scheme similar to Han's 48 idea. They simpli ed the formula and multiplied the di erence of pressure gradient stencils by a factor to be used to control the amount of dissipation in the continuity equation. By using the new formula, Johansson ow of

and Davidson were able to calculate buoyancy-driven cavity which was impossible with the original Rhie & Chow version.

Ra = 107,

They were also

able to compute some turbulent and thermal ventilation applications, which were previously impossible in the colocated grid arrangement.

Gu

44

had also recognized the convergence problems of the colocated grid arows. He introduced a second-order body force

rangement in buoyancy-driven

term into the Rhie & Chow interpolation stencil. A successful test computation was run using Harwell's FLOW3D computer program with a case

Ra = 1010, i.e.

it was turbulent. Using the original PWIM the case did not converge. Rahman et al. 49 also proposed body force correction to PWIM, but their correction rst-order accurate.

made the cell-face velocity stencil

Date 28 proposed methods for preventing pressure oscillations in the colocated grid arrangement. At rst, he introduceded a non-symmetric PWIM scheme,

which was no more second-order accurate. Next, he proposed that cell-face velocities could be calculated using central di erence stencils if the pressure gradients in momentum equations were calculated in a proper manner. the methods, Date calculated lid-driven cavity driven ow in a corner of ow of

Ra = 1 000. Date's latter idea was also tried by Russell


12

Re = 100 and buoyancy-

To demonstrate

and Abdallah 50 . They proposed that higher order e.g. 4th order central di erence approximations for the gradient operators of pressure could prevent oscillation and make solutions dilation-free. Lid-driven cavity ow of Re = 103 was calculated as a computational test. Unfortunately, Acharya and Moukalled 42 found that an oscillation-free solution could be obtained for lid-driven cavity ow of Re = 100 without damping terms. Thus there is no uniform knowledge of any modi ed pressure gradient stencil to prevent the oscillation of results in the colocated grid arrangement. Furthermore, Russell's proposal leads to a very complex pressure Poisson equation, especially in general coordinates. In the later article, Date 29 corrected the conclusions of the earlier paper. He pointed out that, either by evaluating the cell-face velocities by the momentuminterpolated principle or by evaluating an e ective pressure gradient in the nodal momentum equations, this might lead to spurious results when the true pressure variation considerably departed from linearity. What was required instead was a fresh derivation of the pressure correction equation appropriate for a nonstaggered grid. The mass had to be conserved and smoothing components were needed to prevent oscillation. In Date's proposal, the damping terms were directly included in the pressure correction equation. The details, especially pressure correction equation, deviated from ordinary practice, but the idea was earlier used by Han 48 . MWIM has also been used in conjunction with PPEM 35 , 36 and 37 . It is quite natural, because the momentum equations are substituted into the continuity equation, i.e. the momentum equations have to be interpolated at the cell-faces. Sotiropoulos and Abdallah 35 have found that the amount of dissipation cannot be controlled in MWIM. Therefore, they have developed the pressure Poisson equation into a form in which the dissipation term is explicitly present, which makes the control possible, as in Han's 48 and Johansson and Davidson's 26 pressure correction approaches. Reggio and Camarero 51 presented a strange colocated grid arrangement. It was an overlapping grid with forward and backward di erencing for mass and pressure gradients, respectively. The structure allowed the use of the same computational cell for both the continuity and momentum equations and storage of the pressure and velocity components at the same grid location. Even though they calculated a non-orthogonal case to demonstrate the grid arrangement, the system seems to be quite clumsy. The use of the Rhie & Chow interpolation has not been limited to incompressible ow simulations. Compressible ows have been calculated using it in conjunction with the pressure correction approach, e.g. Lien and Leschziner 30 and Rincon and Elder 52 . Also as a part of an ordinary compressible ow solver, MWIM has been used by Zhang et al. 47 . The concept of cell-face velocity has been adopted in conjunction with PWIM 13

and MWIM. The cell-face velocities are needed to calculate mass uxes through the cell-faces. In general coordinate systems, contravariant values are almost always used, except Choi et al. 17 , who has used covariant components. The cell-face values are always updated after the pressure correction equation. Most of the solvers seem to contain arrays for cell-face velocities, because the Rhie & Chow interpolated values are corrected directly without corrected cell values. Only Date 29 has demonstrably abandoned conserving the cell-face values into the arrays, and he always uses a central di erence stencil for cell-face velocities. In the commercial and industrial environment, geometric exibility ranks high in importance. Therefore, even in university codes a general coordinate system must be used. Otherwise, newly developed models and numerical techniques cannot be subjected to testing and validation in a challenging geometric environment within which an increasing amount of experimental data are being generated, e.g. 30 . In many articles the general coordinates are presented using the nite di erence approach with Jacobian. Only Peric 14 , Johansson and Davidson 26 , Jessee and Fiveland 16 , and Mathur and Murthy 20 have used the pure control volume approach with vector notations. But, cartesian velocity components are always used as variables in momentum equations. The di usion terms are always discretized using the central di erence stencil. The variety of convection models is wide. Approximately half of the writers have used the rst-order accurate methods, like rst-order upwind, HYBRID and powerlaw. Also some analytical methods, which in practice are rst-order accurate, have been used, e.g. exponential scheme by Spalding 53 and nite analytic FA method by Chen 54 . The most popular higher order scheme is QUICK, proposed by Leonard 55 . It has been used by Miller and Schmidt 27 , Tamamidis et al. 23 , 24 , Johansson and Davidson 26 , Rahman et al. 49 , and Zhang et al. 47 . The second-order upwind scheme has been adopted by Melaaen 10 , 11 , Sotiropoulos et al. 36 , 37 , and Mathur and Murthy 20 . Han 48 , Peric 14 and Williams 32 have applied the central di erence stencil for convection terms. Only Lien and Leschziner 30 and Jessee and Fiveland 16 have used the MUSCL scheme with ux limiters. The acronym stands for onotone pstreamcentred chemes for onservation aws, after the name of the rst code applying this method 56 . It contains an in nite number of upwind-weighted convection schemes including central di erence, QUICK and second-order upwind stencils.

In this study, the SIMPLE pressure correction approach has been chosen. For simplicity, the equations are solved in a segregated manner. An e ective multigrid solver is developed for pressure correction equation. Methods for calculating cell-face velocities are studied. Convection and di usion terms are modelled using MUSCL and the central di erence stencil, respectively. The variables are located in the cell-centred colocated grid. The general coordinates are described using the pure control volume approach with vector notations. Cartesian velocity components are used as variables in the momentum equations, and the contravariant components are used to calculate convection uxes through the cell-faces. 14

2 Calculation of Incompressible Flows


2.1 Governing Equations
The numerical computations of incompressible ows are based on the conservation of mass, momentum and internal energy, and they are usually presented as a set of partial di erential equations. In this study, two-dimensional equations will be utilized. In cartesian coordinates, the conservation of the mass, i.e. continuity equation, is as follows 2
@ @t @ @ + @x  u + @y  v = 0

1

~ where is the density and velocity vector is de ned ~ v =u i + v~ j . The components of the momentum equation in the x- and y-directions are in conservation form 2

@  u+  @t @x @

2 u

@ + @y 

uv

=

@ @u @ @u +   +    @x @x @x @y @y

@p

jg j

cos T

1 1 2

@  v +  @t @x @

uv

@ + @y  v2 =

@ @v @ @v +   +    @y @x @x @y @y

@p

jg j

sin T

1 1 3

where p is pressure,  is dynamic viscosity, gravitational acceleration vector is ~ g, is temperature, is volumetric thermal expansion coe cient, T1 is ambient temperature and is angle between gravitational acceleration vector ~ g and the y -axis. The uid has been considered as Newtonian. The dynamic viscosity has been assumed to be constant, and the continuity equation 1 has been used to manipulate the viscous terms. Assumption of a constant density has been used except the gravitational body force terms which has been modelled using Boussinesq's appoximation, see e.g. 57 . Equations 2 and 3 are often called Navier-Stokes equations. The temperature occurring in momentum equations is solved from the internal energy equation, which is as follows 2
T @  T +  @t @x @ uT @  + @y  vT @  @T @  @T  = @x  c @x  + @y  c @y 

4

where  and cp are thermal conductivity and speci c heat at constant pressure, respectively. Equations 1, 2, 3 and 4 form a closed system from which 15

the

ow properties

u, v , p

and

can be solved as a function of space and time. ows will be calculated. An extension to the

But, in this study only steady-state

general coordinates will be studied later.

2.2 Numerical Calculation of the Equation System


Simultaneous numerical calculation of Eqs. 1, 2, 3 and 4 is computa2, 3 and 4 is similar ow tionally complex. Therefore, the equations are solved one after another, i.e. in a segregeted manner. The basic structure of Eqs. to each other, containing an unsteady term, convection, di usion and possibly source terms, and they are often called convection-di usion equations. The properties quantity.

u, v and T

are solved from Eqs. 2, 3 and 4 respectively. There-

fore, the continuity equation 1 is to be modi ed for pressure or pressure-like

2.3 Convection-Di usion Solver


The rst stage is to derive a convection-di usion equation to nite form. In this , study the control volume approach is utilized. The process is to be studied by the aid of a general convection-di usion equation for quantity

@  @t
where

+

@  u @x

+

@  v @y

=

@
@ @
@ + +S @x @x @y @y S

5

is di usion coe

cient and the source term

could contain, for instance, 5 is integrated over the

pressure gradient and or body force, etc. Next, Eq. control volume

Vi;j .

The control volume

Vi;j

with its surrounding is represented

in Fig. 1 by dashed lines. After the rearrangements the integrated equation can be obtained in the form,

Vi;j

@

@t
+

i;j

A u

A v

@ i @y i;j +1=2

@ i @x i+1=2;j

A u

A v

@ i = Si;j @y i;j 1=2

@ i @x i 1=2;j

6

where

is a surface area of a face of the cell. Surface integrals has been ap2 . By ux as follows

proximated by using the second-order accurate midpoint rule, see e.g. de ning quantity

16

1=2;j

= A u

@ i @x

1=2;j

7

Eq. 6 can be simpli ed to form


V @  @t F

i;j

i;j

1=2;j

+ F +1 2
i

= ;j

i;j

1=2

+F

i;j +1=2

=S

i;j

8

Next, the integrated Eq. 6 is to be discretized.

i,j+1
(i,j+1)

vi,j+1/2

i1,j
(i1,j)

ui1/2,j

i,j
(i,j)

ui+1/2,j

i+1,j

(i+1,j)

vi,j1/2
y

i,j1
(i,j1)

Figure 1. Control volume i; j  and its surroundings in Cartesian


coordinates.

17

2.3.1 Discretization
The present solver is based on time integration, but the ows are assumed to be independent of time. Thus, a rst-order accurate approximation for the time +1 derivative will be accepted. First, Taylor series for n i;j , where superscript n + 1 denotes to new time tn+1 = tn + t, about the moment tn is developed. After the rearrangements, the approximation of the time derivative is obtained in the form
h@

+1  n   in =  n i;j i;j + Ot @t i;j t

9

where Ot is the magnitude of the truncation error, which is now proportional to t. If the ow depends on time, the approximation 9 should be replaced by a second-order accurate stencil, i.e. the truncation error should be of the order Ot2 . The approximation of convection terms is no more so straight-forward. From the physical point of view, the convection transfers the information downstream. Therefore, the approximation of convection terms must be weighted to the upstream-side. The simpliest stencil is the rst-order upwind FOU
i i

1=2;j 1=2;j

= =

1;j + Ox when ui 1=2;j when ui 1=2;j i;j + O x

0 0

10 11

i.e. the truncation error is of magnitude x. The FOU stencil was popular during the 1970s and the 1980s. It was used in itself or as a part of the HYBRID scheme developed by Spalding 53 . The main problem with the FOU scheme is its di usive e ect on the solution. The rst truncation error term can be written as follows
num @ i 1=2;j

@x i 1=2;j

12

where the numerical di usion can be written in the form inum 1=2;j =  ui 1=2;j x=2, which increases the physical di usion. Peaks or rapid variations of will be smeared out if the grid is not dense enough. From the practical point of view, the grid becomes too ne. Thus, the convection must be approximated by using higher-order schemes. Perhaps the most versatile way is to use the MUSCL approach, e.g. 56 . The 18

MUSCL scheme can be derived as follows: The original convection term in the partial di erential equation is approximated as follows
h@ i

@x

i;j

+1=2;j 1=2;j

The approximations for


i

+1=2;j

and
i i

12 13 x are searched for in the form


i = ;j

1=2;j i+1=2;j

=b =b

2;j 1;j

+c 1 +d + c + d +1
i ;j i;j i

i;j ;j

14 15

where the convection velocities are assumed to be positive at both faces, i.e. u +1 2 0 and u 1 2 0. The approximations 14 and 15 are substituted into the formula 13, and after that the Taylor series are developed for 2 , and +1 about the point i,j. Finally, the formula 13 can be 1 , obtained in the form as follows
i = ;j i = ;j i ;j i ;j i;j i ;j

h@ i

@x

i;j


x  3b c + d
@2 = b + c + d @ + @x 2
3 @x2 2 + Ox3  + 6x 7b + c + d @ 3 @x
i;j i;j i;j

16

Next the approximation 16 is required to be at least second-order accurate, that leads to the pair of equations
b

+c+d=1 3b c + d = 0

and the next limitation can be obtained: = d 1 =2 c = 2d + 3=2


b

By choosing d = 1 + =4 the better known form of the MUSCL scheme is obtained 19

1=2;j

1;j

+1 4 1



1;j

2;j

 + 1 + 

i;j

1;j

17

where the convection velocity at the face i 1=2; j  is assumed to be positive. If the convection velocity is negative, the upwind-weighted MUSCL scheme can be written in the form = 1 1 +4   + 1 +   18

1=2;j

i;j

i;j

i+1;j

1;j

i;j

An interesting feature of the MUSCL discretization is that it is always secondorder accurate and the number of stencils is in nite. However, it is upwindweighted in the region 1   1. Some of the well-known discretization methods are obtained with the choices
 

= -1 =0  = +1 3  = +1 2  = +1

    

second-order upwind Fromm's scheme third-order upwind QUICK central di erence

convection terms of incompressible ows, including commercial CFD packages. The reason for the popularity of the QUICK scheme is perhaps its third-order accuracy to approximate the cell-face value, but in fact it approximates the original convection term second-order accurately. And nally, the QUICK scheme is a stencil that can be written by choosing the value  = 1=2 in the general MUSCL scheme. The di usion terms of Eq. 6 are approximated by using the central di erence scheme
h@ i
@x
i+1=2;j

Nowadays, the QUICK scheme Quadratic-Upstream Interpolation for Convective Kinematics, 55  is perhaps the most popular higher-order stencil for modelling

i+1;j

x

i;j

+ Ox2

19

where the grid is assumed to be uniform. Also the pressure gradient terms of the momentum equations are approximated by using central di erencing. After the integration over the control volume V they are obtained in the form
i;j

20

Z
Vi;j

@p d @x

= p Ai+1=2;j p Ai

1=2;j

20

where central di erence schemes like,


pi+1=2;j

=1  pi;j + pi+1;j  + O x2  2

21

are applied. In the cartesian grid, the gravitational body force terms can be integrated straight-forwardly over the control volume.

2.3.2 Extension to Non-orthogonal Grids


When the geometry of the domain is complex, the uniform cartesian grid is unsuitable for e ective computation. Therefore, the solver must be able to handle "non-orthogonal control volumes". The control volume approach itself provides a straight-forward approach to the problem, 2 . The integral formula of the convection-di usion equation of the scalar is
Z

@ d V @t

Z
V

r r

~ v + S d

22

and by using Gauss's theorem, e.g. 58 , it can be modi ed to form


Z

@ d V @t

Z
A

 r

~ v ~ n0 d

Z
V

Sd

23

where V is an arbitrary volume, A is its surface and ~ n0 is the exterior unit normal at the face A. The integral formula 23 can directly be applied to any control volume. A typical two-dimensional non-orthogonal control volume is presented in Fig. 2. The details of the formula 23 can be studied by the aid of the face i + 1=2; j . The reference point of i + 1=2; j  lies in the middle of the surface i + 1=2; j 1=2-i + 1=2; j + 1=2. The central point of the volume i; j  is de ned as the intersection of the medians i 1=2; j -i+1=2; j  and i; j 1=2-i; j +1=2, which are shown as dashed. The unit normal vectors at the points i 1=2; j  and i; j 1=2 are not the exterior unit normals needed in the common integral formulas. 21

This mathematically uncommon choice is practical from the implementation point of view, and it must be kept in mind during the derivation of the discretized formulas.

i,j+1
(i,j1)
(i+1/2,j+1/2)

0 (i,j+1/2)

t0 n
0

i+1,j
(i+1,j)

i,j
(i1/2,j+1/2)

(i+1/2,j)

n0
(i1/2,j)

(i,j)
n0
(i+1/2.j1/2) (i,j1/2)

i1,j
(i1,j)

(i1/2,j1/2)

i,j1
(i,j1)

j i x

Figure 2. A typical 2-dimensional control volume with the cells surrounding it, and the notations used.

The mass ux through the face i + 1=2; j  can be approximated by the formula
mi

+1=2;j

Z
Ai

+1=2;j

~ v ~ n

~ v ~ n

 +1 2
i

= ;j

24

where the unit normal and the surface area can be calculated exactly in 2-D from the coordinates of the corner points. The convective ux of the scalar quantity can be calculated by assuming that the mass ux is known, which leads to
c Fi

+1=2;j

Z
Ai

+1=2;j

~ v ~ n

m _  +1 2
i

= ;j

25

22

where the value of +1 2 is approximated by using the MUSCL formula 17 or 18, i.e. changes in the distances between cell centres and in the directions of the unit normal vectors are not taken into account. These simpli cations restrict the grid generation: the spacing between the points must expand with the factor as near unity as possible, and the directions of the unit normal vectors at the cell-faces are allowed to turn as slowly as possible. The third limiting factor is the number of control volumes available. The details of the grid must always be compromised among these three items.
i = ;j

The di usive ux on the face i+1=2; j  can be approximated by using the formula
Fi
d

+1=2;j

Z
Ai

+1=2;j

r 

~ n

 r 

~ n

 +1 2
i

= ;j

26

The gradient of at the cell-face i + 1=2; j  can be expressed either in the terms of the derivatives with respect to global cartesian coordinates or local orthogonal coordinates n0; t0 :

r =

~ i @x

@ @ ~ ~ n ~0 + t +@ j = 0 @y @n @t

27

where n and t represent the coordinate directions normal and tangential to the surface i + 1=2; j , respectively. There are many ways to approximate the derivative normal to the cell-face. By using the cartesian coordinates the di usive ux can be calculated in a straightforward way from formula 26, which is not di cult to implement explicitly, but the implicit version may be complicated depending on the number of nodes involved. A good approximation is easily found by using the local orthogonal coordinate system n0 ; t0. The di usive ux can be calculated using the formula,
d Fi

+1=2;j

@ @n

+1=2;j

28

where the derivative in the n-direction can be calculated by using the central di erence scheme:

@ @n

+1=2;j

j

+1;j ~ ri+1;j
i

i;j

~ ri;j

29

where j~ r +1 ~ r j is the distance between nodes i + 1; j  and i; j . In fact, the points i; j , i + 1=2; j  and i + 1; j  in formula 29 are assumed to lie on
i ;j i;j

23

the same line, which is not true in the general case. This simpli cation is called thin-layer approximation, 59 . The limitation mentioned earlier concerning the directions of the unit normal vectors is also valid when using formula 29. Simple source terms, e.g. body force etc., can be easily integrated over the control volume. The cross product of the diagonal vectors of the cell can be used to calculate the volume of the two-dimensional cell. The order of the product is chosen so that the volumes are always positive if the counter-clockwise angle in region 0::: rad. If the angle is less than 0 or more than  the topology of the cell is not proper, and thus, the volume is negative. The formula to calculate the volume can be written as follows, between the vectors ~ ri+1=2;j +1=2

~ ri

1=2;j 1=2 

and ~ ri

1=2;j +1=2

~ ri+1=2;j

1=2 

is

Vi;j =

1 2

j~ ri

+1=2;j +1=2

~ ri

ri 1=2;j +1=2 1=2;j 1=2   ~

~ ri+1=2;j

1=2 j

30

The pressure terms occurring in momentum Eqs. Gauss's theorem

2 and

3 are treated as

source terms. The volume integral can be modi ed to surface integral by using

Z
Vi;j

@p @x

Z
d
=
r
Vi;j

p~ i d

Z
=
Ai;j

p~ i~ n0 d

31

i.e. the pressure at each face of the cell must be interpolated, e.g using the central di erence scheme, and the rest of data of the grid. 31 can be calculated using the coordinate

24

2.3.3 Time Integration


In this study the implicit time-integration is used. It is divided into two steps 60 . During the rst step an explicit residual is calculated using the approximations represented earlier. It is de ned as follows,

 i;j

expl:

t Vi;j

Fi 1=2;j

Fi+1=2;j + Fi;j 1=2

Fi;j +1=2 + Si;j 

32

where the di usion

uxes and source terms are calculated using the values from the preux is calculated using the MUSCL scheme, the ux using central di erencing and the pressure gradient using formulas

vious time-step. The convection

20 and 21. Other source terms are integrated in a straight-forward way. The explicit residual 32 is utilized during the implicit step. The implicit convection-di usion equation can be written in the form

Vi;j

n+1 i;j

n i;j

t

n+1 = Fi 1=2;j

n+1 n+1 Fi+1=2;j + Fi;j 1=2

n+1 Fi;j +1=2

33

where the source term has now been ignored for simplicity. If the convection was now approximated by using any upwind-weighted second-order accurate stencil, Eq. 33 would contain ve unknows in each direction. In two-dimensional case the total number of unknowns would be nine, which would be computationally too expensive. Therefore some simpli cations are necessary. Next, Eq. 33 is written in

delta

form.

A Taylor series for the

uxes with

respect to time about the moment n is developed. di erentiation the next formula is obtained,

By using the chain rule of

n+1

=F

@F @ n

2  + O t 

34

where  is an unknown vector :::; the ux terms 34 into Eq. manipulations.

n+1 i;j ; ::: and  = 

 . By substituting

33 the next formula can be written after some

t Vi;j

Lxx

A
x i;j

Lx  Aui;j + Lyy

A
x i;j

Ly  Av i;j

io

 i;j

impl: expl:

=  i;j

35

25

where Lxx and Lx are di usion and convection operators, respectively. The de nitions of operators will be presented later. In the right-hand side of Eq. 35 there is the explicit residual 32, i.e. the source terms are included in the implicit formula 35 via the explicit residual. Next, the approximative factorization is applied to Eq. 35 which leads to the form
n

1 1

t hL
A L  Au io i;j Vi;j xx x i;j x t hL
A L  Av io i;j Vi;j yy x i;j y

impl: i;j

=

expl: i;j

36

Eq. 36 contains extra terms compared with Eq. 35. The magnitude of these terms is Ot2, which does not decrease the accuracy. By adopting a new auxiliary quantity  # i;j , Eq. 36 can be split to form
n

t hL
A L  Au io i;j V xx x i;j x
i;j

i;j

=

expl: i;j

37

t hL
A L  Av io i;j Vi;j yy x i;j y

impl: i;j

=

i;j

38

The di usion operators Lxx and Lyy use the central di erence stencil as in the explicit part. But the convection operators Lx and Ly use the rst-order upwind stencil 10 to keep the computational costs low enough, and thus the implicit part can be solved using a tridiagonal solver, e.g. Thomas algorithm 61 . This choice does not decrease the accuracy of the convection terms, which can be seen if formula 35 is written term by term. In practice, the rst term on the righthand side of the MUSCL scheme, 17 or 18, is treated implicitly and the rest explicitly. Thus, the explicit part determines the accuracy of the solver. Now the convection and di usion operators can be written as follows,
Lxx

A

x i;j 

i;j

= +

A
A

x x

i+1=2;j i 1=2;j

 

i+1;j i;j

 i;j  
i 1;j

39

26

Lx  Aui;j 

i;j

  +  + 

Ai 1=2;j Ai 1=2;j Ai+1=2;j Ai+1=2;j

MAX 0; ui 1=2;j MIN 0; ui 1=2;j MAX 0; ui+1=2;j MIN 0; ui+1=2;j

 i 1;j  i;j  i;j  i+1;j

40

where the maximum and minimum operators MAX 0; u and MIN 0; u , respectively, are used to nd the upwind-side of the convection, see e.g. 8 .

27

2.3.4 Boundary Conditions


To close the partial di erential equation system 1- 4 for a certain problem, the boundary conditions must be speci ed. Usually the value of f ace or its at the face of the domain is given, which is called the Dirichlet derivative @ @n f ace or Neumann boundary condition respectively, e.g. 62 .
j

In the present solver, the boundary conditions are handled by using ghost cells, which are illustrated in Figs. 3a and 3b. The principal idea is to use the ghost cell values to give the xed boundary value at the boundary of the domain, i.e. the ghost cell values are extrapolated from the face and domain values. Thus, the Dirichlet and Neumann conditions are given in the form
2;j

=2
3;j

f ace

3;j

41 42

2;j

x @ @x f ace

respectively. Formulas like 41 and 42 are used to update the ghost cell values which are needed as boundary conditions during the next iteration.

ghost cells
ghost cells

domain

ghost cells

ghost cells

in next figure

Figure 3a. Ghost cells around the domain.


28

boundary of domain ghost cells domain

(1,j+1)

(2,j+1)

(3,j+1)

(1,j)

(2,j)

(3,j)

x/2 x/2 x x

Figure 3b. The notations of the ghost cells.

29

2.3.5 Flow Chart of the Convection-Di usion Solver

CD solver
Compute residual
n

Perform the implicit stage n+1

Update the solution n+1=n+n+1

Figure 4. Flow chart of the convection-di usion solver.

The ow chart of the convection-di usion solver is shown in Fig. 4. During the rst stage, the explicit residual  n is computed using formula 32. It is used during the implicit stage to compute  n+1 using formulas 37 and 38. During the third stage, new n+1 values are calculated, and when necessary the ghost cell values, i.e. boundaries, are updated. The convection-di usion solver is used to solve momentum and internal energy equations. Next, the treatment of pressure is described.

30

3 Treatment of Pressure in Incompressible NavierStokes Equations


In this study the pressure correction approach is utilized. The velocity components are calculated from the momentum equations where the pressure gradient terms are treated explicitly, i.e. the stencils are calculated using the values from the previous time-step. If the pressure eld and the cell-face velocities in the moeld satis es the continuity mentum equations are correct, the resulting velocity new pressure

equation. But if not, the mass imbalance occurs in the continuity equation. Thus, eld and cell-face velocities are needed until all the equations are satis ed accurately enough. The pressure correction method presented by Patankar and Spalding 7 was

constructed into the staggered grid arrangement, which is shown in Fig. 5.

vi,j+1/2

pi1,j

ui1/2,j

pi,j

ui+1/2,j

vi,j1/2

pi,j1
centre of ucell centre of vcell centre of pcell ucell vcell pcell

Figure 5.
u

Staggered grid arrangement 8 .

It can be noticed that, with respect to the pressure or generally scalar grid points, the the

locations are staggered only in the

locations are staggered only in the

x-direction. In a similar way, y -direction, and a corresponding threev,


and no interpolation is required.

dimensional pattern can be imaged in a straight-forward manner. The key issue of the staggered grid arrangement can be seen from Fig. 5. The cell-face velocities of

cell are the cell-centred values of

and

Furthermore, the face pressures of the momentum equations do not need any interpolation. The interpolation needed in the colocated grid arrangement see Fig. 1. may cause an oscillating solution. The spatial oscillations occur when central di erencing is applied to both the continuity equation and the pressure gradient term in the momentum equations. The momentum equations at the even-numbered nodes depend only on pressures at odd-numbered nodes, and vice versa. The same holds for the continuity equation. Because no pressure-density 31

coupling exists with an incompressible ow, no interaction exists between the pressures at adjacent nodes. This situation permits two di erent pressure elds to co-exist, which is known as "checkerboard" pressure eld, e.g. 27 , 28 , 29 , etc. The price of the staggered grid arrangement is not cheap. In the two-dimensional case, three grids are needed, one for each variable, u, v and p, and respectively, in the three-dimensional domain the number of grids is four. Nowadays, the staggered grid arrangement is no more necessary. The remarkable turn toward the colocated arrangement was the study presented by Rhie and Chow 19 , which will be investigated after introducing the pressure correction approach.

3.1 Velocity-Pressure Coupling Using Pressure Correction Approach in Colocated Grid Arrangement
The colocated grid arrangement is shown in Fig. 1. All variables use the same grid. The velocities u and v are computed from discretized momentum equations where the pressure gradient stencils and the cell-face velocities are calculated using values from the previous time-step. The discretized u momentum equation for the cell i; j  can be written in a simpli ed form as

ai;j ui;j

ai;j 1ui;j

+ ai 1;j ui 1;j + ai;j +1 ui;j +1 + ai+1;j u

 i+1;j

Z
Vi;j

@p d @x

Si;j43

where the FOU scheme can be used for convection components. This choice does not decrease the accuracy of the solver, which will be shown later. Thus, the coe cients can be written in the form

ai;j

= + + + +

A x A x A y A y A

1=2;j

 +  +

A MIN 0; u i

1=2;j

i+1=2;j

A MAX 0; u i+1=2;j A MIN 0; v i;j


1=2

i;j

1=2

i;j +1=2

A MAX 0; v i;j +1=2

44

ai;j

y

i;j

1=2

+ 32

A MAX 0; v i;j

1=2

45

ai

1;j

= = =

A x A y A

1=2;j

+  

A MAX 0; u i

1=2;j

46 47 48

ai;j +1

i;j +1=2

A MIN 0; v

i;j +1=2

ai+1;j

x

i+1=2;j

A MIN 0; u i+1=2;j

For the coe cients it holds that


ai;j

= +a  +  + = +a

i;j

i;j

+ a +1 + a +1 A MAX 0; u  1 2  A MIN 0; u  1 2 A MIN 0; u  +1 2 +  A MAX 0; u  +1 2 A MAX 0; v  1 2  A MIN 0; v  1 2 A MIN 0; v  +1 2 +  A MAX 0; v  +1 2 _ 1 + a 1 + a +1 + a +1 + m


1
i

+a

1;j

i;j i

;j

= ;j

= ;j

= ;j =

= ;j =

i;j

i;j

i;j

i;j

;j

i;j

;j

i;j

49

where m _ is the mass imbalance of the cell i; j , which is calculated using the cell-face velocities from the previous iteration. As it will be later seen, this mass imbalance is zero, if the pressure correction has been computed accurately. Otherwise, mass imbalance occurs, and its amount depends on the accuracy of the computation of the pressure correction. Eq. 43 is usually simpli ed to the form as follows
i;j

ai;j u i;j

i;j 

X a

 nb unb 

Vi;j

@p d @x

Si;j

50

where p is pressure from the previous time-step also called guessed pressure, and u is corresponding velocity. The next de nitions are adopted:
i;j

u = u + u0 v = v + v0 p = p + p0

51 52 53

where u0 , v 0 and p0 are correction terms. The key issue is to correct the velocity eld, u and v  , calculated from the momentum equations to satisfy the continuity 33

equation using the correction terms. The continuity equation 1, where the unsteady term is ignored, is integrated over the control volume i; j 

A ui

+1=2;j

A ui

1=2;j + 

A v i;j

+1=2

A v i;j

1=2 = 0

54

Formulas 51, 52 and 53 can be substituted into the integrated continuity equation 54. The velocity correction at the cell-face is the new unknown quantity. Next, the connection between pressure and velocity corrections needs to be developed. The cell-face velocities occur in Eq. 54 as in the staggered arrangement, but now the "staggering" is done by interpolating the values at the cell-faces, which will be studied in detail later. The discretized momentum equation of the cticious "cell" i 1=2; j  can be written as follows,

ai

1=2

;j ui

1=2;j

i 1=2;j 

anb u  nb

Z
Vi

@p


d

1=2;j @x

Si

1=2;j

55

and the required interpolations at the position i 1=2; j  will be studied later. The corresponding momentum equation can be written for the corrected velocity ui 1=2;j in which the partitions 51 and 53 are substituted. After subtracting Eq. 55 from the described equation the next formula for the correction terms can be derived

ai

1=2;j

0 u

1=2;j

i 1=2;j 

anb

0 u

nb

Z
Vi

@p

0
d

1=2;j @x

56

i.e. the source term disappears. Formula 56 connects velocity and pressure corrections. If it is substituted into the continuity eqution 54 the resulting equation for pressure correction p0 will become too complicated. Therefore, the term i 1=2;j  anb u0  is ignored yielding nally the formula, nb

0 u

1=2;j

Vi ai

1=2;j 1=2;j

pi;j

pi

x

1;j

57

where central di erencing has been applied to the pressure gradient. Formula 57 is the essential connection between velocity and pressure corrections in the SIMPLE algorithm 8 . A discussion of previous simpli cation will be presented later. By using formulas 54, 51, 52, 53 and formulas like 57, the next pressure correction equation can be derived,

awi;j  pi

1;j + as

i;j 

pi;j

1 + ap

i;j 

pi;j

+ aei;j  p0 + ani;j  p0 = m _ i+1;j i;j +1 i;j 34

58

where the rst subscripts w, s, e and n denote to the faces i 1=2; j , i; j 1=2, i + 1=2; j  and i; j + 1=2, respectively. The second subscript i; j  denotes to the cell over which the continuity equation has been integrated. The following equality holds ap
i;j  i;j 

= aw

i;j 

+ as

i;j 

+ ae

i;j 

+ an 
i;j 

59

_ where, e.g. aw =  AV =ax i 1=2;j and m i;j is the mass imbalance of the solutions of the momentum equations. The term ai 1=2;j in the coe cient aw stems from the momentum equation. The momentum based coe cients, e.g. ai;j and ai 1=2;j , have contained and will contain only one subscript, and the coe cients of the pressure correction equations will contain two subscripts, e.g. aw , to avoid misunderstandings.
i;j  i;j 

The motivation to drop the term Pi 1=2;j anbu0nb on the way was convenience. If the term was retained, its component would have to be expressed in terms of the pressure corrections at the neighbouring cells of u0nb. These neighbours would, in turn, bring their neighbors, and so on. Ultimately, the velocity correction formula would involve the pressure correction at all grid points in the domain, and the resulting pressure correction equation would become unmanageable. By droping the contribution of the neighbour velocity corrections, the pressure correction equation becomes manageable. But the simpli cation must also be taken into account when updating pressure. The solution of Eq. 58 overestimates the pressure correction p0 , and therefore, under-relaxation is needed, p = p +
p

The pressure correction equation 58 can be understood as an auxiliary equation to guide the solution to the right direction. It leads to a non-zero contribution when the mass imbalance is not zero. When the converged solution has been found the mass inbalance m _ i;j is zero and the solution of the pressure correc0 tion equation is p 0 over the whole domain. Thus, the pressure correction equation has no contribution to the solution itself, but it has only a guiding role during iterations. This role gives quite a free hand to construct the equation, and therefore the FOU scheme is used for convection terms and the central differencing for di usion terms when calculating the coe cients ae , etc. for the pressure correction equation. Furthermore, the simpli cation made during the derivation has no e ect on the converged solution.
i;j 

p0

60

where the relaxation factor is chosen 0 p 1. The optimal value depends on the case under investigation, and computational experiments have shown that the value p = 0:25 0:5 is a good guess, which is lower than Patankar recommended 8 . Patankar also proposed to under-relax the momentum equations in the form 35

ai;j
v

+1 un i;j

i;j 

Xa

nb unb 

Z
Vi;j

@p d @x
v

Si;j 1

a

i;j v

un i;j

61

where the relaxation factor is chosen 0 1. The action increases diagonal dominance of the matrices. In the present Navier-Stokes solver, the diagonal dominance can be increased by decreasing the time-step, and thus, the underrelaxation of the momentum equations need not be performed. The under-relaxation of pressure ensures the convergence. The optimal value of depends on the case and is probably not a constant value over the whole domain. Thus, it can be expected, that simpli cation 57 combined with underrelaxation 60 of the pressure will decrease the convergence rate. This was noticed by Patankar 8 , who proposed a revised version, SIMPLER SIMPLE Revised.
p

Doormaal and Raithby 9 proposed the SIMPLEC algorithm to improve the convergence of SIMPLE. The cost of obtaining a solution depends critically on how close to the optimal values and have been chosen. The search for the optimal values may, however, be more expensive than simply using non-optimal values. In Doormaal and Raithby's version, attention is paid to the simpli cation 57. The velocity corrections in the term  1 2  a u0  are of the same order as u0 1 2 . The SIMPLE approximation that the sum term can be ignored can therefore be seen as inconsistent. To introduce a "consistent" approximation, Doormaal and Raithby subtracted the term  1 2 a u0 1 2  from both sides of equation 56, which yields
p v

P
i

= ;j

nb

nb

= ;j

= ;j

nb

= ;j

1=2;j

i 1=2;j 

X a u0 1 2
nb i

= ;j

i 1=2;j 

anb u0nb

u0i 1=2;j 

Z
Vi

1=2;j

@p0 d @x

62

where the underlined term is ignored in SIMPLEC algorithm because the velocity corrections are of the same order. The nal form can be written as follows,
u0
i

1=2;j

ai 1=2;j

Vi 1=2;j
i

P

p0

i;j

1=2;j  anb

p0i 1;j x

63

Doormaal and Raithby maintained that SIMPLEC does not need an underrelaxation of the pressure correction and that it converges more rapidly than the SIMPLE method. The latter is also reported by Peric 63 . But the applications utilizing SIMPLEC have shown that under-relaxation is worth using, e.g. 25 , 26 and 27 , and thus the merits of SIMPLEC are partly questionable. 36

3.2 Calculation of Cell-Face Velocity in Colocated Grid


As it was mentioned, the colocated grid arrangement causes problems when central di erencing is applied to cell-face velocities. Rhie and Chow 19 presented a method for avoiding the usage of the staggered grid arrangement. They added a di erence of pressure gradients to the central-di erenced cell-face velocity. The magnitude of the pressure term is the same as the truncation error of the central di erence stencil, i.e. the accuracy is not decreased. A more detailed presentation of the formulas has been written by Miller and Schmidt 27 . The formula for the cell-face velocity u 1 2 can be derived as follows. The starting-point is the discretized u-momentum equations of the cells i; j  and i 1; j .
i = ;j

ai;j ui;j

i;j 

Xa

nb unb 

@p
@x
i;j

Vi;j

64
Vi
1;j

ai

1;j

ui

1;j

i 1;j 

X a X

nb

unb 

@p
@x
i

1;j

65

Next, a similar u-momentum equation is written for the cticious control volume i 1=2; j  overlapping half of the cells i 1; j  and i; j 
ai
1=2;j

ui

1=2;j

i 1=2;j 

a

nb

unb

@p
@x
i

1=2;j

Vi

1=2;j

66

which can also be understood as staggering. The last equation contains coe cients a 1 2 etc., which are di cult to calculate without staggering, and thus, they are interpolated. The Taylor series for the velocities u 1 and u about the point i 1=2; j  are
i = ;j i ;j i;j

ui

1;j

=u

1=2;j

x @u 2 @x 

1=2;j 

x + 8

@2u @x2

i 1=2;j 

x3 @ 3 u 48 @x3 

1=2;j 

+Ox4  67

ui;j

=u

1=2;j

x @u + 2 @x

i 1=2;j 

x + 8

@2u @x2

i 1=2;j 

x + 48

@3u @x3

i 1=2;j

+Ox4  68

By summing up formulas 67 and 68 one gets 37

ui

1=2;j

1 = ui 2

1;j + ui;j 

x2 @ 2 u 8 @x2

i 1=2;j 

Ox4 

69

and appling the central di erence stencil to the second derivative of the right-hand side one gets, 1 = ui 2 x2 ui 8

ui

1=2;j

1;j

+ ui;j 

1;j

+ ui;j 2ui x2 =4

1=2;j

70

Next, the discretized momentum equations 65, 66 and 64 are substituted into the last term of formula 70. After the manipulations one nally gets the formula, 1 Vi 2 ai
i

ui

1=2;j

1 = ui 2

1;j

+ ui;j 

1;j 1;j

@p + 2 ai 1;j @x
1;j

1 V

i

Vi;j @p ai;j @x i Vi;j @p + 1;j  ai;j @x

1=2;j 

i;j 

71

where the next central di erence approximations have been used:

i 1=2;j  

ai

anb unb 

1=2;j

1 = 2

i 1;j  

ai

anb unb 

1;j

i;j  

anb unb + O x2  ai;j

72

Vi ai

1=2;j 1=2;j

1 Vi = 2 ai

1;j 1;j

Vi;j + O x2  + ai;j

73

The Rhie & Chow interpolation of the convection velocity at the control volume face gives formula 71. It can be interpreted as the central di erence approximation of the velocity plus pressure gradient terms created by momentum interpolation. In fact, the formula can be derived in a more straight-forward manner. At rst, the momentum equation of the staggered cell i 1=2; j  is divided by ai 1=2;j . Next, the approximation 72 is substituted into it yielding the formula,

ui

1=2;j

1 = 2

i 1;j  

ai

anb unb 

1;j

i;j  

anb unb  ai;j

Vi ai

1=2;j 1=2;j

@p
@x
i

1=2;j

74

Finally, the sum terms are replaced by the sum terms extracted from momentum Eqs. 64 and 65, and formula 71 is obtained. This reveals the same origin 38

of PWIM or Rhie & Chow interpolation and MWIM. In fact, formula 74 is the "momentum-weighted interpolation" for the cell-face velocity. Thus, PWIM can be understood as "one step forward from MWIM". MWIM or the pressure gradient terms in PWIM are said to strengthen the link between the pressure and the velocity, and thus the pressure oscillations of the solution are avoided. Some modi cations have been proposed. Johansson and Davidson 26 have simpli ed the formula. By using appropriate interpolations, formula 71 can be written in the form

ui

1=2;j

1 = ui 2

1 1 1 1 + 1;j + ui;j  + Vi 1;j + Vi;j  2 2 ai 1;j ai;j  1 @p @p  @p  2 @x + @x i 1=2;j  i 1;j  @x i;j 

75

which can be simpli ed to form

ui

1=2;j

= u + CVi

1 1=2;j ai 1=2;j

 1 @p

2 @x

i 1;j 

@p @x


i;j 

@p @x


i 1=2;j 

76

Johansson and Davidson added the factor C into their formula so that the amount of damping could be altered. Good results were obtained with values  1 2 . The cell-face velocity formulas like 76 can be treated as a sum of a central-di erenced velocity and a damping term. Later on, the latter is called a damping or a local damping term, and it can also contain other components, e.g. body force terms. A relative damping term is de ned as a ratio between a damping term and a central-di erenced cell-face velocity. The pressure term of the last formula can be written in a new form by using Taylor series about the point i 1=2; j  for the pressure gradients in the points i 1; j  and i; j . After the rearrangements the next formula can be written, 1 @p 2 @x

i

@p + 1;j  @x


i;j 

@p @x

i 1=2;j 

x2 @ 3 p 8 @x3

i 1=2;j 

77

i.e. its magnitude is O x2  having probably a minor e ect on the accuracy of the cell-face velocity, which nally depends on the coe cient Vi 1=2;j ai 11=2;j . Term 77 is often called the pressure dissipation term 26 . The motivation for the modi cation proposed by Johansson and Davidson was the convergence problems in buoyancy-driven ows. A more straight-forward 39

approach has been presented by Gu 44 and Rahman et al. proposal for cell-face velocity is

49 . Gu's

rst

ui

1=2;j

V = ui 1=2;j + a

1=2;j

@p @x

1=2;j

@p @x

1=2;j

+ bi

1=2;j

bi

1=2;j 

78

where the overbars denote averages of cell values. The second-order stencil for the body force term bi 1=2;j bi 1=2;j  involves four points, which caused problems near the boundaries 44 . Therefore, Gu proposed another formula. It can be presented as

ui

1=2;j

V = ui 1=2;j + C a

@p 1=2;j @x @u + @x

1=2;j

1=2;j

@p @x @u @x

1=2;j

1=2;j

79

@p @p where C is equal to zero if the V=a @x  is greater than ui 1=2;j , i.e. if a @x damping term higher than the corresponding central-di erenced cell-face velocity is detected, it will be replaced by zero. Otherwise C is equal to one. The extra @u , is used to smooth the velocity directly. A dynamic sensor is term,   @u @x @x used to detect the oscillation in the velocity eld. It has the form

=

u 2ui;j + ui 1;j j jui+1;j j + 2jui;j j + jui 1;j j


j i+1;j

80

It is a normalized second di erence of the velocity component in the x-direction. Its value increases automatically in the region with oscillating velocity. In fact, Gu's damper with the dynamic sensor is similar to Jameson's arti cial dissipation 56 . Gu used the latter proposal 79 successfully for a turbulent buoyancy-driven ow of Ra = 1011 . Rahman et al. 49 proposed an even more straight-forward formula for cell-face velocity. They derived the formula that can be simpli ed as

ui

1=2;j

= ui

1=2;j +

x2 V a 8

@3p @x3 i

1=2;j

x V 4 a

1;j

V a


i;j

@b @x

1=2;j

81

40

where b is a body force term. As can be seen, the problem of Rahman's formula 81 is its rst-order accuracy which is caused by the last term. Rahman applied this formula successfully to thermal ows up to Ra = 107.

3.2.1 A Short Discussion about the Cell-Face Velocity Interpolation


The pressure dissipation term is of the order x2 . Finally, its magnitude depends on the coe cient V 1 2 a 11 2 . The coe cient a 11 2 comes from the momentum equation. It can be calculated using the FOU scheme and central di erencing for convection and di usion terms, respectively. The coe cient at the cell-face is calculated as
i = ;j i = ;j i = ;j

ai 1=2;j

1 =2

1 + 1 a a
i

1;j

i;j

82

ai;j

Assuming the grid is dense enough, only the term a needs to be studied, because ' a 1 . In two-dimensional space it can be written as follows
i;j i ;j

ai;j

 A

1=2;j M IN
= ;j

0; u 0; v

1=2;j
= ;j

+  A +1 2
i

M AX

0; u +1 2
i i;j

 A

i;j

1=2 M IN

1=2

A +  1 2 x
A + x +1 2
i

= ;j

+  A

i;j +1=2

M AX

0; v

i;j +1=2

+ y

A y
A

= ;j

i;j

1=2

i;j +1=2

83

i.e. there are mass ows out of the cell i; j  and the sum of di usion uxes through the cell-faces. In a uniform grid the coe cient V a 1 can be roughly simpli ed to form
i;j i;j

1 Vi;j ai;j

m _ out A

+P

L
f aces

L

84

where it has been assumed that x = y = L and V = AL. The magnitude of the coe cient decreases when the cell becomes smaller. Thus, in a dense enough grid, the converged solution x,y ! 0 should not depend on the pressure dissipation. The denominator contains the convection and di usion 41

term. Usually, the convection term is dominant. In boundary layers or in stagnant regions, the convection term decreases noticeably, which may lead to a very high damping term compared with a central-di erenced cell-face velocity. At the beginning of calculation, the term possible divergence.

Vi;j ai;j1

may get poor values causing

This risk could be eliminated by using the product of a

xed characteristic velocity and the density to replace the convection term in the denominator. The characteristic velocity could be the maximum velocity of the ow eld. This choice would also limit the growth of the coe cient 84 decreasing possible errors in boundary layers and in stagnant regions. In formula 84, the FOU scheme was applied to convection terms. If the FOU scheme was replaced by central di erencing, the mass cell-face velocity ow out of the domain cient 84, could be replaced by zero, or more precisely, by the mass balance of corrected eld. This choice could increase the value of the coe which might cause problems. The characteristic feature of the Rhie and Chow interpolation is dissipation. Its magnitude is of the same order as truncation error, and it should not have an e ect on the solution itself. oscillations. But It could be treated as a damping that prevents nally, it is not well known, if the Rhie and Chow interpolation

causes divergence at the beginning of iteration, or how much e ect it has on the nal solution, especially in regions of low velocity. Another open question is the damping term itself. As it was shown, the coe cient cient

ai;j

is quite complex and three pressure gradient stencils are needed at each face.

In fact, the coe driven

ai;j

is also used in the pressure correction equation, and no

additional computing is needed. But convergence problems reported in buoyancyows, e.g. 44 and 26 , may make the formula quite complicated, as shown by Gu 44 and Rahman et al. 49 . In this study the magnitude of the damping term is studied. It is compared with central-di erenced cell-face velocity. If damping seems to be to high, the usage of some simple limiters is tried. Also, a simpler damper with a a calculation. For buoyancy-driven studied, and xed convection velocity is tested if it can prevent convergence problems at the beginning of ows, the need of the body force terms is nally, as simple cell-face velocity stencil as possible is sought.

42

3.3 Solution Sequence


The ow chart of the present Navier-Stokes solver is presented in Fig. 6. The CD solver denotes Fig. 4, where its ow chart was presented.
Initial guess
p*i,j ui1/2,j vi,j1/2
new iteration begins

umomentum
> u*i,j (CD solver)

vmomentum
> v*i,j (CD solver)

Cellface velocities
> u*i1/2,j v*i,j1/2

Mass balance
u*i1/2,j v*i,j1/2 > m*i,j ||m*i,j||2< ||m*i,j||2>

Result

Pressure correction eq.


> pi,j > pi,j > p*i,j:= pi,j

End

Velocity correction
> ui1/2,j vi,j1/2 > ui1/2,j vi,j1/2 isothermal flow thermal flow

Internal energy equation


> Ti,j (CD solver)

Figure 6. Flow chart of the present Navier-Stokes solver. 43

In the beginning, an initial guess for pressure, and

1=2 , are used to compute new momentum equations, respectively. The convection-di usion solver previously described is used.  1=2;j and vi;j 1=2 , are calculated based on the solutions of momentum equations. Rhie and Chow interpolDuring the fourth stage, new cell-face velocities, ation 76 or its modi cation is used. During the next stage, these cell-face  velocities are used to calculate the mass balance m _ i;j . L2 norm of mass bal ance, jjm _ i;j jj2 , can be used to check the convergence. If the value is smaller than convergence criteria will be continued. During the sixth stage, the pressure correction equation will be studied later. The new pressure, 58 is solved, which

vi;j

p , and cell-face velocities, ui 1=2;j i;j   velocities ui;j and vi;j from u- and v-

u i

",

the solution is accepted, otherwise the computation

is calculated using formula 60.  This new pressure is used during the next iteration as guessed pressure, pi;j . During the seventh stage, the velocities are corrected using formulas 51 and The

pi;j ,

52.

The velocity corrections are calculated using formulas like

57.

corrected cell-face velocities

ui

1=2;j

and

vi;j

1=2 will be used during the next

iteration. On priciple, these cell-face velocities satisfy the mass balance. But in practice, some error occurs which depends on the accuracy of the solution of the pressure correction equation. The cell-face velocities could also be updated using corrected cell values, but as it was mentioned earlier, it is quite unusual. If the ow is isothermal, the new iteration cycle begins. Otherwise, the internal The corrected uxes. The values of the

energy equation is solved using the convection-di usion solver. cell-face velocities are used to calculate the convection new temperature during the next iteration. This cycle is repeated until a convergence solution is obtained.

eld are used in the source terms of the momentum equation

44

4 Pressure Correction Equation


When computing incompressible Navier-Stokes equations using a pressure correction approach, a Poisson-type partial di erential equation for the pressure correction eld is derived 8 . During the derivation of the equation, simpli cations have been made and thus there is no need to compute the machine-accurate solution. Furthermore, the nonlinear momentum equations are solved iteratively and, therefore, the requirement for accuracy of the solution to the pressure correction equation can be decreased. In addition to these two observations, the necessity to under-relax the pressure correction caused by the simpli cations mentioned earlier has a similar e ect. And after all, the pressure correction is an auxiliary quantity to guide the Navier-Stokes solver to converge, having nally no e ect on the solution itself. The nal requirement for accuracy of the pressure correction eld probably depends on the ow eld to be solved. Therefore, perhaps the easiest way to avoid troubles is to develop as e ective a Poisson-solver as possible. The solution techiques can be divided into two branches, single-grid and multigrid techniques. In this study the multigrid technique is chosen. In the family of multigrid technique there are some named methods such, as a full multigrid FMG and a full approximation storage FAS 64 , 2 . The former seeks a starting solution from coarser grid levels and the latter computes the unknown quantity in all grid levels. In this paper the rest are simply called multigrid MG techniques. The multigrid method needs a smoother to compute the solution for each grid level. A great number of iteration methods are available, from simple to more sophisticated and complex ones: Jaboci, Gauss-Seidel, Successive Overrelaxation SOR, Conjugate Gradient method etc, 40 . The convergence rate di ers noticeably, depending on the largest eigenvalue of the iteration matrix associated with the method 2 . Because the idea is to keep the solver as simple as possible some words about the Jacobi and Gauss-Seidel methods can be mentioned. For Laplace's equation, which is the Poisson equation with the zero source term, the two largest eigenvalues absolut values of the Jacobi method are real and of the opposite sign. The corresponding eigenvectors are smooth and rapidly oscillating functions of the spatial coordinates, which make the convergence error of the Jacobi method a mixture of very smooth and very rough components. This makes the acceleration of the convergence rate di cult. On the other hand, the largest eigenvalue of the Gauss-Seidel method is single real positive and, therefore, the convergence error is smooth. The convergence error of the Gauss-Seidel method is a smooth function of the spatial coordinate. Therefore, an approximation of the convergence error can be computed on a coarser grid. In a two-dimensional case on a grid twice as coarse 45

as the original one, the computational cost per iteration is 1 4 as much and, furthermore, iterative methods converge much faster on coarser grids. GaussSeidel converges four times as fast on a grid twice as coarse 2 . A part of the iteration work can be done on coarser grid levels. To construct a solid MG-solver for the Poisson equation, some de nitions are needed. Firstly, the relationship between two grid levels must be de ned. Next, the nite di erence operator of the nest grid level must be tranformed consistently to the coarser levels. Calculated data of the nest grid level must be tranformed to the next coarser level. It is done using a proper restriction operator, which must be de ned. The restriction operator smooths data to the coarser grid level, and therefore, it is sometimes called the smoothing operator. Data must also be interpolated from a coarser grid level to the next ner one, which is done using a prolonging operator. With proper choices, much computational cost can be saved. This study concentrates on the computationally economical implementation of FAS and MG techniques utilizing the line-Gauss-Seidel LGS method as a smoother. These methods are compared with each other and with the single-level line-Gauss-Seidel SL-LGS method in two-dimensional space. An extension of the present Poissonsolver for three-dimensional space will be performed for another Navier-Stokes solver, and a short discussion is presented.

4.1 Fine-Grid Problem


The pressure correction equation 58 was derived by substituting partitions 51 and 52 into the continuity equation integrated over a control volume i; j . Velocity corrections were replaced by formulas like 57 yielding to partition

ui 1=2;j

= =

ui 1=2;j ui 1=2;j

+ u0i

1=2;j
@p

V a 1 i

0
i

1=2;j

@x

1=2;j

85

and then the continuity equation can be written as

AV a

1 @p

0
i+1=2;j

@x

AV a

1 @p

0
i

AV a

0 @p
@y
i;j

@x

1=2;j
Au

AV a

1 @p

0
i;j +1=2

@y Au

1=2

= +



+ Av  i;j +1=2 46

i+1=2;j



 Av  i;j

1=2;j 1=2

86

which was earlier simpli ed to form 58. The term V 1 2 is the volume of the staggered control volume i 1=2; j , and the term a 11 2 is the coe cient of the u-momentum equation at the same location. In fact, the "staggering" in the colocated grid arrangement is performed by interpolating the required staggered values from cell-centred values. At the actual control volume i; j  the content of a was presented in formula 83.
i = ;j i = ;j i;j

The relevant boundary conditions for the pressure correction equation can be derived from the boundary conditions of velocities and pressure. A xed velocity on a boundary yields to zero pressure correction gradient, which can be seen from formula 85, i.e. velocity correction at the face is zero. On an open boundary, where the pressure is xed, the corresponding pressure correction becomes zero.

4.2 Multigrid Operators


4.2.1 Grid Transfer
The coarse cell-centred grid is presented in Fig. 7. In two-dimensional space four ne-grid cells de ne a coarse-grid cell, e.g. the cell I,J contains ne-grid cells i 1; j 1, i; j 1, i 1; j  and i; j .

j+2

(I1,J+1) (I1,J)

(I,J+1)
(i1,j+1) (i,j+1)

(I+1,J+1) (I,J) (I+1,J)

j+1

(i2,j)

(i1,j)

(i,j) (i+1,j)

(i2,j1)

(i1,j1)

(i,j1)

(i+1,j1)

j1

(i1,j2)

(i,j2)

j2

(I1,J1)

(I+1,J1) (I,J1)

j3

i3

i2

i1

i+1

i+2

Figure 7. De

nitions between the ne and coarse-grids. 47

4.2.2 Coarse-Grid Problem and Restriction Operator


The Poisson equation of the coarser grid could be derived in the same manner as on the ner grid, i.e. by calculating the geometrical properties and cell-face velocities for the coarser grid. This is called Discretization Coarse-Grid Approximation DCA 41 . But the pressure correction equation is linear and therefore the coarse-grid problem can be approximated accurately enough from the ner one. The process is called Galerkin Coarse-Grid Approximation GCA 41 . The mass balance of the coarser grid cell I; J  can be written using ner grid cells i 1; j 1, i 1; j , i; j 1 and i; j . The integration over the cell I; J  can be performed by adding the integrated continuity equations of the ner grid cells, which is naturally valid for discretized equations. After manipulations a discretized continuity equation for the cell I; J  can be written as follows.
asi p0 awi 1;j 1 p0 ani 1;j 1 p0 aei 1;j 1 p0 i 1;j 2 i 2;j 1 i 1;j i;j 1 0 0 0 api 1;j 1 pi 1;j 1 asi;j 1 pi;j 2 awi;j 1 pi 1;j 1 ani;j 1 p0 i;j 0 0 0 aei;j 1 p0 a p a p a p p s w i+1;j 1 i;j 1 i;j 1 i 1;j  i 1;j 1 i 1;j  i 2;j
0

1;j

+an

+ +

1

i 1;j 

+anw +anw +anw

pi 1;j +1 aei 1;j  pi;j api 1;j  pi 1;j asi;j  pi;j 1 ani;j  p0 aei;j  p0 api;j  p0 aswi i;j +1 i+1;j i;j
0 0 0

i 1;j i;j

p0 i 2;j 0 p 1 i 1;j
1

i 1;j 

p0 i

+ ane pi;j + ase pi;j 2 + asw + ane pi+1;j + ase pi+1;j 2 + asw pi 2;j 1 pi;j +1 + ase pi;j 1 + asw pi 1;j 1 2;j +1 + ane +anw pi 1;j+1 + ane pi+1;j+1 + ase pi+1;j 1 = m _ i 1;j 1 + m _ i 1;j + m _ i;j 1 + m _ i;j
0 0

+
0

+ aw
1;j i;j

i;j 

p0 i 1;j

1

i 1;j

1

i 1;j

1
0

p0 i 2;j 2 0 p 1 i 1;j 2
0 0

i;j

1

i;j

1

i 1;j 

i 1;j 
0

i 1;j  i;j 

i;j 

i;j 

i;j 

87

where the coe cients of the intercardinal points, e.g. asw , are zero when moving from the nest grid level to the next coarser one, but otherwise not necessarily. It depends on the interpolation method to be used. The left side of Eq. 87 can be written as a mass balance of a cell I; J , i.e. m _ I;J . The corresponding restriction operator, R : p ! p , where the overbar indicates the coarse-grid quantities, is
i 1;j 1
 0 0

p0 I;J

=1 4 pi
0

1;j 1

+ pi;j 1 + pi
0 0

1;j

+ pi;j 
0

88

where the grid has been assumed to be uniform, and the fraction 1 4 has been cancelled from both sides of the equation 87. The restriction operator 88 is simpler than the operators used by Crumpton et al. 65 or Wesseling 66 , e.g. 48

pI;J
0

1 2p = 16 i 1;j 1 + 3pi;j 1 + 3pi 1;j + 2pi;j + pi;j 2 + pi+1;j 2 + pi+1;j 1 + pi 2;j + pi
0 0 0 0 0 0 0 0 0

2;j +1

+ pi
0

1;j +1

89

which interpolates exactly polynomials of degree at most 1, and the corresponding value of the operator 88 is 0. The order of the operator mR is de ned to be the maximum degree of exactly interpolated polynomials plus 1 66 ; hence for 88 we have mR = 1 and for 89 mR = 2. The later operator is more accurate, but its implementation is more complicated and problems might be faced beside the boundaries. Therefore the less accurate formula 88 is chosen. The extension of formulas 87 and 88 to three-dimensional space is straightforward. The coarse-grid value pI;J;K is calculated using eight surrounding negrid values pi 1:::i;j 1:::j;k 1:::k with common coe cient 1 8 , which can be cancelled from both sides of Eq. 87.
0 0

4.2.3 Prolongation Operator


Next, the prolongation operator, P : p ! p , must be de ned. The pressure correction values of the ner grid are interpolated using the values of the coarser grid, e.g. pi;j is interpolated using the values pI;J , pI +1;J , pI;J +1 and pI +1;J +1. The Taylor series of these four values are developed around the point i; j  giving
0 0 0 0 0 0 0

pI;J
0

@p = pi;j 1  x2 4 @x
0

i;j

1 y @p + 1 x y @ 2 p 4 2 @y i;j 16 2 2 @x@y
0 0

i;j

90 91 92 93

pI +1;J
0

@p = pi;j + 3  x2 4 @x
0

i;j

1 y @p 4 2 @y

i;j

3 x y @ 2 p 16 2 2 @x@y
0

i;j

pI;J +1
0

@p 3 y @p  x2 + = pi;j 1 4 @x i;j 4 2 @y
0 0 0

i;j

3 x y @ 2 p 16 2 2 @x@y
0 0

i;j

pI +1;J +1
0

@p 3 y @p + 9 x y @ 2 p = pi;j + 3  x2 + 4 @x i;j 4 2 @y i;j 16 2 2 @x@y


0 0

i;j

where the grid is assumed to be uniform and the subscript 2 in x and y denotes the coarser grid. The four-point interpolation formula of pi;j can be derived as follows
0

49

pi;j
0

9p + 3p 3p 1p = 16 + + +1 +1 16 16 16 +1 +1  ! 2 @ p @2p 3 2 2 + 32 x2  @x2 +y2 @y2


0 0 0 0

I ;J

;J

I ;J

;J

i;j

i;j

94

The interpolation is illustrated in Fig. 8. The point i; j  is inside the interpolation area de ned by the points I; J , I + 1; J , I; J + 1, I + 1; J + 1. The process can also be understood as nite element interpolation. The gure also explains why the coe cients of intercardinal points of the continuity equation of the coarse grid are always non-zero when using the four-point formula. Thus, these points are needed when transforming the problem from the second grid level to the third level etc. The contributions of intercardinal points are zero if a three-point formula is used. The formula is illustrated in Fig. 9, and it can be written as
pi;j
0

1 1 = 1 p + p +1 + p +1 2 4 4  ! @2p @2p @2p 3 1 2 2 + 16 x2 y2 @x@y + 32 x2  @x2 +y2  @y2
0 0 0

I ;J

;J

I ;J

i;j

i;j

i;j

95

pi+1;j
0

3p 1 = 4 +1 + 4 p +1  ! 3 x y @ 2 p + 3 x 2 @ 2 p +y 2 @ 2 p 96 2 @x2 2 @y 2 16 2 2 @x@y 16


0 0

;J

I ;J

i;j

i;j

i;j

Now the interpolation area is a triangular element de ned by coarse-grid points I; J , I + 1; J  and I; J + 1, which are used to interpolate the values of the ne-grid at points i; j , i + 1; j  and i; j + 1. Otherwise the intercardinal coe cients are no more zero. The three-point formula must be used in the corners of the domain to avoid the usage of the corner ghost cell. Other cells beside the boundary do not cause any problems with the four-point formula.

50

(I,J+1)

(I+1,J+1)

(i,j)

(I,J)

(I+1,J)

border of coarsegrid cell border of finegrid cell border of interpolation area

Figure 8. Four-point interpolation.

(I,J+1)
(i,j+1)

(i,j)

(i+1,j)

(I,J)

(I+1,J)

border of coarsegrid cell border of finegrid cell border of interpolation area

Figure 9. Three-point interpolation.


Thus the interpolations, or prolongation formulas, needed to transform the negrid problem to a coarse one are available and the restricted continuity equation 87 can be written using coarse-grid points. The simpli ed form can be written as follows,
aswI;J  pI 1;J 1 + asI;J  pI;J 1 + aseI;J  pI +1;J 1 + awI;J  pI 1;J +apI;J  pI;J + aeI;J  pI +1;J + anwI;J  pI 1;J +1 + anI;J  pI;J +1 +aneI;J  pI +1;J +1 = m _ I;J
0 0 0 0 0 0 0 0 0 

97

51

where the coe cients asw , etc. are calculated using ne-grid values asw , etc., and the weighting is based on the restriction and prolongation operators used. In this study, the four-point formula is used except for the corners where the three-point scheme is applied. The GCA process is often presented in compact form
I;J  i;j 

Ap = m _ ! RAPp = Rm _
0  0

98

i.e. both sides of the ne-grid problem are rst restricted and then the vector of unknowns is prolongated. An extension of the four-point interpolation to three-dimensional space is straightforward. Formula 94 is replaced by an analogous eight-point formula. A problem arising is the number of ne-grid points involved to the coarse-grid continuity equation. When coarsening the nest grid level, the number is 32 and otherwise 64. Taking into account the eight points of the interpolant and the number of the coe cients of each ne-grid pressure correction term, a very "long" coarsened continuity equation is obtained. Its implementation becomes tedious. But easier interpolants do not help much. The simplest interpolant is a four-point formula, i.e. tetrahedral element. The rst problem is that a tetrahedron contains only one ne-grid point. Much savings are not achieved, because 32 or 56 ne-grid points are involved in the coarse grid continuity equation. Furthermore, in the uniform grid, the coe cient of the four-point interpolant is a common 1 4 yielding to the coarse-grid continuity equation, where the diagonal coe cient is identically zero. Thus the four-point formula must be forgotten. Five-, six- and seven-point formulas do not give much savings in implementation when compared with the eight-point formula. And furthermore, the eight-point formula seems to be the most stable choice in three-dimensional space.

52

4.2.4 Accuracy of the Operators


As noted, the accuracy of the restriction operator 88 is mR = 1. The prolongation operators 94, 95 and 96 interpolate exactly polynomials of degree at most 1: hence we have mP = 2. To become a successful process, the GCA must satisfy the following simple condition 41 :

mP + mR 2 m

99

where 2m is the order of the partial di erential equation to be solved, and now 2m = 2. Thus, in the present investigation, the condition 99 is satis ed, i.e. 2 + 1 2. Practical experiences have shown the necessity of 99. It is also mentioned 41 that the theory would claim more accurate operators, but the condition 99 is found to su ce in practice. More about the theory of the accuracy is discussed in reference 41 .

53

4.3 Multigrid Technique


Three techniques are used in the computations. An SL-LGS iteration is the simplest but also the most time-consuming method. The easiest way to speed up the convergence is to utilize an FAS technique combined with the Gauss-Seidel solver. The MG-LGS solver is developed from the FAS-LGS solver with small modi cations. In the two next subsections the solution sequences of both solvers are presented.

4.3.1 The Solution Sequence of the FAS Method


The basic idea of the FAS technique is to transfer the unknown variable between di erent grid levels. V-cycles with a xed number of iterations on each grid level are used, which is the simpliest way to drive the solver. The solution sequence of the V-cycle downward, i.e. from ner grid to coarser one, is as follows:

0. initialization: as a guess, use previous result or zero. 1. compute LGS-iterations on the rst  nest grid level. 2. make the coarser grid approximation of ner grid problem, i.e. the
process 98. 3. restrict the result of the ner grid level to the coarser one 88. 4. compute LGS-iterations on the coarser grid level starting from the interpolated result of step 3.

Steps 2-4 are repeated until the coarsest grid level is reached. From the second V-cycle, step 2 can be skipped because the problem of each grid level can be preserved during the rst V-cycle. The solution sequence of the V-cycle upwards is as follows:

5. interpolate the ner grid result from the coarser one, operator 94. 6. compute LGS-iterations on the ner grid level by using the interpolation of step 5 as the starting-point.
Steps 5-6 are repeated until the nest grid level is reached. The number of LGSiterations on each grid level is xed. The equations to be computed during step 6 are the same as were computed during step 4 of the corresponding grid level. Figure 10 illustrates the FAS-LGS V-cycle.

54

1. level

LGS

LGS

Illustrates the Galerkin coarsegrid approximation (GCA) of the problem.

2. level

LGS

LGS

Illustrates the restriction process of the result.

Illustrates the prolongation process of the result. 3. level

LGS
LGS

Illustrates the lineGauss Seidel iterations.

Figure 10. A schematic gure of the three-level FAS-LGS solver.

55

4.3.2 The Solution Sequence of the MG Method


The principal task is to compute the solution of equation 58 on the nest grid. The nal result should not depend on the solution of the coarse-grid levels. After performing n iterations on the nest grid, an approximative solution p n is n: obtained, and equation 58 is satis ed to within the residual i;j
0

asi;j pi;j
0

n n n n 100 _ i;j i;j + aw i;j pin 1;j + an i;j pi;j +1 + ae i;j pi+1;j + ap i;j pi;j = m
0 0 0 0 

Subtracting this equation from 58 gives


asi;j
n i;j 1 n i;j

+ aw i;j


n i 1;j

+ an i;j


n i;j +1

+ ae i;j


n i+1;j

+ ap i;j


n i;j

n = i;j

101

where the

is the convergence error


n i;j
n = pi;j pi;j
0 0

102
0

where pi;j is "the machine accurate numerical solution". The coarse-grid problem of 101 is derived by using formula 98, where the unknown p is replaced by n and the source term m _ by n. It is practical to note that the coe cients of the coarser grid matrix of MG are the same as they are in the FAS application. The boundary conditions of the convergence error n i;j are as follows:
0 

p1:gc = fixed 
0

1:gc

=0
1:gc

103 =
lc

@p @x

face

= 0:  p1:gc = plc 
0 0

104

where subscript 1:gc denotes the rst ghost cell and lc denotes the last calculated cell. The implementation of boundary conditions of is as straightforward as it is for p .
0

56

The solution sequence of the MG-LGS solver is as follows:


0. 1. 2.

initialization: as a guess, use previous result or zero. compute LGS-iterations on the rst grid level. calculate the residual n from equation 100 or from the corresponding one. 3. make the coarser grid problem by approximating equation 101, the process 98. 4. compute LGS-iterations on the coarser grid level.

Steps 2-4 are repeated until the coarsest grid level is reached. It is worth noting that on the third grid level the convergence error is being computed to correct the result of the problem on the second grid level; the problem of the second grid level seeks the correction to the problem of the rst grid level. The process is recursive, and the deeper in the V-cycle, the more hazy the unknown quantity is. During the rst V-cycle, the coe cients of the equations on each grid level are preserved, and thus, from the second V-cycle, step 3 can be skipped. When coming upwards in the V-cycle, the result of the next ner grid is corrected. New iterations are not executed in the present solver.
5.

interpolate the ner grid correction, i.e. convergence error 102, from the result of the coarser grid by using the prolongation operator 94. 6. correct the unknown on the ner grid, which was calculated during the downward steps, by adding the interpolant of step 5, by using the idea of formula 102.

Steps 5-6 are repeated until the nest grid level is reached. Figure 11 illustrates the MG-LGS V-cycle.

57

1. level

LGS n LGS n

C
Illustrates the Galerkin coarsegrid approximation (GCA) of the problem

2. level

Illustrates the interpolation process of the convergence error .

C 3. level

LGS
LGS

Correction of the unknown of present grid level.

lineGaussSeidel iterations.

LGS n

lineGaussSeidel iterations & the calculation of the resi dual of discretized problem.

Figure 11. A schematic gure of the three-level MG-LGS solver.

58

5 Test Calculations
Firstly, the test calculations of Poisson-solvers are presented. The most e ective solver is applied to tests of the Navier-Stokes solver, which contains the twodimensional cases: lid-driven cavity ow, inviscid channel ow, inviscid ow in a diverging channel, and buoyancy-driven cavity ow, which are used to study di erent cell-face velocity stencils.

5.1 Poisson Equation


The test problem used is a pressure correction equation taken from the fth iteration of a 2-D lid-driven cavity ow, which will be described on page 68. The rst-level grid is 64  64, and the maximum number of grid levels to be used is 5. The grids are 64  64, 32  32, 16  16, 8  8 and 4  4. All the grids are uniform. In a cavity ow, the mass ux through the boundaries of the domain is zero, which yields the boundary condition @p =@n = 0 around the domain. But then the pressure correction is nowhere speci ed, which might cause problems. Therefore it is xed in the ghost cells above the moving wall, which limits only the absolute values but not the relative ones.
0

Figure 12. The machine-accurate results of the original 64  64 problem and its GCA-problems 32  32, 16  16, 8  8 and 4  4 at X = 0:5.
59

First, the machine-accurate results of the original problem and its GCAs of four coarser grids were computed. The distributions of the results at mid-line X = 0:5 are shown in Fig. 12. A reason for the deviations seems to be the xed pressure correction p = 0 of the ghost cell row above the moving top wall. The zero value was given in the rst ghost cell, which can be seen in Fig. 12 by extrapolating the result to the centre of the ghost cell. And of course, the approximation of the original problem also contributes to the results. In the gure there are also the CPU times used and the number of LGS iterations needed to get the machine-accurate result.
0

When an SL-LGS-solver is used, the result develops very slowly, which is illustrated in Fig. 13. The iterations were started using zero as the initial condition over the whole domain.

LGS-solver was used. The number of iterations and CPU time as parameters.

Figure 13. The development of the result at X = 0:5 when SL-

60

The simplest technique to speed up the convergence rate is FAS-LGS. In Fig. 14 the results after one V-cycle are shown. The studied parameters are the number of grid levels and the number of LGS iterations on each level. The latter was not varied on di erent grid levels, e.g. in a three level V-cycle the number of LGS iterations was the same on the rst, second and third grid. The more grid levels used, the better can be achieved the result within a CPU time of the same order. But this is just the rst side of the method. Another side can be seen in Fig. 15, where the number of V-cycles is increased. The solution does not develop after some cycles due to the disturbance of the coarser grids: the result of the coarser grid changes the ner one in the wrong "direction". This was also found by Salminen 67 . He solved this weakness of FAS by using a more complex control structure than a xed V-cycle. Nevertheless, the control system was found to be not intelligent enough, sometimes causing ine ective computing. Thus he put the solver to compute SOR-iterations as the nal step of the solution sequence. But the solver became quite complex and vulnerable, indicating that a simpler solution technique should be developed.

Figure 14. The results of the FAS-LGS solver after one V-cycle at X = 0:5. The number of grid levels used and the number of LGS iterations on each level as the parameters. 61

Figure 15. The development of a 4-level FAS-LGS solution at X = 0:5. The number of V-cycles as a parameter. On all the levels 10 LGS iterations are used.

62

In Figs. 16-18 the development of the MG-LGS results is shown. The parameters are the number of grid levels, the number of LGS iterations in each grid level and the number of executed V-cycles. Any increase in the previous parameters improves the result, as can be expected. In Fig. 16 the three-level MG-LGS results are presented. The number of V-cycles are 1, 4, 7, 15 and 40. During each cycle the number of LGS iterations was 1, 3 and 5 on the 1st, 2nd and 3rd grid levels, respectively. The LGS iterations are computed when moving downwards in the V-cycle. When coming upwards only the corrections are executed. A clear improvement compared to a singlelevel LGS computation can be seen. Also the improvement in convergence rate is clear compared with the two-level MG-LGS execution, which is not shown in the gures.

Figure 16.

X = 0:5. The parameter varied is the number of V-cycles.

The results of the MG-LGS solver of 3 grid levels at

63

In Fig. 17 the four-level MG-LGS results are presented. The number of V-cycles are 1, 2, 5, 8 and 20. During each cycle the number of LGS iterations was 1, 2, 4 and 5 on the 1st, 2nd, 3rd and 4th grid levels, respectively. The improvement, compared with the three-level MG-LGS computations can be observed, but it is not dramatic. After the rst and the second V-cycle, the result contains an interesting feature. In region 0 7

1 0 the results turn clearly towards zero. xed zero values above the top

This is probably caused by the ghost cells having wall. This feature can also be seen in Fig. 18.

Figure 17.

The results of the four-level MG-LGS solver at

= 0 5.

The parameter varied is the number of V-cycles.

64

In Fig. 18 the

ve-level MG-LGS results are presented. The number of V-cycles

are 1, 2, 4, 8 and 16. During each cycle the number of LGS iterations was 1, 2, 3, 4 and 5 on the 1st, 2nd, 3rd, 4th and 5th grid levels, respectively. The fourand ve-level MG-LGS computes the converged solution within a quite similar CPU time. However, using the 5-level scheme, the correct shape of the pressure correction pro le is obtained in a few V-cycles. This may in practice increase the robustness of the solver in comparison to the four-level method.

Figure 18.

The results of the

ve-level MG-LGS solver at

The parameter varied is the number of V-cycles.

X = 0:5.

65

In Fig. 19, all the methods are compared by

xing the CPU time of the solver.

The power of the MG technique is notable, which is also demonstrated in Figs. 20 and 21. The convergence of a single-level LGS solver is very slow. The FAS-LGS solver using a xed V-cycle cannot improve the result after some cycles. The MGLGS solver is clearly the most e ective. In this computational experiment, the machine-accurate result can be obtained using the MG-LGS solver within 1 145 CPU time compared with SL-LGS solver. The number of grid levels should be at least four so that the power of the MG-LGS method can be utilized. In Figs. 20 and 21 the dimensional residuals are given using SI units. And later on, dimensional quantities are always given using SI units

Figure 19.
time is

The results of di erent solvers at

xed at about 0.2 seconds.

= 0 5. The CPU

66

Figure 20. The convergence history of di erent solvers.

Figure 21. The convergence history of the MG-LGS solver with di erent number of grid levels. Each V-cycle has been marked. 67

5.2 Lid-driven Cavity Flow


A lid-driven cavity ow, Fig. 22, is the most commonly used computational experiment to demonstrate the performance of an incompressible Navier-Stokes solver. The top wall of the closed cavity moves with a uniform velocity UW causing a ow eld via the viscous e ect at the uid-solid interface. The big vortex is sometimes called primary vortex and the three smaller ones, secondary vortices. The existence and the size of the secondary vortices depend on the Reynolds number de ned as
U L ReL = W 

105

where L is the length of the top wall and  is kinematic viscosity. A comprehensive collection of results for di erent Reynolds numbers has been presented by Ghia et al. 68 . An extraordinary feature of the article is that numerical results are presented, which makes the comparison more accurate than usual.

u=Uw v=0

u=0 v=0

u=0 v=0 L

u=0 v=0 L

Figure 22. A schematic gure of the lid-driven cavity ow.

68

5.2.1 Re = 400
The grid used in test computations for Reynolds number 400 is shown in Fig. 23. The number of control volumes is 64  64, the height of the rst cell adjacent to the wall is L=250 yielding to expansion factor 1:0764. The biggest aspect ratio of a cell is almost ten.

Figure 23. Computational grid used in lid-driven cavity ow of


Re=400. All computations were performed using non- xed pressure correction, i.e. boundary condition @p @n = 0 was used over the whole domain. During each outer iteration, two multigrid V-cycles were computed to obtain the pressure correction eld. Five grid levels were used. During each V-cycle the number of executed LGS iterations was 2, 3, 4, 5 and 10 on the 1st, 2nd, 3rd, 4th and 5th grid levels respectively.
0

69

First, the modi ed Rhie and Chow interpolation 76 by Johansson and Davidson 26 was tested. The damping term was multiplied by factor C with values 1.0, 0.5, 1 2 0.1, 0.01 and 0. In Figs. 24-26 the dimensionless u=UW , v=UW and p= 2 UW  pro les on the vertical midplane are presented. The u=UW and v=UW pro les do not depend visibly on the damping term. The di erences in dimensionless pressure pro les are quite constant, i.e. only the absolute pressure levels di er and have no e ect on the ow eld itself. If the damping term is zero, a very lightly oscillating pressure eld will be obtained, which is demonstrated in Fig. 27, where a strongly magni ed section of Fig. 26 is shown. In practice, the oscillation of the pressure beside the bottom wall is insigni cant and cannot be identi ed from Fig. 26.

Figure 24. Dimensionless velocity pro les u=UW on the vertical midplane x=L = 0:5, Re = 400.

70

Figure 25. Dimensionless velocity pro les v=UW on the vertical mid-

plane x=L = 0:5, Re = 400.

midplane x=L = 0:5, Re = 400.

1 2 UW  on the vertical Figure 26. Dimensionless pressure pro les p= 2

71

2 Figure 27. Dimensionless pressure pro le p= 1 2 UW  at x=L = 0:5

without damping, Re = 400.

The detailed statistics of the damping terms over the whole domain prevails as follows. By using the value C = 1, which corresponds to the original Rhie and Chow interpolation, the largest absolute value of the relative damping term
CVi
1=2;j ai 1 1 2

@p @p = ;j @x i 1;j+ @x i;j 1 2 ui 1;j + ui;j 


1 2

@p @x i 1=2;j

106

was about 16. Statistics for each value of factor C are presented in Table 1, where the number of the relative damping terms are shown in groups 1=100 ::: 1=10, 1=10 ::: 1 and 1. C=1.0 C=0.5 C=0.1 C=0.01 C=0.0 1 100 ... 1 10 1 10 ... 1 154 26 123 13 44 4 9 1 1 3 1 -

terms. The total number of u-face-velocities is 4 032.

Table 1. The statistics of the relative magnitude of the local damping

It is natural that the number of relatively large damping terms decreases with decreasing factor C , and the damping is zero when C = 0. But an interesting 72

point is that the number of big damping terms in the original Rhie and Chow interpolation method is quite high; almost 5  of u-face-velocities have a damping term at least 1  of its absolute central-di erenced value. Relatively large damping terms are found near the boundaries, in the middle of the primary vortex and at the interfaces of the vortices. Low velocity is a common feature for these three regions. But what is the component of the damping term that produces the increase? A damping term consists of a volume, a coe cient of a momentum equation and a subtraction of a pressure gradient stencils. In this particular case these values vary as follows

X @p

Vi ai
i

1=2 1=2

= 1:6  10 5 ::: 1:5  10 = 2:5  103 ::: 1:5  104 = 2:7  10


12

@x

1=2

::: 1:1  10

when C = 1. The absolute values of the subtraction of pressure gradient terms vary in the interval of 10 decades wide, which may cause problems. A more detailed analysis of the cell-face velocities containing high damping terms now C = 1, and see table 1. prevails as follows. All three damping terms higher than the central-di erenced cell-face velocity are at the rst cell-face after the boundary. Also the velocity is always relatively small de ned as juj UW =100. 18 high damping terms of the next group 1 10 ... 1 are found near the boundaries three cell-rows beside the boundaries, where the velocity is low 11 except the moving top wall seven, and where the cells are small. The remaining eight faces are in the inner part of the domain and have small velocities and big volumes. In this group seven high central-di erenced cell-face velocities are found near the upper corners where the cells are small, and thus, large damping terms are caused by subtractions of pressure gradient terms. 118 high damping terms of the next group 1 10 ... 1 100 are found near the boundaries, where the velocity is low 93 except the cells near the moving top wall 25. The remaining 36 faces are in the inner part of the domain, and 29 of them have small cell-face velocity. Thus, the number of small velocity-faces is 122. The rest 25 high velocity-faces are in the upper corners of the domain, where subtractions of pressure gradients are high. More detailed statistics of the previous paragraphs are shown in Table 2, where 154 high damping terms of Rhie & Chow interpolation C = 1 are divided into four subgroups columns: near boundary & low velocity, near boundary & high velocity, far from boundary & low velocity, and far from boundary & high velocity. The "near boundary" faces are the rst, second and third cell-faces after 73

the boundary of the domain. The rest are "far from boundary" cell-faces. The limit for velocities is in Table 1. In this particular case, the high damping terms are mainly found near the boundaries, where the velocities are small except the moving top wall, and where, however, the control volumes are small. If a large subtraction of pressure gradient stencils exists, the high damping term occurs in spite of high cell-face velocity and small control volume. This can be recognized in the stagnation region of the upper corner of the right side. And minor e ect. nally, it seems that the term 1 groups of the relative magnitudes of the local damping terms, which were used

UW =100.

The rows 1 100

:::1=10, 1=10:::1 and

1 are the

=a has a

Rhie & Chow C=1.0


= :::1=10 1=10:::1:0 1:0
1 100

near b. low v. 93 11 3

near b. high v. 25 7 0

far from b. low v. 29 8 0

far from b. high v. 7 0 0

terms in the particular test case with 64 v=velocity.

Table 2. The statistics explaining possible reasons for high damping


 64 grid: b=boundary and

The number of high damping terms can be decreased by increasing the grid density, which is demonstrated in Table 3. By using a 128 of large damping terms decreases noticeably 5 

 128 grid, the portion ! 1 .


far from b. high v. 2 0 0

Rhie & Chow C=1.0


= :::1=10 1=10:::1:0 1:0
1 100

near b. low v. 97 11 1

near b. high v. 31 5 0

far from b. low v. 13 2 0

terms in the particular test case with 128

Table 3. The statistics explaining possible reasons for high damping


 128 grid:
b=boundary

and v=velocity. The number of cell faces is 16256.

The damping term also has an e ect on convergence history, which is represented in Fig. 28. If the coe cient

is at least 0.1, 74

L2 -norm of mass residual becomes

machine-accurate within 1 000 iterations. The corresponding numbers of iterations are 1 250 and about 10 000 for C = 0:01 and C = 0 respectively. This does mean that the internal damping of central di erencing is not su ciently large and extra damping is needed. The damped cell-face velocities have a direct e ect on mass balance convergences, but the e ect can no more be recognized in L2 norms of momentum residuals, see Fig. 29: machine accuracy is always obtained within 1 250 iterations without dependence on the damping. During the next test, the damping terms are limited locally. The local relative damping 106 was de ned as the ratio between the local damping term and the local central-di erenced cell-face velocity. Limited damping means that the absolute value of the relative damping term is not allowed to exceed the limiting value. If any excess is observed, only the excess is cut o , i.e. in practise corresponding limiting value is used as a damper, and the original sign is used. Hereafter this will be referred to as the limited damping or the limited damping term. If the limiting value tends to zero, the resulting cell-face velocity stencil becomes a pure central di erence scheme. If the limiting value tends to in nity, the traditional Rhie and Chow interpolation method is obtained. Limited damping has similar e ects as the varied coe cient C had in the previous test. The e ects on the solutions themselves are minor, but can be seen in the histories of the mass residuals, see Fig. 30. If the limit for the relative damping term is not lower than 0.001, the convergence speed is not decreased.

Figure 28. L2 norms of mass residuals during iterations, lid-driven cavity ow of Re = 400.

75

Figure 29. L2

norms of u-momentum residuals during iterations, lid-driven cavity ow of Re = 400.

cavity ow term.

Figure 30. L2 norms of mass residuals during iterations, lid-driven


Re = 400. DAMP LI is local limit for relative damping
76

Next, the damping term was simpli ed SIMP1. Version 84, where the convection term was replaced by 2 Uchar , i.e. the damping term was as follows

@p @p  L C 2 Uchar + Pf aces L @x @x

107

where C = 1, performed like the modi ed Rhie and Chow interpolation 76. The characteristic velocity was UW . In the Rhie and Chow interpolation, the term ai 11=2;j was suspected to cause convergence problems at the beginning of iteration. But by replacing its convection term by 2 Uchar it was not possible to increase the Courant number, i.e. suspicion was probably groundless. Also the detailed statistics of high damping terms were like those shown in Table 2. By decreasing the coe cient C the number of high damping terms could be reduced, but the convergence speed became signi cantly slower when using value C = 0:01. Next, the damping term was further simpli ed SIMP2. The viscous terms were dropped and the convection term was replaced by k Uchar , i.e. it was as follows
k Uchar @x

x

@p

@p @x

108

where Uchar = UW and the coe cient k was varied. With the value k = 1 the damper performed like the modi ed Rhie and Chow interpolation 76. By choosing the value k = 100, the number of high damping terms was decreased signi cantly, and the numbers were 0, 1 and 26 in groups 1, 0:1 ::: 1 and 0:01 ::: 0:1 respectively. If a higher k value was used, the convergence speed became slower, as can be seen from Fig. 31. Hereafter the simpli cation 108 will be referred to as the simpli ed damping or the simpli ed damping term.

77

cavity ow

Figure 31. L2 norms of mass residuals during iterations, lid-driven


Re = 400, damping term version SIMP2.

The damping term in the cell-face velocity formula seems to be necessary to decrease the number of iterations and to prevent the pressure oscillation. The original Rhie & Chow interpolation scheme generates quite high damping terms. The dampers over the whole domain can be multiplied by a constant to reduce high values. But to make sure that the high values are cut o , the locally limited damping should be used.

5.2.2 Re = 10 000
A reference result of Reynolds number 10 000 was calculated by Ghia et al. 68 . They used a uniform grid with 128  128 cells, i.e. the cell size was L=128  L=128 L=128  0:007813L. The boundary layer thickness of the developed laminar ow on a at plate can be written as, 69
L=

5:0 x

Re

109

In this particular case, the boundary layer thickness beside the top wall would be 0:025L, 0:035L and 0:050L at the positions 0:25L, 0:5L and L, respectively. The grid used by Ghia et al. was probably not dense enough. The rst grid used in this study is 128  128. The height of the rst cell adjacent to the wall is L=1000 yielding to an expansion factor 1:053. 78

In Fig.

32, dimensionless

proposed by Johansson and Davidson 76 is used. The damped and non-damped result do not deviate visibly, but they do deviate slightly from Ghia's result. It is expected to be caused by Ghia's uniform grid, which is probably too coarse beside 1 2 the moving top wall. Dimensionless and 2  pro les are shown in Figs. 34 and 35. Damped  = 0 5 and non-damped  = 0 do not 1 2  pro les deviate visibly, and thus only the former one is shown. But 2 deviate from each other. The strongest di erences are found near the walls, where

u=UW

pro les are represented.

The damping term

v=UW

p=

UW

v=UW p= UW

the high relative damping terms occur. Especially in industrial applications, the grid density is quite often too low due to the computational power available. Such a case can be studied by decreasing the number of cells of the domain. Comparison between a damped and a non-damped solution prevails, if a damping term has visible e ect to the The 32 wall is nal solution. rst cell beside the

the vertical midplane are presented in Figs. 33-35. The deviation is visible in all three gures. This does mean that the damping term contributes to the nal solution, if the grid density is low.

C = 1 corresponding to original Rhie and Chow 1 2 interpolation and non-damped C = 0 u=UW , v=UW and p= 2 UW  pro les on
The damped 

L=100 high.

 32 grid with expansion factor 1:138 is adopted. The

Figure 32. Dimensionless velocity pro


plane

les u=UW x=L = 0:5, Re = 10 000, 128  128 grid.

on the vertical mid-

79

Figure 33. Dimensionless velocity pro les u=UW on the vertical mid-

plane x=L = 0:5, Re = 10 000, 32  32 grid.

Figure 34. Dimensionless velocity pro les v=UW on the vertical mid-

plane x=L = 0:5, Re = 10 000, 32  32 grid. 80

midplane x=L = 0:5, Re = 10 000.

1 2 Figure 35. Dimensionless pressure pro les p= 2 UW  on the vertical

The statistics of the high damping terms Table 4, prevail the same as Figs. 33-35. The number of relatively high damping terms is large, about 46  of dampers are at least 1  of the central-di erenced cell-face velocity. 1 100 ... 1 10 1 10 ... 1 1 C=1.0 324 115 16
Table 4. The statistics for the high damping terms, Re = 10 000,

32  32 grid, total number of u-face-velocities is 992.

When the relative damping is limitted to be at most 0.1  of central di erenced cell-face velocity, the mid-plane pressure deviates still visibly from the non-damped solution, Fig. 36. The convergence history of mass residual is quite similar as using Rhie & Chow interpolation C = 1, Fig. 37. An interesting feature is that the limiter value 0.01  leads to the solution similar to without damping but simultaneously ensuring better convergence rate like using Rhie and Chow interpolation. An interesting feature is that no oscillation occurs in pressure pro les, Figs. 35 and 36, even without damping. Compared with the computations of Re = 400 the 81

only di erences are the Reynolds numbers, used grids and driving parameters. The boundaries are handled in the same way in both cases. Based on these computations the feature cannot be explained.

1 2 UW  on the vertical les p= 2 midplane x=L = 0:5, Re = 10 000, 32  32 grid.

Figure 36. Dimensionless pressure pro

cavity ow.

Figure 37.

L2 norms of mass residuals during iterations, lid-driven 82

5.3 Two-dimensional Inviscid Channel Flow


In this test case, the smoothing viscous terms of the momentum equations are ignored to highlight the performance of the pressure correction approach. Flow is allowed to slip on the walls, and therefore boundary layers do not exist beside them. Thus, the high grid-densities beside the walls are not needed, and the and the main interest is focused on the convergence. A schematic description of the case is shown in Fig. 38. For simplicity, dimen3 sional variables are used. The ow of inviscid uid, = 1 000 kg m , entrances the domain from the left side with uniform velocity exiting uid is xed, ow of any velocity remains stable. Furthermore, the solution is known beforehand,

through them. The dimensions of the domain are

L = 10 m and H = 1 m. In the x-direction 24 non-uniform control volumes with expansion factor 1:05 are used. The rst and the last cells are 0:2247 m and 0:6902 m long, respectively. In the y -direction 12 cells of uniform distribution are used. The grid is shown in Fig.

p = 0 Pa.

u = 10 m=s.

The pressure of

The inviscid walls are impermeable, i.e.

v = 0 m= s

39.

y inviscid wall entrance: u = 10 m/s v = 0 m/s fluid: = 1000 kg/m3 = 0 kg/(ms) inviscid wall: du/dy=0 v=0 m/s L = 10 m exit: p= 0 Pa

H=1m

Figure 38.

Schematic description of the inviscid channel

ow.

Figure 39.

Computational grid used in inviscid channel 83

ow.

The nal solution is u = 10 m=s, v = 0 m=s and p = 0 Pa over the whole domain. If it is used as an initial value during the rst iteration, it will be accepted as the nal solution. Therefore, any other initial guess must be used to study convergence rates. If the initial guess is not far from the nal solution, there is no need to x the pressure correction eld, i.e. the boundary condition @p =@n = 0 can be used over the whole boundary of the domain. But if they do deviate from each other much, the xed pressure correction eld must be used to ensure convergence. Otherwise, very strong under-relaxation factors must be used, which cause a very slow convergence rate. In practical applications the initial guess often deviates quite much from the nal solution, and thus only the latter case will be studied.
0

Now the exit pressure is xed, and thus the pressure correction is zero at the face. It is the most natural way to x the level of the pressure correction to prevent the overshoots of the pressure eld ensuring the convergence. The constant initial guess u0 = 0:1 m=s, v0 = 0:1 m=s and p0 = 500 000 Pa has been given over the domain, i.e. the initial u-velocity is 1  of the nal solution, a disturbing transverse component exists and the initial pressure is many magnitudes higher than the nal solution. If any candidate of the proposed cellface velocity stencils maintains its robustness in this particular test case, the method can be assumed to manage in a wide range of applications. The studied stencils are the traditional Rhie and Chow interpolation method without and with limitation, a simpli ed damping term and the pure central di erence formula without damping.
0

The number of variable parameters is huge, and thus this computational experiment cannot cover everything. First, the number of the multigrid levels is 3 due to the grid introduced earlier. Ten V-cycles are always used, and the number of Gauss-Seidel iterations is 2, 3 and 4 on the 1st, 2nd and 3rd grid level respectively. These choices are based on preliminary tests. The under-relaxation factor of the velocities, v , will be changed from value one only if it is necessary to ensure convergence. The new velocity is a linear combination of the old and the new values where the former is weighted by 1 v  and the latter by v , i.e. the de nition is not the same as proposed by Patankar 8 . Patankar uses the under-relaxation of the velocities to strengthen diagonal dominance of discretized momentum equations to ensure convergence. In Patankar's proposal, the corrected velocity eld always satis es the continuity equation. In the present solver, the diagonal dominance is guaranteed by unsteady terms. If v 6= 1, the corrected velocity eld does not necessarily satisfy the continuity equation depending on the initial conditions. If the guessed velocity eld satis es the continuity equation, the mass balance always holds, otherwise after attaining the converged solution. 84

The optimal values of p, in formula 60, and the Courant number t CFL = u x 110

are sought for di erent cell-face velocity stencils. In formula 110, u is a convection velocity and x is the length of the cell, and the formula de nes the time-step t of the momentum equations. In this particular case, the convection velocity has to be xed to avoid convergence problems caused by the initial values, and the characteristic value u = 10 m=s is used. The traditional Rhie and Chow interpolation method needs strong under-relaxation for velocities. Regardless of the under-relaxation of the pressure, the best convergence rate is attained using relatively small Courant number and a relatively high underrelaxation factor, e.g. when p = 0:25 the optimal values are CFL = 3:5 and v = 0:25. These values are found by decreasing the Courant number and increasing the factor v simultaneously. The computation converges also with higher Courant numbers ! 20, but the convergence becomes slower because then the under-relaxation factor of the velocity must be decreased. Typical upper limits of v are 0.06 and 0.034 for Courant numbers 8 and 14 respectively. An interesting feature is that, for all the used p, a local optimum is found using a high Courant number, CFL  15 and v  0:03, but the number of needed iterations is about double compared with the global optimum. In Fig. 40 the convergence rates of the mass residuals of various Courant numbers are presented for p = 0:25. For each Courant number the optimal v is used. In Fig. 41 the e ect of p on convergence rate is presented. Corresponding values of Courant number and v are optimal. Di erences in convergence rates are not signi cant, but the trend is quite clear. The higher the value of p, the better the convergence rate becomes. But in practice, it is quite di cult to nd rapidly converging p, CFL and v values for the traditional Rhie and Chow interpolation method. It can be suspected that the convection components in the damping term cause convergence problems during the rst iterations. In the nal solution the damping terms in the x-direction are very small compared with the central-di erenced cell-face velocities, on the average 2:7  10 6  . In the ydirection relative damping terms are high, on the average 39 , but the e ect on the nal solution is invisible due to very small values of v-velocities  10 6 m=s compared with u-velocities = 10 m=s.

85

Figure 40. Convergence histories of mass residuals of inviscid channel ow using traditional Rhie and Chow interpolation: p = 0:25, various Courant numbers with optimal v .

Figure 41. Convergence histories of mass residuals of inviscid channel ow using traditional Rhie and Chow interpolation: various pvalues with optimal Courant number and v . 86

Next, the performance of the limited Rhie and Chow interpolation method is studied. The practice was described earlier. In brief, the absolute value of the relative damping term 106 is not allowed to exceed the limiting value DAMPLI . If any excess is observed, only the excess is cut o . The limited Rhie and Chow interpolation method does not need under-relaxation of velocities, i.e. value v = 1:0 can be used. The relative damping term has to be smaller than  15 to ensure the convergence. The optimal limiting value is approximately 10, regardless of the pressure under-relaxation factor and Courant number. In Fig. 42 the convergence histories of mass residuals are presented for various limiting values. Other parameters are p = 0:5, v = 1:0 and CFL = 3:0. Up to the limiter value of one, the convergence rate does not deviate much from the optimal choice 10. But if the value of the limiter is decreased below one, the convergence rate becomes clearly slower. Also, too high a value of the limiter makes the convergence slower oscillating, and soon diverging limiter values are achieved,  15. The limited damping terms are needed in the x-direction, i.e. for u-velocity. The y-direction can even be unlimited. The limiter value below one impairs the convergence rate. If the limiter in the y-direction is higher than one, the limiter in the x-direction can be even 0.01 without impairing the convergence.

Figure 42. Convergence histories of mass residuals of inviscid channel ow using the Rhie and Chow interpolation method with limited damping: p = 0:5, CFL = 3 and various limiters of relative damping term.

87

In Fig. 43, the e ect of the Courant number on the convergence rate is shown. The higher the Courant number is, the better the convergence rate becomes. The corresponding values of limiters are optimal ones and the under-relaxation factor of pressure is 0.5 . The optimal Courant number is the upper limit ensuring the convergence. In practice, the same holds for the limiter. The range of converging limiter values of the highest Courant number is always narrow, but it can be found quite easily.

Figure 43.
nel damping: iters.

Convergence histories of mass residuals of inviscid chan-

ow using the Rhie and Chow interpolation method with limited

p = 0 5 and various Courant numbers with optimal lim-

88

In Fig. 44, the e ect of the under-relaxation of pressure on the convergence rate is shown. The higher the under-relaxation factor is, the better the convergence rate becomes. If the value 0 75 is exceeded, the convergence rate becomes no

better. The corresponding values of other parameters are optimal ones, i.e. the upper limits of the Courant number and the limiter. By adopting a limiter into the Rhie and Chow interpolation stencil, the convergence rate can be clearly improved. In this particular case, the number of iterations needed can be decreased approximately to one third. The test computations indicate, that the high damping terms of the traditional Rhie and Chow interpolation cause divergence at the begining of computation. The used limiter simply cuts too high damping terms, and thus the under-relaxation of the velocities can be ignored. The values of the other parameters are similar in the unlimited and the limited cases. The limiter does not see the reason for too high a damping. Limited damping terms are necessary only in the direction where high pressure gradients appear. The practical information about the e ect of the non- xed convection components in the damping term is still unclear.

Figure 44.
nel ing term.

Convergence histories of mass residuals of inviscid chan-

ow using the Rhie and Chow interpolation method with limited

damping: various

p with optimal

CFL and limiter of relative damp-

89

Next, the simpli ed Rhie and Chow interpolation stencil 108 is tested. The optimal value for the coe cient k is to be sought. The smaller the coe cient is, the stronger the damping becomes. As it tends to in nity, the damping term tends to zero. According to the previous tests, the damping must be limited to ensure the convergence. Thus, the xed convection velocity of the damping term must have a lower limit. But too high a velocity decreases the convergence rate. Hereafter, the product kUchar will be called  xed convection velocity. In Fig. 45 the e ect of the xed convection velocity on the convergence rate is demonstrated. The values of the velocity are Uin, 10  Uin and 100  Uin, where Uin is the inlet velocity. Other parameters are p = 0:35, v = 1:0 and CF L = 3, i.e. the under-relaxation of the velocities can be ignored. The trend is similar for other values of p, v and CF L. In general, the bigger the velocity is the slower the convergence becomes, which was expected. In this particular case, the optimal value is approximately the same as the inlet velocity. The computation becomes diverging when the velocity is below  0:85  Uin .

Figure 45. Convergence histories of mass residuals of inviscid channel ow using simpli ed damping: p = 0:35, v = 1:0, CF L = 3 with various convection velocities in damping term.

90

An interesting feature is that the higher the velocity is, the smaller the converging Courant numbers are. The converging values of Courant numbers are CFL 4:26, CFL 3:61, CFL 3:55 and CFL 3:54 for the convection velocities U , 10  U , 100  U and 1000  U , respectively.
in in in in

In Fig. 46 the e ect of the Courant number on the convergence rate is demonstrated. The values of the Courant number used are 2, 3 and 4. Other parameters are = 0:35, = 1:0 and the corresponding optimal value of the convection velocity in damping term. An interesting feature is that the optimal Courant number is not the biggest value. Furthermore, regardless of the values of and the xed velocity, the optimal value of the CFL seems to be about three.
p v p

Figure 46. Convergence histories of mass residuals of inviscid channel ow using simpli ed damping: various Courant numbers with = 0:35, = 1:0 and optimal convection velocity in damping term.
p v

91

In Fig. 47, the e ect of the underrelaxation of pressure on the convergence rate is presented. The higher the under-relaxation factor is, the faster the convergence rate becomes. The higher values of ments in convergence rates are no more signi cant. Courant number of about three. The higher the

p can be used, but the practical improveThe values of the other

parameters are optimal ones, i.e. a low convection velocity in damping term and

CFL

p is, the smaller the converging Courant numbers become, e.g. for p = 0 35 the computation converges using Courant numbers 4 26 and the corresponding region for p = 0 75 is
3 90, where the convection velocity in damping term is

Uin.

It is seen that, by replacing the damping term of the traditional Rhie and Chow interpolation by the simpli ed damping term 108, the convergence rate can be noticeably improved. The number of iterations needed can be decreased almost to one third. The most noticeable change in the converging values of the driving parameters is that the underrelaxation of the velocities can be ignored. It can be expected that the xed coe cient of the di erence of the pressure gradient stencils prevents the overshoots of the damping term and makes the driving parameters more convenient for convergence.

Figure 47.
nel term,

Convergence histories of mass residuals of inviscid chan-

p = 0 35 with optimal Courant number and optimal convection velocity in damping v = 1 0.

ow using simpli ed damping: various values of

92

In Fig.

48, the convergence rates of various damping methods are presented.

The chosen parameters are the optimal ones presented earlier. The convergence rate can be noticeably improved, if the traditional Rhie and Chow interpolation scheme is supplied with the limiter of the damping term, or if the coe ation scheme is replaced by a simpler cient of the di erence of the pressure gradient stencils in the Rhie and Chow interpolxed factor. The former converges three times faster than the traditional method, and the latter is a little bit slower than the former. The di erence is smaller than could have been expected. The limited scheme cuts the huge damping terms locally, but the scheme with the coe xed cient contributes to every cell-face velocity. It seems as if the overshoots of

the damping should be cut away to improve the convergence. But, on the other hand, the damping cannot be ignored without signi cantly impairing the convergence rate. In this particular case, the non-damped method converges to the level 5 1 jj _ jj = 10 kgs approximately within 8 000 iterations. The optimal driving parameters are easily found for both the modi ed versions. But the optimal values for the traditional Rhie and Chow interpolation scheme must be sought carefully. Thus, the modi ed Rhie and Chow interpolation is recommended to be used, if the ow eld is expected to contain large pressure di erences, which cient may cause an uncontrollable rise of coe

83.

Figure 48.
nel

Convergence histories of mass residuals of inviscid chan-

ow: various dampers with optimal parameters.

93

The results themselves do not deviate much from each other. The main emphasis can be put on the quality of the results, see Fig. 49. When using the centraldi erenced cell-face velocities in the colocated grid arrangement, the coupling between the pressure and the velocity elds is weak and produces nonphysical oscillations in the pressure eld. The non-damped pressure in Fig. 49 demonstrates this phenomenon. The pressure oscillates within p  8  8 P a. The amplitude is  0:032  of the dynamic pressure, and thus the oscillation is quite insigni cant. An interesting feature is that the pressure oscillates most near the boundary where the pressure is xed and the velocities are extrapolated; and the oscillation is minor near the extrapolated pressure boundary where the velocities are xed. The rst guess would have been opposite. The traditional Rhie and Chow interpolation stencil damps the oscillation and the pressure decreases quite linearly from 0:325 Pa to 0 Pa in the ow direction, and thus the pressure loss is insigni cant. The pressure elds of the modi ed versions are quite similar. The absolute values are approximately 0:01 Pa, and the distributions refer more to noise than oscillation.

Figure 49. Dimensionless pressure distribution of the inviscid channel ow along the plane y=H = 0:458.

94

5.4 Two-dimensional Inviscid Flow in Diverging Channel


A schematic description of the case is shown in Fig. 50. For simplicity, dimensional variables are used. The ow of inviscid uid, = 1 000 kg=m3, enters the domain from the left side with the uniform velocity distributions U = 10 m=s and V = 0 m=s. The pressure of exiting uid is xed, p = 0 Pa. The walls are impermeable, i.e. boundary condition
in in out

v =0
n n

111
in

holds for them, where v is the normal velocity The length of the domain is L = 10 m. The height of the inlet face is H = 1:051 m and the corresponding value of the outlet is H = 1:745 m. The formula de ning the diverging walls is shown in Fig. 50. The channel is symmetrical with respect to the x-axis. In the x-direction 48 uniform control volumes, x = 0:417 m, are used. In the y -direction 24 cells with uniform distribution are used, i.e. the heights of the cells are 0:0438 m and 0:0727 m at inlet and outlet faces, respectively. The computational grid is shown in Fig. 51.
out

y y(x)=0.5{1.398 + 0.347tanh(0.8x 4)}

entrance: u = 10 m/s v = 0 m/s

exit: p = 0 Pa

fluid: = 1000 kg/m3 = 0 kg/(ms) Hin =1.051 m vn = 0 m/s L = 10 m

Hout =1.745 m

Figure 50. Schematic description of inviscid ow in the diverging channel.

95

Figure 51. Computational grid of inviscid ow in the diverging channel.

Now, four multigrid levels can be used to compute the pressure correction equation. According to the test computations of the multigrid method, the number of grid levels should be at least four to obtain a signi cant reduction in computational costs when calculating a machine-accurate result. But now the multigrid Poisson-solver is working as a unit of the Navier-Stokes solver, and thus it is not necessary to compute the pressure correction equation with machine-accuracy during each outer iteration. In fact, the optimal numbers of V-cycles and LGS iterations on each grid level depend on many things, e.g. the case to be computed, the number of cells, the initial values, the driving parameters, etc. In this particular case, the driving parameters are xed to be CFL = 13:0, p = 0:6, v = 1:0 based on preliminary tests. The initial values for the rst iteration are u0 = 0 m=s, v0 = 0 m=s and p0 = 0 Pa. The parameters of the multigrid solver are to be optimized. The cell-face velocities are calculated using the traditional Rhie and Chow interpolation scheme.

96

Figure 52. Convergence histories of mass residuals. In Fig. 52, the mass residuals are presented as the function of CPU time. By using the optimized MG-parameters, the converged solution is computed during 59 outer iterations. The corresponding number of iterations for the SL-LGS-solver is approximately 200. The optimal values for the multigrid method are 10 V-cycles with 1, 1, 1, and 12 LGS iterations on the 1st, 2nd, 3rd and 4th grid levels, respectively. The number of V-cycles is quite high, but the number of inner iterations on each grid level is only one, except for the coarsest grid level where the computational cost is minor. If the LGS iterations on each grid level are xed to be 2,2,2,12, or 2,3,4,12 the optimal number of V-cycles is 9 and 8 respectively, but the required CPU time is increased approximately 20 . When using a SL-LGS solver the optimal value of the inner iterations is 15. The CPU time used can be reduced to one third by replacing the SL-LGS Poisson-solver by the MG-LGS-solver.

97

Figure 53. Relative pressure eld Pa of the inviscid uid ow in the diverging channel.

Figure 54. Momentum, de ned as ~ v with dimension tours of the inviscid uid ow in the diverging channel.
j j

kg m2 s

, con-

In Figs. 53 and 54 constant pressure and momentum contours of the solution are presented. The total energy is conserved locally within an accuracy of 2.0  without dependence on the damping of cell-face velocities. 98

Figure 55. Inviscid


damper the

uid

ow in the diverging channel: Dimensionless

pressure along the surface

Rhie and Chow interpolation scheme xed convection

DAM P LI velocity is Uin

= 0:042 near the exit.

In the limited

= 1 and in simpli ed

In Fig. 55, the pressure distribution along the surface in section 7:5 m the

= 0:042 m is presented

10 m. The cell-centred pressure distribution is smooth The strongest oscillation is found beside

except for the non-damped version.

xed pressure boundary. The amplitude of the oscillation is very small, 3 dimensionless value 10 , compared with the pressure di erence between the exit and the inlet, dimensionless pressure di erence = 0:64, and it vanishes in the upstream direction. In practice, the oscillation is insigni cant. In Fig. 56, convergence histories of mass residuals with the di erent magnitudes of the damping term are presented. To ensure fast convergence, strong relative damping  1 is necessary for contravariant velocities in the J-direction. Even The transverse the damping terms are quite high compared with the central-di erenced cellface velocities, the e ect on the velocity vectors is invisible. component. With the choice between the solutions. In Fig. 57, the convergence histories of mass residuals with the simpli ed damping term are presented. The smaller the in the xed convection velocity is, the stronger the damping is and the faster the convergence becomes. High damping terms occur nal solution, but their e ect is minor, as discussed earlier. 56 and 57, the convergence rates are quite similar 99 velocity component is always small compared with the corresponding main-stream

DAM P LI

= 10 the solution does not converge to

the same level as with other limiter values. But in practice, there is no deviation

As can be seen from Figs.

for all three methods with optimal parameters. In this test case, the convection velocity of the damper has an optimal value about the inlet velocity as it was in earlier test case. one if the limiter

DAMPLI

The convergence rates do not deviate much from the optimal is at least 1.

Figure 56. Inviscid


ing term,

uid

ow in the diverging channel: Convergence

histories of mass residuals with the di erent magnitudes of the damp-

DAMPLI

= maximum of relative damping term

Figure 57. Inviscid

uid

ow in the diverging channel: Convergence

histories of mass residuals with the simpli ed damping term containg only convection component.

100

5.5 Buoyancy-Driven Cavity Flow


A buoyancy-driven cavity ow, see Fig. 58, is the most commonly used computational experiment for demonstrating the preformance of a thermal Navier-Stokes solver. The enclosure has a hot left vertical wall with uniform temperature Th and a cold right vertical wall with uniform temperature Tc. The height and the width of the enclosure is L. The characteristic number of this problem is the Rayleigh number,
Ra = g

Th

Tc  L3

112

where is volumetric thermal expansion coe cient and is thermal di usivity. The uid properties are evaluated at the mean temperature Tc + Th=2. For vertical plates the critical Rayleigh number is approximately 109, 57 . de Vahl Davis 70 has reported the computational benchmark solutions for air up to Ra = 106. Barakos et al. 71 have reported laminar computations up to Ra = 1011.

dT/dy=0 isolated wall

u=0 v=0

hot wall

T=Th u=0 v=0 L


y,v

cold wall

Tc

T=Tc u=0 v=0

isolated wall x,u

dT/dy=0

u=0 v=0

Figure 58. Schematic description of a buoyancy-driven cavity ow.

101

Firstly, the case Ra = 106 is calculated to con rm the propriety of the implementation. The original Rhie & Chow interpolation scheme is used. The 64  64 grid shown in Fig. 23 is chosen. de Vahl Davis 70 has reported his results in the form shown in Table 5, where only his results of the nest grid uniform 80  80 are given. Also the results of Barakos et al. are presented. They used an 80  80 non-uniform grid. This work Nu 8.835 Numax at y=L 17.63 0.0375 Numin at y=L 0.968 1.00 Umax at y=L 0.077 0.850 Vmax at x=L 0.261 0.0380 de Vahl Davis 8.799 17.93 0.0378 0.989 1.00 0.079 0.850 0.262 0.0380 Barakos et al. 8.806 17.44 0.0368 1.000 1.00 0.077 0.859 0.262 0.0390

Table 5. Comparison of results for Ra = 106.


As can be seen, the present results agree well with the reference results, and the implementation can be expected to be error-free. Dimensionless velocities U and V are de ned as

U=q

u g Th Tc L

113 114

v V =q g Th Tc L

where the denominator is the characteristc velocity of buoyancy-induced ow, and the Nusselt number is de ned as

Nu = T L T  @T h c @x wall

115

and its average value over the wall is Nu. The gradient in de nition 115 is approximated by using a one-sided second-order formula. The driving parameters used are not optimized, and thus convergence is not very rapid. But, on the other hand, convergence problems are not observed. Next, the same grid is used to calculate the case Ra = 108 to nd out if the original Rhie & Chow interpolation scheme has convergence problems with high 102

26 . Again, the computation converges. But, from Fig. 59, it can be seen that the ow tends to behave timedependently. The present, Henkes's 72 and Barakos's averaged Nusselt numbers are 30.7, 30.4 and 30.1, respectively.

Ra values, as reported by Johansson and Davidson

Buoyancy-driven cavity ow upper corner.

Figure 59.

Ra = 108,

ow eld in left

Next, the same grid is used to calculate the case Ra = 109 . Based on convergence histories, the ow eld is time-dependent. Qualitatively the ow eld is relevant. The averaged Nusselt number is 57.5 Barakos et al. 54.4 and Henkes et al. 54.1, and thus the ow eld is also quantitatively quite proper. Residuals do not decrease to the same levels as with smaller Rayleigh numbers, but the computation does not diverge during 30 000 iterations. Thus, convergence problems are not found with the original Rhie & Chow interpolation scheme. In the Ra = 106 case, the number of high relative damping terms of u-face1 1 1 :::1 and 100 ::: 10 velocities are 1, 3, 18 and 162 in the groups 10, 1:::10, 10 respectively, and the total number of faces is 4 032. Most of the high damping terms are found near the boundaries and in low velocity regions. The corresponding values in the y-direction are 2, 52, 265 and 502 in groups 10, 1:::10, 1 1 1 10 :::1 and 100 ::: 10 respectively. Most of the high damping terms are found in low velocity regions, mainly in the middle of the domain. Thus, the number of high 103

damping terms in the main ow direction is approximately 4:5 times higher than in the transverse direction. In the Ra = 108 case with the same grid, the numbers of high relative damping terms are approximately 3:5 times higher, and most of them are found in the middle of the domain. Thus, it seems as if the original Rhie & Chow interpolation scheme does not cause divergence when applied to laminar and near-turbulent body force driven ows. Therefore, there is no motivation to study any body force source term in the face velocity scheme. Furthermore, extra terms in the formulas will in all probability increase the number of high damping terms. Gu 44 and Rahman et al. 49 have expressed that the body force terms in momentum-weighted cellface velocity formulas are necessary to ensure convergence. The nal reason for their conclusions is unclear, e.g. chosen cell-type cell-centred versus cell-vertex and or the treatment of the boundaries and or the details of the implementation approximative factorization versus traditional steady-state treatment etc., may in uence more than usually expected. Now a 32  32 grid with expansion factor 1.14 is adpoted to study convergence. The grid is shown in Fig. 60.

Figure 60.

A 32  32 computational grid. 104

In Fig. 61, the convergence histories of mass residuals are presented. The 6 Rayleigh number is 10 . The original Rhie & Chow interpolation scheme, its limited version, and the simpli ed version converge in a similar way. The convergence of the non-damped version is slower than others. The limiter value is 0.01, and the simpli ed version uses characteristic velocity, denominator of meters of the methods are 113, as the convection term, which in practice is the optimal value. Other driving para-

CFL = 12,

p =

v = 1 and 5 LGS iterations without

MG. The parameters are not carefully optimized. The average Nusselt number of all four methods is 8 89 de Vahl Davis 8.799.

Figure 61.

Buoyancy-driven cavity

ow,

Ra

6 = 10 : Convergence

histories of mass residuals.

The

solution was not found. number was

Ra = 108 case was also calculated with the 32  32 grid, but the stationary :

 32:8  1, which deviates only about 10  from the benchmark

Despite strong time-dependency, the average Nusselt

solution 30 4 of de Vahl Davis.

105

6 Conclusions
In this study, incompressible Navier-Stokes equations have been solved using the SIMPLE pressure correction approach in the cell-centred colocated grid arrangement. The general coordinates have been described using the pure control volume approach with vector notations. Cartesian velocity components have been used as variables in the momentum equations, and the contravariant velocity components have been used to calculate convection uxes. For simplicity, the system The solution method is of equations has been solved in a segregated manner.

based on time integration, and the approximative factorization has been applied. Convection and di usion terms have been modelled using MUSCL and central di erence stencils, respectively. Methods for calculating cell-face velocities have been studied, and an e ective multigrid-based solver for the pressure correction equation Poisson-type has been developed. The single-level =SL line-Gauss-Seidel =LGS solver, the full approximation storage = FAS combined with the LGS smoother, and the multigrid =MG combined with the same smoother have been compared with each other. Crumpton et al. to interpolate the or Wesseling 66 . The simplest method, SL-LGS, is computationally expensive and cannot be used if the number of control volumes is large. The FAS-LGS solver is a more complicated case. When using xed V-cycles, the development of the solution stops quite soon after the start. The problem is that the machine-accurate solutions of the di erent grid levels are not the same. It is natural that after some V-cycles a coarser grid always gives the same "guess" of the next ner grid level, and thus the solution does not develop. Therefore, the FAS technique needs controlled solution cycles that makes the solver vulnerable. It was accomplished in Salminen's study 67 . The computational experiments show that the MG technique combined with the LGS smoother is clearly the most e ective. Also there is no need to tune up the MG-LGS solver. The MG-LGS solver with of grid levels should be at least four. The xed V-cycles is recommended when solving larger Poisson problems. Test calculations indicate that the number fth grid level does not improve the convergence rate much, but the proper shape of the solution is obtained more quickly than by using a four-grid-level scheme, i.e. the 5th grid level enhances the robustness of the MG solver. In practical ow calculations, the grid should be such that as many grid levels as possible could be used. Test computations of case three, inviscid ow in the diverging channel, show The restriction operator used is simpler, and thus less accurate, than that used by 65 or Wesseling 66 . But the applied prolongation operator The idea is to devote more e ort ner grid values from the coarser grid solution. The present is more accurate than those used by them.

choices also lead to a simpler solver than that developed by Crumpton et al. 65

that the MG-LGS pressure correction solver can decrease the total CPU time

106

signi cantly.

The number of grid levels has to be as high as possible.

The

number of V-cycles has to be quite high, and the number of LGS sweeps on each grid level is to be low, except for the coarsest grid. How general this instruction is, must be further studied. The central di erence stencil for cell-face velocities causes in all probability oscillating pressure eld. This choice decreases also the convergence rate of the mass balance compared with the momentum-based interpolation methods, pressureweighted interpolation method =PWIM and momentum-weighted interpolation method MWIM. PWIM is better known as Rhie & Chow interpolation. These two methods have been found to be e ective to prevent pressure oscillation. Even though the methods have di erent names they have the same origin, "staggered momentum equation". In fact, the MWIM can be treated as an intermediate stage toward the PWIM. But it is not well reported if other traces of the PWIM or the MWIM can be identi ed in the nal solution.

In this study, Rhie & Chow interpolation scheme is treated as a sum of a centraldi erenced cell-face velocity and a damping term. The number of high damping terms can be quite large when applying the original Rhie & Chow interpolation scheme for cell-face velocities. It may have a contribution to the when the resolution of the grid is low, as it was seen in the cavity ow simply cuts o nal solution rst case, lid-driven

Re = 10000. A limiter has been introduced to prevent it. The limiter


the excess of the damping locally. It removed the e ect of the nal solution of the previous case, and the convergence rate

damping from the

was not slowed down. In this case, the limiter value for the absolute value of the ratio between a damping term and a central-di erenced cell-face velocity had to be at least 0 001 to prevent a decreased convergence rate. The e ect of poor initial values has been investigated in the inviscid channel ow case. If the initial values are far from the right level, the original Rhie & Chow interpolation scheme has convergence problems. The troubles are caused by high convection terms in the coe cient of the pressure gradient terms. The convergence can be ensured by using weak driving parameters at the cost of CPU time. The limiter can be used to avoid the problems. Also a simpli ed damping term with a xed convection velocity has been introduced. The optimal the damping term is allowed to be 13 limiter value is about 13, i.e.

than central-di erenced cell-face velocity. But in practice, high damping terms are found only in a transverse direction, which has no e ect on the inlet velocity, i.e. the characteristic velocity of the nal solution. The optimal choice for the convection velocity of the simpli ed damper is the ow. Both modi ed versions converge approximately three times faster than the original Rhie & Chow version. The limited version is a little bit faster than the simpli ed one. The case of inviscid ow in the diverging channel has also been used to investigate The limited The original Rhie & Chow

 higher

the convergence rates of two modi ed cell-face velocity schemes. version is slightly faster than the simpli ed one. 107

version has a similar convergence history. The optimal value of the limiter is about 10, and the optimal convection velocity in the simpli ed damper is the inlet velocity. The behaviour of the original Rhie & Chow scheme has been studied in the buoyancy-driven cavity ow case. The test computations have shown that no convergence problems have been experienced. Thus, there is no need to add any body force term to the Rhie & Chow scheme to ensure the convergence. A proper body force term might improve the convergence rate, but it has not been studied during this investigation. All three previous cell-face velocity stencils work in a similar way. The characteristic velocity can again be used as the convection term of the simpli ed damper. In this study, only laminar and inviscid test cases have been calculated. For more general conclusions, the cell-face velocity stencils should be tested in threedimensional and turbulent ows.

108

References
1 Salonen, E.-M., Mekaniikan Kasitteita ja Kaavoja," Helsinki University of Technology, Laboratory of Mechanics, Espoo, Finland, 1987. in Finnish. 2 Ferziger, J. and Peric, M., Computational Methods for Fluid Dynamics. Berlin: Springer-Verlag, 1996. ISBN 3 540 59434 5. 3 Chorin, A., A Numerical Method for Solving Incompressible Viscous Flow Problems," Journal of Computational Physics, Vol. 2, 1967, pp. 12 26. 4 Kettleborough, C., Husain, S., and Prakash, C., Solution of Fluid Flow Problems with the Vorticity- Streamfunction formulation and the ControlVolume-Based Finite-Element Method," Numerical Heat Transfer, Part B, Vol. 16, 1989, pp. 31 58. 5 Rahman, M., Numerical Study of Natural Convection from a Vertical Surface Due to Combined Buoyancies," Numerical Heat Transfer, Part A, Vol. 28, 1995, pp. 409 429. 6 Harlow, F. and Welsh, J., Numerical Calculation of Time Dependent Viscous Incompressible Flow with Free Surface," Physics of Fluids, Vol. 8, 1965, pp. 2182 2189. 7 Patankar, S. and Spalding, D., A Calculation Procedure for Heat, Mass and Momentum Transfer in Three-Dimensional Parabolic Flows," Int. J. of Heat and Mass Transfer, Vol. 15, 1972, pp. 1787 1806. 8 Patankar, S., Numerical Hemisphere, 1980.

Heat Transfer and Fluid Flow.

Washington, D.C:

9 Doormaal, J. and Raithby, G., Enhancements of the SIMPLE Method for Predicting Incompressible Fluid Flows," Numerical Heat Transfer, Vol. 7, 1984, pp. 147 163. 10 Melaaen, M., Calculation of Fluid Flow with Staggered and Nonstaggered Curvilinear Nonorthogonal Grids - the Theory," Numerical Heat Transfer Part B, Vol. 21, 1992, pp. 1 19. 11 Melaaen, M., Calculation of Fluid Flow with Staggered and Nonstaggered Curvilinear Nonorthogonal Grids - a Comparison," Numerical Heat Transfer Part B, Vol. 21, 1992, pp. 21 39. 12 Kobayashi, M. and Pereira, C., Calculation of Incompressible Laminar Flows on a Nonstaggered, Nonorthogonal Grid," Numerical Heat Transfer, Part B, Vol. 19, 1991, pp. 243 262. 13 Peric, M., Kessler, R., and Scheuerer, G., Comparison of Finite-Volume Numerical Methods with Staggered and Colocated Grids," Computers and Fluids, Vol. 16, 1988, pp. 389 403. 109

14 Peric, M., Analysis of Pressure-Velocity Coupling on Nonorthogonal grids," Numerical Heat Transfer Part B, Vol. 17, 1990, pp. 63 82. 15 Lapworth, B. L., Examination of Pressure Oscillations Arising in the Computation of Cascade Flow Using a Boundary-Fitted Co-ordinate System," International Journal for Numerical Methods in Fluids, Vol. 8, 1988, pp. 387 404. 16 Jessee, J. and Fiveland, W., A Cell Vertex Algorithm for the Incompressible Navier-Stokes Equations on Non-orthogonal Grids," International Journal for Numerical Methods in Fluids, Vol. 23, 1996, pp. 271 293. 17 Choi, S., Nam, H., Lee, Y., and Cho, M., An E cient Three-Dimensional Calculation Procedure for Incompressible Flows in Complex Geometries," Numerical Heat Transfer Part B, Vol. 23, 1993, pp. 387 400. 18 Jeng, N. and Liou, Y., On the Open Boundary Condition for the SIMPLE algorithm Using Nonstaggered Grids," Numerical Heat Transfer Part B, Vol. 27, 1995, pp. 23 42. 19 Rhie, C. M. and Chow, W. L., Numerical Study of the Turbulent Flow Past an Airfoil with Trailing Edge Separation," AIAA Journal, Vol. 21, November 1983, pp. 1525 1532. 20 Mathur, S. and Murthy, J., A Pressure-Based Method for Unstructured Meshes," Numerical Heat Transfer Part B, Vol. 31, 1997, pp. 195 215. 21 Thiart, G., Finite Di erence Scheme for the Numerical Solution of Fluid Flow and Heat Transfer Problems on Nonstaggered Grids," Numerical Heat Transfer, Part B, Vol. 17, 1990, pp. 43 63. 22 Thiart, G., Improved Finite-Di erence Scheme for the Solution of Convection-Di usion Problems with the SIMPLEN Algorithm," Numerical Heat Transfer, Part B, Vol. 18, 1990, pp. 81 95. 23 Tamamidis, P. and Assanis, D. N., Three-Dimensional Incompressible Flow Calculations with Alternative Discretization Schemes," Numerical Heat Transfer Part B, Vol. 24, 1993, pp. 57 76. 24 Tamamidis, P., Zhang, G., and Assanis, D. N., Comparison of PressureBased and Arti cial Compressibility Methods for Solving 3D Steady Incompressible Viscous Flows," Journal of Computational Physics, Vol. 124, 1996, pp. 1 13. 25 Rodi, W., Majumdar, S., and Schnung, B., Finite Volume Methods for TwoDimensional Incompressible Flows with Complex Boundaries," Computer Methods in Applied Mechanics and Engineering, Vol. 75, 1989, pp. 369 392. 26 Johansson, P. and Davidson, L., Modi ed Collocated SIMPLEC Algorithm Applied to Buoyancy-a ected Turbulent Flow Using a Multigrid Procedure," Numerical Heat Transfer Part B, Vol. 28, 1995, pp. 39 57. 110

27 Miller, T. F. and Schmidt, F. W., Use of Pressure-Weighted Interpolation Method for the Solution of the Incompressible Navier-Stokes Equations on a Nonstaggered Grid System," Numerical Heat Transfer, Vol. 14, 1988, pp. 213 33. 28 Date, A., Solution of Navier-Stokes Equations on Non-Staggered Grid," Int. J. Heat Mass Transfer, Vol. 36, No. 7, 1993, pp. 1913 1922. 29 Date, A., Complete Pressure Correction Algorithm for Solution of Incompressible Navier-Stokes Equations on a Nonstaggered Grid," Numerical Heat Transfer, Part B, Vol. 29, 1996, pp. 441 458. 30 Lien, F. and Leschziner, M., A General Non-orthogonal Collocated Finite Volume Algorithm for Turbulent Flow at all Speeds Incorporating SecondMoment Turbulence-Transport Closure, Part 1: Computational Implementation," Computer Methods in Applied Mechanics and Engineering, Vol. 114, 1994, pp. 123 148. 31 Jang, D., Jetli, R., and Acharya, S., Comparison of the PISO, SIMPLER, and SIMPLEC Algorithms for the Treatment of the Pressure-Velocity Coupling in Steady Flow Problems," Numerical Heat Transfer, Vol. 10, 1986, pp. 209 228. 32 Williams, M., Methods for Calculating Incompressible Viscous Flows," Numerical Heat Transfer Part B, Vol. 20, 1991, pp. 241 253. 33 Abdallah, S., Numerical Solutions for the Incompressible Navier-Stokes Equations in Primitive Variables Using Nonstaggered Grid I," J. Comp. Physics, Vol. 70, 1987, pp. 182 192. 34 Abdallah, S., Numerical Solutions for the Incompressible Navier-Stokes Equations in Primitive Variables Using Nonstaggered Grid II," J. Comp. Physics, Vol. 70, 1987, pp. 193 202. 35 Sotiropoulos, F. and Abdallah, S., The Discrete Continuity Equation in Primitive Variable Solutions of Incompressible Flow," Journal of Computational Physics, Vol. 95, 1991, pp. 212 227. 36 Sotiropoulos, F. and Abdallah, S., A Primitive Variable Method for the Solution of Three-Dimensional Incompressible Viscous Flow," Journal of Computational Physics, Vol. 103, 1992, pp. 336 349. 37 Sotiropoulos, F., Kim, W., and Patel, V., A Computational Comparison of Two Incompressible Navier-Stokes Solvers in Three-Dimensional Laminar Flows," Computers and Fluids, Vol. 23, No. 4, 1994, pp. 627 646. 38 Aksoy, H. and Chen, C.-J., Numerical Solution of Navier-Stokes Equations with Nonstaggered Grids Using Finite Analytic Method," Numerical Heat Transfer, Part B, Vol. 21, 1992, pp. 287 306. 111

39 Kim, J. and Moin, P., Application of a Fractional Step Method to Incompressible Navier-Stokes Equations," Journal of Computational Physics, Vol. 59, 1985, pp. 308 323. 40 Atkinson, K., An Introduction to Numerical Analysis. New York: John Wiley and Sons, 2nd ed., 1982. ISBN 0 471 50023 2. 41 Wesseling, P., An Introduction to Multigrid Methods. New York: John Wiley and Sons, 1991. ISBN 0 486 65251 3. 42 Acharya, S. and Moukalled, F., Improvements to Incompressible Flow Calculation on a Nonstaggered Curvilinear Grid," Numerical Heat Transfer, Part B, Vol. 15, 1989, pp. 131 152. 43 Wu, J. and Rath, H., Finite-Di erence Method of Incompressible Flows with Rotation and Moving Boundary in a Nonstaggered Grid," Numerical Heat Transfer Part B, Vol. 26, 1994, pp. 189 206. 44 Gu, C., Computation of Flows with Large Body Forces," in Numerical Methods in Laminar and Turbulent Flow: vol. VII Part 2, Swansea U.K., 1991. 45 Parameswaran, S., Srinivasan, A., and Sun, R., Numerical Aerodynamic Simulation of Steady and Transient Flows around Two-Dimensional Blu Bodies Using the Nonstaggered Grid System," Numerical Heat Transfer, Part A, Vol. 21, 1992, pp. 443 461. 46 Majumdar, S., Role of Underrelaxation in Momentum Interpolation for Calculation of Flow with Nonstaggered Grids," Numerical Heat Transfer, Vol. 13, 1988, pp. 125 132. 47 Zhang, G., Assanis, D., and Tamamidis, P., Segregated Prediction of 3-D Compressible Subsonic Fluid Flows using Collocated Grids," Numerical Heat Transfer Part A, Vol. 29, 1996, pp. 757 775. 48 Han, T., Computational Analysis of Three-Dimensional Turbulent Flow around a Blu Body in Ground Proximity," AIAA Journal, Vol. 27, No. 9, 1989, pp. 1213 1219. 49 Rahman, M., Miettinen, A., and Siikonen, T., Modi ed Simple Formulation on a Collocated Grid with an Assessment of the Simpli ed QUICK Scheme," Numerical Heat Transfer, Part B, Vol. 30, 1996, pp. 291 314. 50 Russell, P. and Abdallah, S., Dilation-Free Solutions for the Incompressible Flow Equations on Nonstaggered Grids," AIAA Journal, Vol. 35, No. 3, 1997, pp. 585 586. 51 Reggio, M. and Camarero, R., Numerical Solution Procedure for Viscous Incompressible Flows," Numerical Heat Transfer, Vol. 10, 1986, pp. 131 146. 112

52 Rincon, J. and Elder, R., A High-Resolution Pressure-Based Method for Compressible Flows," Computers and Fluids, Vol. 26, No. 3, 1997, pp. 217 231. 53 Spalding, D., A Novel Finite-Di erence Formulation for Di erential Expressions Involving Both First and Second Derivatives," Int. J. Num. Methods Eng., Vol. 4, 1972, pp. 551 559. 54 Minkowycz, W., Sparrow, E., Schneider, G., and Pletcher, R., Handbook of Numerical Heat Transfer. New York: John Wiley & Sons Inc., 1988. ISBN 0 471 83093 3. 55 Leonard, B., A Stable and Accurate Convective Modelling Procedure Based on Quadratic Upstream Interpolation," Comput. Meth. Appl. Mech. Eng., Vol. 19, 1979, pp. 59 98. 56 Hirsch, C., Computational Methods for Inviscid and Viscous Flows, Vol. 2 of Numerical Computation of Internal and External Flows. Chichester: John Wiley & Sons Ltd, 1990. ISBN 0 471 92351 6. 57 Incropera, F. and DeWitt, D., Fundamentals of Heat and Mass Transfer, 2. ed. New York: John Wiley & Sons Ltd, 1985. ISBN 0 471 82561 1. 58 Adams, R., Calculus, A Complete Course, 3. ed. Don Mills, Ontario, Canada: Addison-Wesley Publishers Limited, 1995. ISBN 0 201 82823 5. 59 Siikonen, T., An Application of Roe's Flux-Di erence Splitting for the k Turbulence Model," International Journal for Numerical Methods in Fluids, Vol. 21, 1995, pp. 1017 1039. 60 Siikonen, T., Laskennallisen virtausmekaniikan perusteet," Helsinki University of Technology, Laboratory of Aerodynamics, Espoo, Finland, 1991. in Finnish. 61 Anderson, D., Tannehill, J., and Pletcher, R., Computational Fluid Mechanics and Heat Transfer. New York: Hemisphere Publishing Corporation, 1984. ISBN 0 89116 471 5. 62 Zachmanoglou, E. and Thoe, D., Introduction to Partial Di erential Equations with Applications. New York: Dover Publications, Inc., 1986. ISBN 0 486 65251 3. 63 Peric, M., Numerical Methods for Computing Turbulent Flows," in An Introduction to the Modeling of Turbulence, von Karman Institute for Fluid Dynamics Lecture Series 1997, 1997. 64 Brandt, A., Multigrid Techniques: 1984 Guide with Applications to Fluid Dynamics," Gesellschaft fur Mathematik und Datenverarbeitung, Bonn, 1984. 113

65 Crumpton, P., Shaw, G., and Ware, A., Discretization and Multigrid Solution of Elliptic Equations with Mixed Derivative Terms and Strongly Discontinuous Coe cients," Journal of Computational Physics, Vol. 116, 1995, pp. 343 358. 66 Wesseling, P., Cell-Centered Multigrid for Interface Problems," Journal of Computational Physics, Vol. 79, 1988, pp. 85 91. 67 Salminen, H., Osittaisdi erentiaaliyhtaloiden diskretoinnissa syntyvien lineaaristen yhtaloryhmien iteratiiviset ratkaisumenetelmat," Matematiikan erikoistyo, 1995, pp. 33 + Appendices in Finnish. 68 Ghia, U., Ghia, K., and Shin, C., High-Re Solutions for Incompressible Flow Using the Navier-Stokes Equations and a Multigrid Method," Journal of Computational Physics, Vol. 48, 1982, pp. 387 411. 69 Schlichting, H., Boundary Layer Theory. New York: McGraw Hill Book Co., 7th ed., 1979. ISBN 0 07 055334 3. 70 de Vahl Davis, G., Natural Convection of Air in a Square Cavity: a Bench Mark Numerical Solution," International Journal for Numerical Methods in Fluids, Vol. 3, 1983, pp. 249 264. 71 Barakos, G., Mitsoulis, E., and Assimacopoulos, D., Natural Convection Flow in a Square Cavity Revisited: Laminar and Turbulent Models with Wall Functions," International Journal for Numerical Methods in Fluids, Vol. 18, 1994, pp. 695 719. 72 Henkes, R., van der Vlugt, F., and Hoogendoorn, C., Natural Convection Flow in a Square Cavity Calculated with Low-Reynolds-Number Turbulence Model," International Journal of Heat and Mass Transfer, Vol. 34, 1991, pp. 1543 1557.

114

You might also like