Sie sind auf Seite 1von 13

Computers & Fluids 42 (2011) 13–25

Contents lists available at ScienceDirect

Computers & Fluids


j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / c o m p fl u i d

An efficient solver for the RANS equations and a one-equation turbulence model
R.C. Swanson ⇑,1, C.-C. Rossow
DLR, Deutsches Zentrum für Luft- und Raumfahrt, Lilienthalplatz 7, D-38108 Braunschweig, Germany

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

Article history: A three-stage Runge-Kutta (RK) scheme with multigrid and an implicit preconditioner has been shown to
Received 23 October 2009 be an effective solver for the fluid dynamic equations. Using the algebraic turbulence model of Baldwin
Received in revised form 13 July 2010 and Lomax, this scheme has been used to solve the compressible Reynolds-averaged Navier–Stokes
Accepted 19 October 2010
(RANS) equations for transonic and low-speed flows. In this paper we focus on the convergence of the
Available online 28 October 2010
RK/Implicit scheme when the effects of turbulence are represented by the one-equation model of Spalart
and Allmaras. With the present scheme the RANS equations and the partial differential equation of the
Keywords:
turbulence model are solved in a loosely coupled manner. This approach allows the convergence behavior
Runge-Kutta
Implicit preconditioner
of each system to be examined. Point symmetric Gauss-Seidel supplemented with local line relaxation is
Multigrid used to approximate the inverse of the implicit operator of the RANS solver. To solve the turbulence equa-
Reynolds-averaged Navier–Stokes equations tion we consider three alternative methods: diagonally dominant alternating direction implicit (DDADI),
SA turbulence model symmetric line Gauss-Seidel (SLGS), and a two-stage RK scheme with implicit preconditioning. Compu-
tational results are presented for airfoil flows, and comparisons are made with experimental data. We
demonstrate that the two-dimensional RANS equations and a transport-type equation for turbulence
modeling can be efficiently solved with an indirectly coupled algorithm that uses RK/Implicit schemes.
Ó 2010 Elsevier Ltd. All rights reserved.

1. Introduction When developing an improved numerical method for solving


the Reynolds-averaged Navier–Stokes (RANS) equations, a neces-
Reliable and sufficient convergence for steady-state computa- sary consideration is the coupling of the RANS equations and the
tions of turbulent flows continues to be a challenge in computa- equation or equations of the turbulence model being applied. If
tional fluid dynamics. Here sufficient convergence means that the both the fluid dynamic and turbulence equations are directly cou-
residuals of the fluid dynamic equations and the equation set of pled, then the characterization of the discrete system can change.
a turbulence model are reduced to the level of the truncation error That is, with appropriate discretization the fluid dynamic equa-
of the numerical scheme. In many applications a turbulence model tions are positive definite (sometimes called a vector positive sys-
has one or more partial differential equations (PDEs) which have a tem [1]), making them amenable to relaxation, but the directly
transport form and represent the effects of turbulence on the flow. coupled system may not be, due to the equation set for the turbu-
When solving the transport-type equations of turbulence models, lence model [2]. The numerical stiffness of the entire system is also
either directly or indirectly coupled to the flow equations, the much higher due to the source terms of the turbulence model. An
residuals are frequently reduced only two orders of magnitude. alternative is to use indirect coupling of the two equation sets.
In addition, the poor convergence of these transport-type equa- Generally, in an iterative solution process with this approach the
tions adversely affects the convergence of the flow equations. Of flow variables are updated while the turbulence variables are fro-
course, when adequate convergence is not achieved, there is no zen; and then, the turbulence variables are updated while the flow
assurance that the results obtained represent an acceptable variables are treated as fixed quantities. Strategies for implement-
approximation of the solution even from an engineering perspec- ing indirect coupling depend on the algorithm being used. For
tive. Thus, there is a strong need for improved numerical methods example, when applying multigrid methods, there are two princi-
for not only obtaining steady-state solutions but also unsteady pal strategies for indirect coupling. The first approach [3,4] is to
solutions when using a dual time-stepping scheme. solve the mean flow and turbulence equations in sequence on each
grid level. This can augment the coupling effects, which may be
beneficial for certain types of problems. The second strategy [5,6]
⇑ Corresponding author.
is to use the eddy viscosity determined on the finest grid on all
E-mail addresses: r.c.swanson10@gmail.com (R.C. Swanson), cord.rossow@dlr.de
(C.-C. Rossow).
coarser grids in the multigrid procedure for the mean flow
1
Corresponding author was visiting scientist at the Center for Computer Applica- equations. Then, the turbulence model equation set is solved with
tions in AeroSpace Science and Engineering (C2A2S2E), DLR, Braunschweig, Germany. multigrid or another algorithm. This method can provide an

0045-7930/$ - see front matter Ó 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compfluid.2010.10.010
14 R.C. Swanson, C.-C. Rossow / Computers & Fluids 42 (2011) 13–25

advantage in flexibility for independent evaluations of the schemes 2. Spalart–Allmaras turbulence model
for solving the two equation sets. In addition, since there is latitude
in choosing the iterative scheme and update procedure for the tur- Here we provide a sufficient description of the SA model to al-
bulence quantities, one can also investigate methods to enhance low implementation. A detail discussion explaining the modeling
the coupling effects. of the physical terms in the single transport-type equation is given
By indirectly coupling the equations one can focus on the in the paper by Spalart and Allmaras [10]. Let mt be the eddy viscos-
specific properties of each equation set to obtain the best possible ity, which is defined by
convergence of the two systems of equations. Furthermore, the
essential properties of an algorithm for efficiently solving the di- v3 m~
mt ¼ m~fv 1 ; f v 1 ¼ ; v ; ð2:1Þ
rectly coupled system can be identified. There are common design v þ Cv1
3 3 m
criteria for the algorithms of both equation sets. These require-
ments are as follows: (1) high Courant–Friedrichs–Lewy (CFL) lim- where m is the kinematic viscosity. The transport-type equation for
it, (2) convergence with weak dependency on mesh density, (3) e
m given in Ref. [10] is written as
suitable for stiff discrete systems. In addition, for the equation
@m
~ @m
~
set of the turbulence model there must also be appropriate treat- þ uj ¼ C b1 ð1  ft2 Þe
Sm~
ment of any source terms so that convergence is not adversely @t @xj
   
affected. 1 @ @m
~ @m
~ @m~
þ ðm þ m~Þ þ C b2
A candidate for the flow solver of the loosely coupled system is r @xj @xj @xj @xj
an RK/Implicit scheme with three stages and three evaluations of   2
C b1 m~
the numerical dissipation. Subsequently, this scheme is designated  C w1 fw  2 ft2 þ S; ð2:2Þ
as the RK3/Implicit scheme. Previously, Rossow [7] and Swanson j d
et al. [8] demonstrated that fast convergence can be obtained for
where t is time, xj and uj are Cartesian coordinates and velocity
both the two-dimensional (2-D) and three dimensional (3-D) RANS
components, respectively, and
equations with the RK3/Implicit scheme with multigrid when
using the Baldwin-Lomax (BL) algebraic eddy viscosity model [9]. m~ v
e
S Sþ fv 2 ; f v2 ¼ 1  ; ð2:3Þ
Although there is some slowdown in the convergence rate of the 2 d2 1 þ vfv 1
j
3-D scheme relative to the 2-D scheme, Swanson et al. demon-
strate that this scheme is more than 10 times faster than a well- with S being the magnitude of the vorticity (jXj), d delineating the
tuned standard RK scheme with scalar implicit residual smoothing distance to the closest wall boundary, and j denoting the von
and multigrid. The underlying three-stage RK scheme of this algo- Kármán constant. The first, second, and third terms on the right-
rithm is important for clustering of the eigenvalues associated with hand side of Eq. (2.2) represent the production, diffusion, and
the error components of the iterative process. Preconditioning with destruction terms, respectively. The last term is a source term,
a fully implicit operator, which allows a CFL number of 1000, treats which is defined by
the discrete stiffness problem associated with viscous-layer resolu-
tion. The operator of the discrete implicit system can be approxi- S ¼ ft1 DU 2 ; ð2:4Þ
mately inverted with symmetric Gauss-Seidel (SGS).
The main purpose of this work was to initiate an effort to satisfy where DU is the norm of the difference between the velocity at the
the need to significantly augment the effectiveness (as measured transition location and that at a field point being considered. The
by reliability and efficiency) of algorithms for solving the RANS function fw in Eq. (2.2) is given by
equations and the PDEs of turbulence models. Since establishing !1=6
a highly effective scheme in two dimensions is a prerequisite for 1 þ C 6w3
fw ¼ g ; ð2:5Þ
constructing an efficient scheme in three dimensions, the focus g 6 þ C 6w3
of the present effort is on a 2-D scheme. Moreover, we assess the
performance of an efficient RANS solver (i.e., RK3/Implicit scheme where g and r are defined by
with multigrid) when the turbulent viscosity field is generated by m~
solving a transport-type equation. g ¼ r þ C w2 ðr6  rÞ; r : ð2:6Þ
e
S j2 d
2
To represent the effects of turbulence we use the Spalart–Allm-
aras (SA) model, which is a transport-type equation model that is For large values of r the function fw goes to a constant, and a value of
frequently used in solving a variety of fluid dynamics problems. 10 is appropriate. The function ft2 is defined as
In the first section of this paper this turbulence model is described, ft2 ¼ C t3 expðC t4 v2 Þ: ð2:7Þ
and the specifics of its implementation are given. Next the numer-
ical schemes for solving the mean flow and turbulence equations Spalart includes the transition function given by
are presented and discussed. Modifications of the RK3/Implicit  
x2 2 2
scheme that have produced improved efficiency and robustness ft1 ¼ C t1 g t exp C t2 t 2 ðd þ g 2t dt Þ ; ð2:8Þ
are emphasized. Then three approaches for solving the SA equation DU
are considered. These methods are as follows: diagonally dominant where dt is the distance from the field point to the boundary-layer
alternating direction implicit (DDADI), symmetric line Gauss-Sei- trip (where trip refers to a known location for transition), xt is the
del (SLGS), and a two-stage RK scheme (RK2/Implicit) with implicit wall vorticity at the trip,
preconditioning. In the results section the convergence behavior of
g t  min ½0:1; DU=ðxt Dxt Þ: ð2:9Þ
the methods for solving the RANS equations and the SA equation is
examined. The effectiveness of the loosely coupled algorithm at and Dxt is the grid spacing along the wall at the trip.
high Reynolds numbers and low Mach number is presented. Rapid In the present implementation of the model we do not include
evolution of global quantities such as lift and drag coefficients is the trip function, which is usually neglected when applying the
demonstrated. Furthermore, we show that the convergence of model (e.g., Ref. [11]). In addition, for the purposes of grouping
the RK3/Implicit scheme with the SA model is similar to that ob- terms similar in form and numerical implementation, we rewrite
tained with the BL model. (after some algebra and rearranging of terms), Eq. (2.2) as
R.C. Swanson, C.-C. Rossow / Computers & Fluids 42 (2011) 13–25 15

@m
~ @m
~  To extend the support of the difference scheme we consider
þ uj ¼ SP1 ðm
~Þ þ SP2 ðm
~Þ þ SD ðm
~Þ þ Dðm
~Þ; ð2:10Þ
@t @xj implicit residual smoothing. Applying the smoothing technique
of Ref. [15] we have the following:
where SP1 and SP2 are the two contributions to the production term,
SD is the destruction term, and D is the diffusion term, and they are Li dWðqÞ ¼ dWðqÞ ; ð3:3Þ
given by
where Li is an implicit operator. By approximately inverting the
SP 1 ð m
~Þ ¼ C b1 ð1  ft2 ÞjXjm
~; operator Li we obtain

 2 Dt Dt X ðq1Þ

m
~ dWðqÞ ¼ aq PL; Wðq1Þ ¼ aq P F S; ð3:4Þ
SP 2 ð m ~Þ ¼ C b1 ½ð1  ft2 Þfv 2 þ ft2 j2  C w1 fw
~Þ þ SD ðm ; V V all faces n
d
where P is a preconditioner defined by the approximate inverse
  e 1 , Fn is the normal flux density vector at the cell face, and S is
1 @ @m
~ C b2 @ 2 m
~ L i
Dðm
~Þ ¼ ðm þ ð1 þ C b2 Þm
~Þ  m~ 2 : the area of the cell face. The change dWðqÞ replaces the explicit
r @xj @xj r @xj
update appearing in Eq. (3.1). Thus, each stage in the RK scheme
The constants of the model are as follows: is preconditioned by an implicit operator.
A first-order upwind approximation based on the Roe scheme is
2
C b1 ¼ 0:1355; r ¼ ; C b2 ¼ 0:622; j ¼ 0:41; ð2:11Þ used for the convective derivatives in the implicit operator. To
3
derive this operator one treats the spatial discretization terms in
C b1 1 þ C b2
C w1 ¼ þ ; the flow equations implicitly and applies linearization. For a
j r detailed derivation see Rossow [7]. Substituting for the implicit
C w2 ¼ 0:3; C w3 ¼ 2; C v 1 ¼ 7:1; C t1 ¼ 1; C t2 ¼ 2; operator in Eq. (3.3), we obtain for the qth stage of the RK scheme
C t3 ¼ 1:2; C t4 ¼ 0:5: " #
Dt X Dt X ðq1Þ b ðq1Þ ;
On a solid boundary m ~ ¼ 0. Originally, the free-stream m
~ was set to Iþe An S dWðqÞ ¼ aq F S¼R ð3:5Þ
V all faces V all faces n
1.342m1, where m1 is the free-stream kinematic viscosity. In order
to avoid the possibility of a delayed transition, the free-stream va- where the matrix An is the flux Jacobian associated with Fn at a cell
lue of m
~ is set to 3m1, as suggested by Rumsey [12]. b ðq1Þ represents the residual function for the (q1)th stage,
face, R
and e is an implicit parameter, which will be defined later in this
3. Numerical schemes section.
The matrix An can be decomposed into Aþ 
n and An , which are
To solve the two-dimensional RANS equations we use the RK3/ associated with the positive and negative eigenvalues of An and
Implicit scheme. Complete details of the scheme are presented in defined by
the papers of Rossow [7] and Swanson et al. [8]. The SA turbulence 1 1
model requires the solution of one transport-type equation. In the Aþn ¼ ðAn þ jAn jÞ; An ¼ ðAn  jAn jÞ: ð3:6Þ
2 2
present work we do not directly couple the solution of the fluid dy-
namic equations with the additional equation of the turbulence If we substitute for An in Eq. (3.5) using the definitions of Eq. (3.6),
model. To solve the transport-type equation of the SA turbulence then the implicit scheme can be written as
model we consider the DDADI, SLGS, and RK2/Implicit schemes. " #
In the first part of this section the essential elements of the RK3/
Dt X þ ðqÞ b ðq1Þ  e Dt
X  ðqÞ
Iþe A S dWi;j ¼ R A dWNB S; ð3:7Þ
Implicit scheme are presented. Recent enhancements of the origi- V all faces n i;j
V all faces n
nal RK3/Implicit scheme are also introduced. Then the three meth-
where the indices (i, j) indicate the cell of interest, and NB refers to
ods considered for solving the SA equation are described.
all the direct neighbors of the cell being considered.
To solve the implicit system of Eq. (3.7) for the changes in con-
3.1. RK/implicit scheme ðqÞ
servative variables dWi;j , the implicit operator must be inverted. It
is sufficient to approximate the inverse of the implicit operator.
We apply a finite-volume approach to discretize the fluid
Based on analysis and numerical testing, an adequate approximate
dynamic equations and use the approximate Riemann solver of
inverse is obtained with two pointwise symmetric Gauss-Seidel
Roe [13] to obtain a second-order discretization of the convective
(SGS) sweeps, with each complete sweep followed by one local
terms. The viscous terms are discretized with a second-order cen-
(boundary layer and near wake) symmetric line sweep. Previously
tral difference approximation. To obtain an explicit update to the
[16], we have demonstrated that line relaxation can be used to effi-
solution vector for the flow equations we use a three-stage RK
ciently approximate the inverse. By using local line relaxation, we
scheme. The update for the qth stage of the RK scheme is given by
make the scheme amenable to application on unstructured grids.
WðqÞ ¼ Wð0Þ þ dWðqÞ ; ð3:1Þ To initialize the iterative process the unknowns are set to zero.
The choice of the implicit parameter e in Eq. (3.5) affects the
where the change in the solution vector W is high-frequency damping of the scheme, and thus, its effectiveness
Dt as a smoother for multigrid. Fig. 1a shows the amplification factor g
dWðqÞ ¼ WðqÞ  Wð0Þ ¼ aq LWðq1Þ ; ð3:2Þ of the original RK3/Implicit scheme [7,8] with variation of e. The g
V
was determined with the one-dimensional Fourier analysis of
and L is the complete difference operator for the system of equa- Swanson et al. [8] and is a function of the phase angle hx, which
tions. Here aq is the RK coefficient of the qth stage, Dt is the time is proportional to frequency. At e = 0.5 there is a significant in-
step, and V is the volume of the mesh cell being considered. For crease in g at the highest frequencies, and when e = 0.4, the scheme
the three-stage scheme we use the coefficients is unstable. Thus, for the original scheme we chose e = 0.6. By intro-
ducing non-standard weighting of the explicit numerical dissipa-
½a1 ; a2 ; a3  ¼ ½0:15; 0:4; 1:0
tion on stages of the RK scheme, the lower bound on e can be
from Ref. [14]. decreased, which results in faster convergence. Here, weighting
16 R.C. Swanson, C.-C. Rossow / Computers & Fluids 42 (2011) 13–25

1 1
(a) (b)
ε = 1.0 ε = 1.0
ε = 0.8 ε = 0.8
0.8 ε = 0.6 0.8 ε = 0.6
ε = 0.5 ε = 0.5
ε = 0.4 ε = 0.4
0.6 0.6
g

g
0.4 0.4

0.2 0.2

0 0
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
θx θx

Fig. 1. Effect on amplification factor of RK3/Implicit scheme (applied to 1-D Euler equations) due to variation of implicit parameter e (3-stage scheme). (a) Standard weighting
of numerical dissipation, (b) modified weighting of dissipation.

of the numerical dissipation means that the dissipation in the applied eigensystem analysis to show that the RK/ Implicit scheme
residual function on a given RK stage q (q > 1) is weighted with that with a Gauss-Seidel preconditioner and red-black ordering is an
from a previous stage. In general, there is weighting of both the effective smoother for multigrid. With a different data structure
numerical dissipative and physical diffusive terms, and the opera- such as that employed for unstructured grids, an alternative order-
tor L in Eq. (3.2) is defined by ing can be more convenient to implement and even lead to a more
" # robust iterative method. Due to the larger stencils that are pro-
1 Xq Xq
LW ðqÞ
¼ ðqÞ
Lc W þ ðrÞ
cqr Lv W þ ðrÞ
cqr Ld W ; ð3:8Þ duced on unstructured grids when approximating the spatial
V r¼0 r¼0 derivatives of the governing equations, the red-black ordering
P must be replaced by a multicolor ordering to ensure that each
with cqr ¼ 1 for consistency. The operators Lc ; Lv , and Ld relate to point of a particular color only directly connects to points of a dif-
the convective, viscous, and numerical dissipative terms. The coef- ferent color, and thus, obtain a Gauss-Seidel type scheme.
ficients cqr are the weights of the viscous and dissipative terms on By using multicoloring the algorithm can be highly parallelized
each stage, and for the 3-stage scheme, (see Refs. [21,22]), and the convergence rate is the same as for
c00 ¼ c1 ; c10 ¼ 1  c2 ; c11 ¼ c2 ; c20 ¼ 1  c3 ; sequential processing. We can define each color in the multicolor-
ing ordering as a member of an independent set. Then, each solu-
c21 ¼ 0; c22 ¼ c3 : ð3:9Þ
tion point is only directly connected to points with a different
When the weights ½c 1 ; c
2 ; c
3  are [1, 1, 1], this is called standard color. First, all points of a particular color on all subdomains are
weighting. Based upon analysis and numerical testing we have updated in parallel. Next, another set of points of a different color
determined that the modified weights [1, 0.5, 0.5] lead to improved is updated, using the latest information available. This procedure is
robustness of the smoother. With this weighting there is a shift in continued until all points have been updated. In updating the solu-
the intersection of the locus of the residual eigenvalues to the left tion points of the initial color, Jacobi relaxation is used since no up-
along the negative real axis of the complex plane, which increases dated points are available. However, this would also be true for
the parabolic stability limit of the basic RK scheme by more than red-black (odd-even) Gauss-Seidel on a structured mesh. The num-
a factor of two. A detailed discussion of dissipation weighting is ber of colors required to build independent sets is a function of the
given in Jameson [17] and Swanson and Turkel [18]. Fig. 1b shows number of points in the stencil.
the effect on g due to weighting of the dissipation on the second
and third stages by a factor of 0.5. The modified scheme is now 3.2. Schemes for SA equation
stable when e = 0.4. However, in practice, for stability, good damp-
ing of the highest frequencies, and best convergence the parameter After discretizing Eq. (2.10), we consider the implicit form
e is taken to be 0.5.
In the application of the RK3/Implicit scheme as the smoother of ðI þ Lx þ Ly þ S J ÞDm
~ ¼ Rðm
~Þ; ð3:11Þ
a full approximation storage (FAS) multigrid method, the CFL num-
where Lx and Ly are the linear discrete operators for the terms of the
ber is increased to 1000 after 10 multigrid cycles. The hyperbolic
transport-type equation, S J is a Jacobian of the source term contain-
tangent function defined in
ing the production and destruction of turbulence contributions, and
n o
CFL ¼ 1000 tanh½ð0:005Þ2N1  ; ð3:10Þ Rðm
~Þ is the residual function. The operators for the two coordinate
directions are as follows:
where N is the number of cycles, is used to smoothly increase the Dt  u
CFL number. A W-type cycle (see Ref. [19]) is used to execute the Lx ¼ h dx  ðdx b1 þ b2 dx Þdx ;
V
multigrid. Coarse meshes are created by eliminating every other
Dt h u i
mesh line in each coordinate direction (i.e., full coarsening). Details Ly ¼ h dy  ðdy b1 þ b2 dy Þdy ; ð3:12Þ
V
of the multigrid method are given in the paper by Swanson et al. [8].
Although we have used lexicographic ordering in applying where du is a first-order upwind operator for the convective term, d
Gauss-Seidel, alternative ordering strategies can also be used. For is a standard central difference operator, and the coefficients b1, b2
example, the lexicographic ordering of the solution points can be are defined by the diffusion term of the turbulence model. The
replaced with red-black ordering. Roberts and Swanson [20] have parameter h indicates temporal accuracy. If h = 1/2, then the time
R.C. Swanson, C.-C. Rossow / Computers & Fluids 42 (2011) 13–25 17

derivative is approximated by a central difference, which is second- ½a1 ; a2  ¼ ½0:25; 1:0:


order accurate (i.e., Crank–Nicolson type scheme). When h = 1 the
approximation is a first-order backward difference, and we have a Subsequently this method is designated as the RKI-SGS scheme.
fully (an Euler) implicit scheme. The parameter h may also be Due to the strong nonlinearities of the source terms, we have
viewed as a measure of implicitness with h > 1 and 0 < h < 1 indicat- employed numerical evaluation to determine an appropriate num-
ing under-relaxation and over-relaxation, respectively (to see this ber of relaxation sweeps for the SLGS and RKI-SGS schemes. In the
consider the effect of h for large Dt). The source term Jacobian evaluation we also considered the effect of mesh density on the
and the residual function are defined by number of relaxation sweeps for solving the turbulence equation.
To achieve favorable convergence rates the turbulence equation
Dt @  Dt is solved on each stage of the RK/Implicit smoothing scheme for
SJ ¼ h S ðm
~ Þ þ SD ð m
~Þ ; Rðm
~Þ ¼ Rðm
~Þ: ð3:13Þ
V @m~ P2 V the mean flow equations. The CFL number for all solvers of the
For the convective and diffusive terms of the residual function we SA equation is 1000. When solving the mean flow equations, solu-
use first-order upwind difference and central difference approxima- tion of the turbulence equation is performed on the fine mesh only,
tions, respectively. A first-order approximation of convective terms and the eddy viscosity is frozen on the coarser meshes. For addi-
is frequently applied in the implementation of turbulence models to tional enhancement of efficiency and robustness when solving
promote positivity of the turbulence variables. In general, this is not Eq. (2.2) the three different solution strategies, namely DDADI,
sufficient to ensure positivity, so usually there is also limiting (clip- SLGS, and RKI-SGS, are supported by a V-cycle multigrid algorithm.
ping) of the turbulence quantities and/or certain terms (e.g., pro- The multigrid algorithm is called at each stage of the fine mesh RK/
duction term) in the set of turbulence field equations. Implicit scheme when solving the mean flow equations.
An appropriate linearization of the source term is extremely One issue that can arise in solving the turbulence equation is
important to allow the use of large CFL numbers. One approach the unbounded growth of the solution. Exponential growth of m ~
for solving Eq. (3.11) is to factor the implicit (left-hand side) oper- can occur when there is a sufficiently large imbalance of the pro-
ator and apply the DDADI scheme. Define the diagonal contribu- duction and destruction terms so as to produce an instability. As
tion in Eq. (3.11) as pointed out by Allmaras [27], this is a consequence of the Jacobian
of the production and destruction terms becoming positive. Such a
D ¼ I þ Dx þ Dy þ S J ; ð3:14Þ behavior can become a significant problem especially when using
where Dx and Dy are the diagonal parts of Lx and Ly, respectively. multistage (e.g., RK) relaxation and only updating the precondi-
Then, after factoring out D, we factor the resulting operator, tioner on the zeroth stage. In the current formulation we do not ob-
obtaining serve this type of problem. There are several possible reasons for
  this. One is the form of the weak coupling used, where the multi-
I þ Ly þ Dx þ S J D1 I þ Lx þ Dy þ S J Dm
~ ¼ Rðm
~Þ: ð3:15Þ grid scheme for the turbulence equation is separate from that of
the mean flow equations. Another reason is that the preconditioner
To invert this implicit operator, we solve the sequence of one-
is updated on each stage of the RK scheme. There is also the poten-
dimensional systems corresponding to the two coordinate direc-
tial benefit from reducing the positive contribution to the Jacobian
tions. To prevent deterioration in the allowable CFL number and
of the production and destruction terms (see Eq. (3.13)). These fac-
damping behavior of the DDADI scheme due to the factorization
tors have contributed to the increased reliability of the solver.
error and possible boundary condition lagging error, we use the
During the course of this work we have made the following
subiterative procedure described by Klopfer et al. [23]. Using Fou-
convergence behavior observations, which are similar to the ones
rier analysis and some applications of the iterative DDADI scheme
reported by Walsh and Pulliam [25]. The rate of development of
(also called the modified approximate factorization (MAF) scheme)
the turbulence field can signficantly affect the convergence of the
MacCormack and Pulliam [24] and Walsh and Pulliam [25] have
flow solver. Conversely, how well the flow solver converges can
demonstrated that a few subiterations (e.g., two to four) makes
have an impact on the effectiveness of the scheme for solving the
the DDADI scheme unconditionally stable and improves the damp-
equation set of the turbulence model. Moreover, when the RANS
ing properties. In Pulliam et al. [26] best performance for a diago-
and turbulence equations are being solved in a loosely coupled
nalized DDADI was obtained with three to six iterations, and
manner, an essential requirement for an effective total algorithm
three iterations were recommended. Certainly, the effectiveness
is that the numerical solution vector of each equation set exhibits
and reliability of the DDADI scheme depends on the convergence
a similar evolution rate.
behavior of the subiterative process and on the magnitude of the
implicit parameter h. By numerical testing we have found that four
subiterations produces reliable convergence and best performance 4. Computational results
when DDADI is used by itself or as a smoother for multigrid. Cur-
rently we use four subiterations when performing one outer itera- Computations for turbulent, viscous flow over the RAE 2822 air-
tion. Convergence with iteration and subiteration can be foil were performed to evaluate the convergence behavior of the
enhanced by choosing an appropriate implicit parameter. Over a RK3/Implicit scheme when applying the SA turbulence model.
range of mesh densities a h between 1.2 and 2.0 works well. Addi- The airfoil solutions were primarily calculated with the Cases 1
tional discussion of the present implementation is given in Swanson and 9 flow conditions given in Table 1 from the experimental
and Rossow [16]. investigation of Cook et al. [28]. In the table M1 is the free-stream
With the SLGS scheme the implicit operator of Eq. (3.11) is Mach number, a denotes the angle of attack, Rec represents the
approximately inverted in each iteration with two symmetric Reynolds number based on chord length, and xtr/c is the transition
Gauss-Seidel line relaxation sweeps (line solves performed in ra-
dial direction only). The RK2/Implicit scheme involves two RK
stages and an implicit preconditioner, and it is the same type of Table 1
scheme used to solve the mean flow equations. One point SGS Flow conditions for RAE 2822 airfoil.
sweep and one local (boundary layer + near wake) symmetric line
Cases M1 a (deg.) Rec xtr/c
relaxation sweep are applied twice to obtain an approximate inver-
Case 1 0.676 1.93 5.7  106 0.11
sion of the implicit preconditioner. The coefficients for the two-
Case 9 0.730 2.79 6.5  106 0.03
stage scheme are
18 R.C. Swanson, C.-C. Rossow / Computers & Fluids 42 (2011) 13–25

location divided by the chord length. For Case 1 the flow is Fig. 2 shows convergence histories for Case 9 of the schemes for
primarily subsonic with a relatively small region of supersonic the RANS and turbulence equations. For these three results on the
flow. For Case 9 the flow is transonic, with a shock wave occurring 320  64 grid the SA equation was solved with DDADI, SLGS, and
on the upper surface at approximately the 55% chord location. In RKI-SGS schemes. The L2 norm of the residual of the continuity
addition, an incompressible (M1 = 0.001) airfoil flow calculation equation is used as a measure of convergence for the flow equa-
was made with a = 2.79 and Rec = 6.5  106. tions. With each scheme the residual of the mean flow equations
In solving the flow equations structured meshes with a C-type is reduced 13 orders of magnitude in less than 75 multigrid cycles
topology were used. We primarily considered three mesh densi- (for an average reduction rate of about 0.65). In fact, the residual
ties, with the finest having 1280 cells around the airfoil and 256 histories for the mean flow equations essentially coincide. This is
cells in the radial direction. Successively coarser grids (640  128 not surprising since for all schemes the residuals of the turbulence
and 320  64) were generated with half as many cells as the next equation are reduced between seven and eight orders. The DDADI
finer grid in each coordinate direction. There is clustering of the scheme requires less CPU time than the other two schemes, as seen
grids at the leading and trailing edges of the airfoil and also at in Table 2. However, the residual is reduced about an order of mag-
the surface in the radial direction. The finest mesh has 1024 cells nitude more with the RKI-SGS scheme than the DDADI scheme. The
on the airfoil and a minimum normal mesh spacing of 3  106. RKI-SGS scheme has the advantage of being compatible with the
On the airfoil surface the maximum cell aspect ratio is 2032. The solver of the mean flow equations, allowing the possibility of con-
outer boundary is located at 20 chords away from the airfoil. To structing a fully coupled solver. In addition, it is amenable to appli-
investigate the RANS solver for a range of Reynolds (Re) numbers cation in an unstructured flow solver, since it does not require lines
we used a set of meshes (adapted to the Re of the flow [29]) con- across the entire domain for the solution algorithm. For these rea-
taining 368  88 cells. sons we use the RKI-SGS scheme to solve the SA equation.
In all the applications the same boundary conditions were im-
posed for the fluid dynamic equations. On the surface the no-slip 4.1. Essentially subsonic flow
condition was applied. At the outer boundary Riemann invariants
were used. A far-field vortex effect was included to specify the For Case 1 we first consider the effect of the approximation
velocity for an inflow condition at the outer boundary. A detailed order for the convective terms of the mean flow equations on the
discussion of the boundary conditions is given in Ref. [18]. In the coarse grids in the multigrid method. Usually, only first-order
computations two types of initial conditions were considered. accurate spatial discretization is used for these terms. In Fig. 3
One type uses the free-stream values of the dependent variables. the effect on convergence behavior when using a second-order
The other one uses an initial solution determined by applying approximation is shown. Clearly there is a significant improvement
full multigrid (FMG). With FMG a grid sequencing process is in convergence with the second order, as the number of multigrid
used to generate an initial solution on successively finer meshes. cycles to reduce the residual of the flow equations 13 orders is de-
Multigrid is used to solve the discrete problem on each grid in creased from 82 cycles to 65 cycles. The residual of the SA equation
the sequence. All computations were performed on a Fujitsu (using the RKI-SGS scheme) is reduced to almost the same level
computer with an Intel core two duo CPU 6750 processor at (exceeding nine orders) in both calculations. In all subsequent re-
2.66 GHz. sults for subcritical flows the second-order approximation is used.
When comparing the computational performance of the RK3/ Furthermore, to elimnate the possibility of convergence effects due
Implicit scheme for different turbulence model solvers and for dif- to limiting, no limiter is applied.
ferent mesh densities, the computational time is included. These The effect of mesh refinement on convergence for Case 1 is
computing times provide a reasonable estimate of performance shown Fig. 4. The finest mesh (1280  256) contains over
since all solvers were programmed in Fortran 77 by the same per- 300,000 cells. Similar convergence behavior is obtained on all grids
son using the same coding practices. This also applies to compari- for both the mean flow and turbulence equations. As revealed in
sons that are made with a frequently used scheme for solving the Table 3 the convergence rate in solving the mean flow equations
RANS equations. Furthermore, by providing a description of the is approximately 0.6, and the CPU time is increased by about a fac-
processor used, the computational times required on other com- tor of four as the number of mesh points is doubled in each coor-
puters can be determined. dinate direction, indicating convergence without mesh

(a) 0
RK3/Implicit, SA Model, RAE 2822: Case 9
M ∞ = 0.73, α = 2.79 , Re = 6.5 x 10
o 6
1 (b) 4 SA Model,M RAE 2822: Case 9, 320 x 64
= 0.73, α = 2.79 , Re - 6.5 x 10
o 6

-2 2
0.8 DDADI
Grid: 320 x 64 0 SLGS
-4
Log (||Res tur||2)
Log (||Res|| 2)

RKI-SGS
DDADI -2
-6 CL
0.6
SLGS
CL

CL
-4
-8 RKI-SGS
CL -6
0.4
-10
-8

-12 -10
0.2

-14 -12
0 20 40 60 80 100 120 0 20 40 60 80 100 120
Cycles Cycles

Fig. 2. Convergence histories for solvers of flow and turbulence equations (Case 9, grid: 320  64). SA equation solved with three different methods: DDADI, SLGS, RKI-SGS.
(a) Flow equations, (b) SA equation.
R.C. Swanson, C.-C. Rossow / Computers & Fluids 42 (2011) 13–25 19

Table 2 Table 3
Comparison for Case 9 of solution strategies for solving the turbulence equation of the Effect of mesh density on convergence of RK3/Implicit scheme (Case 1). SA model
SA model (grid: 320  64). solved with RKI-SGS scheme.

Method CPU time (s) MG cycles Mesh size CPU time (s) MG cycles Convergence rate
DDADI 63 69 320  64 71 65 0.629
SLGS 65 70 640  128 299 63 0.619
RKI-SGS 75 69 1280  256 1242 60 0.604

Table 4
dependency. In Table 4 a comparison is made of the computational Comparison of computational efficiency of RK3/Implicit scheme with that of SLGS and
efficiency of the current RK3/Implicit scheme (denoted by RK3/I) tuned RK5/S schemes. Case 1 on the 320  64 grid.
with the SA and BL turbulence models. In addition, the computa-
Scheme Turb. model CPU time (s) MG cycles Convergence rate
tional effort required by a highly tuned standard five stage RK
scheme (RK5/S) with three evaluations of numerical dissipative RK3/I SA 71 65 0.629
SLGS SA 152 344 0.917
and physical diffusive terms, scalar implicit residual smoothing, RK3/I BL 44 64 0.624
and multigrid is given. The BL model was used when applying SLGS BL 128 351 0.918
the RK5/S algorithm. With the BL model the RK3/Implicit scheme RK5/S BL 181 1792 0.983
is about four times faster than the RK5/S scheme. Convergence his-
tories with the BL model are given in Fig. 5. Even with the addi-
tional computing time required by the SA model, the RK3/
Implicit scheme is still about two times faster than the RK5/S To provide an additional perspective on the efficiency of the
scheme. The computing time with the SA model is increased by RK/Implicit algorithm Table 4 also includes a comparison with
roughly a factor of 1.6 relative to that with the BL model. the SLGS scheme when used to solve both the mean flow and SA

RKI-SGS, SA Model, RAE 2822: Case 1


(a) 0 RK3/Implicit, SA Model, RAE 2822: Case 1
M = 0.676, α = 1.93 , Re = 5.7 x 10
o 6
0.7 (b) 4
M ∞ = 0.676, α = 1.93 , Re - 5.7 x 10
o 6

-2 2
0.6
Grid: 320 x 64
0
-4
Log (||Res tur||2)
Log (||Res|| 2)

Grid: 320 x 64 1st order


0.5 -2 2nd order
-6 1st order
CL

CL -4
2nd order 0.4
-8 CL
-6
-10
0.3 -8

-12 -10
0.2
-14 -12
0 20 40 60 80 100 120 0 20 40 60 80 100 120
Cycles Cycles

Fig. 3. Effect on convergence of approximation order of convective terms in the mean flow equations on coarse grids of the multigrid method (SA model, Case 1, grid density:
320  64). (a) Flow equations, (b) SA equation.

(a) 0 RK3/Implicit, SA Model, RAE 2822: Case 1


0.7 (b) 4
RKI-SGS, SA Model, RAE 2822: Case 1
6
M ∞ = 0.676, α = 1.93 , Re - 5.7 x 10
6 o
M ∞ = 0.676, α = 1.93 , Re = 5.7 x 10
o

-2 2
0.6
0 320 x 64
-4 640x 128
Log (||Restur||2)

320 x 64
Log (||Res||2)

1280 x 256
CL 0.5 -2
640 x 128
-6
CL
CL

1280 x 256 -4
-8 CL 0.4
-6
-10 0.3 -8

-12 -10
0.2
-14 -12
0 20 40 60 80 100 120 0 20 40 60 80 100 120
Cycles Cycles

Fig. 4. Convergence histories for solvers of flow and turbulence equations for Case 1 on three grids. SA equation solved with RKI-SGS. (a) Flow equations, (b) SA equation.
20 R.C. Swanson, C.-C. Rossow / Computers & Fluids 42 (2011) 13–25

RK3/Implicit, BL Model, RAE 2822: Case 1 Table 5


0 0.7 Effect of mesh density on computed lift and drag coefficients for Case 1.
M ∞ = 0.676, α = 1.93 , Re = 5.7 x 10
o 6

-2 Mesh size Cycles (FMG) CL CD (CD)p (CD)f


0.6
160  32 50 0.5840 0.010030 0.004012 0.006023
-4 320  64 50 0.5915 0.008541 0.002669 0.005872
Log (||Res|| 2)

320 x 64 0.5 640  128 50 0.5903 0.008298 0.002497 0.005802


-6 CL 1280  256 50 0.5884 0.008261 0.002484 0.005777
640 x 128 640  128 3 0.5900 0.008300 0.002501 0.005798

CL
CL 0.4 640  128 5 0.5901 0.008303 0.002503 0.005800
-8
640  128 10 0.5903 0.008298 0.002497 0.005802
1280  256 3 0.5887 0.008254 0.002484 0.005770
-10 0.3 1280  256 5 0.5885 0.008254 0.002480 0.005775
1280  256 10 0.5884 0.008261 0.002484 0.005777
-12
0.2
-14
0 20 40 60 80 100 120 So far we have presented results for grids with moderately high
Cycles aspect ratio cells. Fig. 8 shows the residual histories for Case 1
when the Re number is varied by more than an order of magnitude
Fig. 5. Convergence histories with BL model for Case 1. Second-order approxima- (from 5.7  106 to 100  106). The SA equation was solved with the
tion of convective terms on coarse grids. RKI-SGS scheme. Even at a Re = 100  106 a good convergence rate
(0.751) is still obtained for the RK3/Implicit scheme. Despite a Rey-
nolds number increase exceeding an order of magnitude, there is
only a factor of about two increase in computational effort. A com-
equations. As a multigrid smoother for the mean flow, the SLGS
parison of the RK3/Implicit and RK5/S schemes (with the SA and BL
scheme, due to stability, required a reduced CFL of 100 and under-
models, respectively) reveals that the RK3/Implicit method is more
relaxation on the coarse grids. When solving the SA equation, the
than five times faster when Re = 100  106.
RK3/Implicit scheme requires less than half the computer time of
the SLGS scheme.
In Fig. 6 convergence plots with FMG, starting on a 160  32 4.2. Incompressible flow
grid, are displayed. Four grids were used on all levels of grid refine-
ment, except for the 1280  256 level, which used five grids. The Since the numerical dissipation matrix of the present scheme is
benefit of the FMG in accelerating the convergence of a global written as a function of Mach number (see Refs. [7,8]), the dissipa-
quantity such as lift coefficient (CL) is evident. In Table 5 the com- tion can be scaled appropriately for low-speed flows. To demon-
puted CL and drag coefficient (CD) are presented for each grid level. strate the effectiveness of the present algorithm at a low Mach
The two contributions to the total drag coefficient, pressure drag number we consider an incompressible airfoil flow. Except for
(CD)p and skin-friction drag (CD)f coefficients, are also given. The the free-stream Mach number of M1 = 0.001, the flow conditions
development of these coefficients after three, five, and 10 multigrid are the same as for Case 9. Fig. 9 exhibits the residual histories.
cycles on each level are included in the table as well. In 10 cycles Here the density residual is decreased by only nine orders of mag-
the CL and CD are obtained to at least four significant digits. Fur- nitude to avoid round-off errors [7]. Removal of round-off errors at
thermore, with just three cycles on each level of the FMG the error low Mach number can be achieved by introducing a gauge pressure
in these quantities is less than 0.1%. [30].
Fig. 7 shows a comparison of the computed surface pressure
and skin-friction distributions on the 1280  256 grid with exper- 4.3. Transonic flow
imental data. In general, there is very good agreement with the
data. The computational pressure distribution does exhibit a weak In Fig. 10 the convergence histories on three grids is presented
shock on the upper surface of the airfoil in the transition region for Case 9. For these results the limiter was activated, and first-
(11% chord location). There is insufficient data in the region to ver- order differencing was used for coarse-grid convective terms. As
ify this behavior. for Case 1, similar convergence behavior is obtained on all grids.

RK3/Implicit, SA Model, RAE 2822: Case 1 RKI-SGS, SA Model, RAE 2822: Case 1
(a) 0 0.7 (b) 4
160 x 32 320 x 64 640 x 128 1280 x 256
-2
160 x 32 320 x 64 640 x 128 1280 x 256 2
0.6
0
-4
Log (||Restur||2)
Log (||Res||2)

0.5 -2
-6
CL

-4
-8
0.4
-6
-10
-8
0.3
-12 Residual -10
CL
-14 0.2 -12
0 50 100 150 200 0 50 100 150 200
Cycles Cycles

Fig. 6. Convergence histories with FMG for solvers of flow and turbulence equations (Case 1). SA equation solved with RKI-SGS. (a) Flow equations, (b) SA equation.
R.C. Swanson, C.-C. Rossow / Computers & Fluids 42 (2011) 13–25 21

(a) RK3/Implicit, SA Model, Case 1


(b) RK3/Implicit, SA Model, Case 1
-1.5 0.008

Exp.
-1 0.006 1280 x 256

-0.5
0.004
Cp

Cf
0
0.002

0.5
Exp.
1280 x 256 0
1

-0.002
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x/c x/c

Fig. 7. Comparison of computed surface pressures and skin friction with experimental data (Case 1, grid: 1280  256). (a) Surface pressures, (b) surface skin friction.

(a) 0 RK3/Implicit, SA Model, RAE 2822: Case 1


(b) 4 RKI-SGS, SA Model, RAE 2822: Case 1
M ∞ = 0.676, α = 1.93 , Re = 5.7 x 10
o 6
M ∞ = 0.676, α = 1.93 , Re = 5.7 x 10
o 6

-2 2
5.7 x 106
Grid: 368 x 88 0 20 x 106
-4
Log (||Res tur||2)

6
Log (||Res|| 2)

57 x 10
-2 100 x 106
-6
-4
-8
-6
-10
-8

-12 -10
5.7 m 20 m 57 m 100 m

-14 -12
0 20 40 60 80 100 120 0 20 40 60 80 100 120
Cycles Cycles

Fig. 8. Effect of Reynolds number variation on convergence of solvers for RANS and turbulence equations (Case 1, grid: 368  88). RKI-SGS scheme used to solve SA equation.
(a) Flow equations, (b) SA equation.

(a) 0 RK3/Implicit, SA Model, Incomp.


0.7 (b) 4 RKI-SGS, SA Model, Incomp.
M ∞ = 0.001, α = 2.79 , Re = 6.5 x 10 M ∞ = 0.001, α = 2.79 , Re - 6.5 x 10
o 6 o 6

2
-2 0.6
320 x 64
0
Log (||Res tur||2)

640 x 128
Log (||Res||2)

0.5
-4
320 x 64 -2
CL

CL
640 x 128 0.4 -4
-6 CL

-6
0.3
-8
-8
0.2
-10 -10
0 20 40 60 80 0 20 40 60 80
Cycles Cycles

Fig. 9. Convergence histories for solvers of flow and turbulence equations for incompressible case on two grids. SA equation solved with RKI-SGS. (a) Flow equations, (b) SA
equation.

The rate of convergence on the three grids is between 0.63 and from the 640  128 grid to the 1280  256 grid is slightly greater than
0.65. From Table 6 we see that the increase in CPU time in going a factor of four, which suggests a weak dependence of convergence
22 R.C. Swanson, C.-C. Rossow / Computers & Fluids 42 (2011) 13–25

(a) 0 RK3/Implicit, SA Model, RAE 2822: Case 9


1 (b) 4 RKI-SGS,M SA Model RAE 2822: Case 9
= 0.730, α = 2.79 , Re - 6.5 x 10
o 6
M ∞ = 0.730, α = 2.79 o, Re = 6.5 x 10 6 ∞

-2 2
0.8 320 x 64
0
-4 640 x 128

Log (||Res tur||2)


320 x 64
Log (||Res|| 2)

CL 1280 x 256
-2
640 x 128
-6 0.6
CL

CL
1280 x 256 -4
-8 CL
0.4 -6
-10
-8

-12 0.2 -10

-14 -12
0 20 40 60 80 100 120 0 20 40 60 80 100 120
Cycles Cycles

Fig. 10. Convergence histories for solvers of flow and turbulence equations for Case 9 on three grids. SA equation solved with RKI-SGS. (a) Flow equations, (b) SA equation.

Table 6 RK3/Implicit, BL Model, RAE 2822: Case 9


Effect of mesh density on convergence of RK3/Implicit scheme (Case 9). SA model 0 1
M ∞ = 0.73, α = 2.79 , Re = 6.5 x 10
o 6

solved with RKI-SGS scheme.


-2
Mesh size CPU time (s) MG cycles Convergence rate
0.8
320  64 75 69 0.648 -4
640  128 308 64 0.626 Log (||Res||2) 320 x 64
1280  256 1307 66 0.632 -6 CL
0.6
640 x 128

CL
CL
-8
Table 7 0.4
Comparison of computational efficiency of RK3/Implicit scheme with that of SLGS and -10
tuned RK5/S schemes. Case 9 on 320  64 grid.
-12
Scheme Turb. model CPU time (s) MG cycles Convergence rate 0.2
RK3/I SA 75 69 0.648 -14
SLGS SA 268 632 0.954 0 20 40 60 80 100 120
RK3/I BL 44 62 0.616 Cycles
SLGS BL 211 599 0.951
RK5/S BL 191 1891 0.984
Fig. 11. Convergence history for Case 9 using the BL model.

The convergence of the RK3/Implicit scheme with the BL model


on mesh density. The computational efficiency of the RK3/Implicit is similar to that obtained with the SA model, as revealed in the
scheme with both the SA and BL models is given in Table 7. convergence plots of Fig. 11.
For this case the computer time of the RK3/Implicit scheme is The convergence behavior with FMG for Case 9 is displayed in
approximately a factor of 3.5 smaller than that of the SLGS scheme. Fig. 12. As in Case 1, we observe a rapid evolution of the CL. Table
Again, the RK3/Implicit scheme with the SA model is about two 8 gives the CL and CD after three, five, 10, and 50 multigrid cycles.
times faster than the standard scheme RK5/S with the BL model. Even for this transonic case these coefficients are obtained to four

(a) 0 RK3/Implicit, SA Model, RAE 2822: Case 9


1 (b) 4 RKI-SGS, SA Model, RAE 2822: Case 9

160 x 32 320 x 64 640 x 128 1280 x 256


-2
160 x 32 320 x 64 640 x 128 1280 x 256 2
0.8
0
-4
Log (||Res tur||2)
Log (||Res||2)

-2
-6 0.6
CL

-4
-8
-6
0.4
-10
-8
-12 Residual -10
0.2
CL
-14 -12
0 50 100 150 200 0 50 100 150 200
Cycles Cycles

Fig. 12. Convergence histories with FMG for solvers of flow and turbulence equations (Case 9). SA equation solved with RKI-SGS. (a) Flow equations, (b) SA equation.
R.C. Swanson, C.-C. Rossow / Computers & Fluids 42 (2011) 13–25 23

Table 8 enhance efficiency and robustness of these schemes multigrid


Effect of mesh density on computed lift and drag coefficients for Case 9. acceleration has also been applied. For both the fluid dynamic
Mesh size Cycles (FMG) CL CD (CD)p (CD)f and turbulence equations a CFL of 1000 has been used.
160  32 50 0.7955 0.01748 0.01192 0.005563 Similar convergence behavior has been observed for the three
320  64 50 0.8185 0.01678 0.01120 0.005576 schemes evaluated for solving the turbulence model equation.
640  128 50 0.8227 0.01665 0.01113 0.005510 Although the computational effort required with the RKI-SGS
1280  256 50 0.8238 0.01655 0.01108 0.005474 scheme is somewhat larger than that needed with the other two
640  128 3 0.8152 0.01650 0.01103 0.005471
640  128 5 0.8213 0.01661 0.01111 0.005500
schemes, it provides important advantages. With appropriate
640  128 10 0.8227 0.01665 0.01113 0.005510 ordering for Gauss-Seidel, this method can be implemented in an
1280  256 3 0.8222 0.01654 0.01107 0.005471 unstructured grid algorithm. In addition, it allows for the possibil-
1280  256 5 0.8235 0.01655 0.01108 0.005474 ity to solve the mean flow and turbulence equations in a fully cou-
1280  256 10 0.8238 0.01655 0.01108 0.005474
pled manner. Thus, the present loosely coupled algorithm is based
on RK/Implicit schemes. This algorithm also has the advantage that
significant digits in just 10 cycles. After five cycles the computed CL it can be readily incorporated into many existing codes that
and CD, on both the 640  128 and 1280  256 grids, are obtained employ RK smoothers for multigrid methods.
to within about 0.25% of their final values. On the finest grid with The performance of the loosely coupled algorithm (RK3/I +
three cycles on each refinement level the coefficients have an error RKI-SGS schemes) has been investigated by computing solutions
of less than 0.2%. In Fig. 13 a comparison is made of the calculated to subsonic and transonic airfoil flows. We have demonstrated that
surface pressure and skin-friction variations on the finest grid at there is no significant slowdown in convergence of the RK/Implicit
three, five, and 10 cycles. With just three cycles on each level there scheme when the SA model is used instead of the algebraic model
are bearly discernible differences on the upper airfoil surface and of Baldwin and Lomax. This is quite important since it suggests
at the shock. The distributions on the finest grid are compared with that for at least similar 3-D problems, such as wing flows, the
the experimental data in Fig. 14. There is fairly good agreement performance of the 3-D scheme for the SA model will be similar
with the data. to that observed for the BL model. In addition, even with the SA
For all the computations the local line solves of the RKI-SGS model, the loosely coupled algorithm is approximately two to five
scheme were terminated at the jl/4 location, where jl is the number times faster, depending on the Reynolds number, than the highly
of cells in the normal direction to the airfoil. In Fig. 15 the effect of tuned standard RK scheme (RK5/S) with the BL model. It should
varying the number of points in the line solves is shown. The con- be emphasized that the RK5/S scheme includes three evaluations
vergence is only slightly faster by doubling the number of points in of the dissipative and diffusive terms, multigrid, and scalar implicit
the normal direction. residual smoothing. The RK/Implicit algorithm has also been com-
pared to the SLGS scheme when applied to both the mean flow and
SA equations. It is between two and 3.5 times faster than the SLGS
5. Concluding remarks scheme, depending on the flow conditions.
Although the indirectly coupled algorithm uses local line solves
In this work the fluid dynamic (RANS) equations and the trans- (in boundary layer and wake) rather than line solves extending
port-type equation of the SA turbulence model have been solved in across the entire domain, there is no significant deterioration in
a loosely coupled manner. The RANS equations have been solved convergence. The RK/Implicit schemes applied to the mean flow
with a RK3/Implicit scheme (RK3/I) and multigrid. This scheme and turbulence equations have exhibited a low sensitivity to dis-
has been enhanced by weighting the numerical dissipative and crete stiffness associated with large aspect ratio mesh cells. Fur-
physical diffusive terms, a smooth initial increase of the CFL num- thermore, it has been shown that the algorithm can also
ber, and local line implicit relaxation. Three different methods have effectively solve a low-speed flow; and thus, the analytical stiffness
been considered for solving the SA equation: diagonally dominant due to disparity in wave speeds has been removed.
alternating direction implicit (DDADI), symmetric line Gauss-Seidel By using FMG to generate the initial conditions on the solution
(SLGS), and a RK2/Implicit with local line solves (RKI-SGS). To grid, we have observed rapid development of the aerodynamic

(a) RK3/Implicit, SA Model, Case 9, 1280 x 256


(b) RK3/Implicit, SA Model, Case 9, 1280 x 256
-1.5 0.008

3 cycles
-1 0.006 5 cycles
50 cycles

-0.5
0.004
Cp

Cf

0
0.002

0.5
3 cycles
5 cycles 0
50 cycles
1

-0.002
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x/c x/c

Fig. 13. Effect of number of cycles in each level of FMG on computed surface pressures and skin friction (Case 9, grid: 1280  256). (a) Surface pressures, (b) Surface skin
friction.
24 R.C. Swanson, C.-C. Rossow / Computers & Fluids 42 (2011) 13–25

(a) RK3/Implicit, SA Model, Case 9


(b) RK3/Implicit, SA Model, Case 9
-1.5 0.008

Exp.
-1 0.006 1280 x 256

-0.5
0.004
Cp

Cf
0
0.002

0.5
Exp.
1280 x 256 0
1

-0.002
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x/c x/c

Fig. 14. Comparison of computed surface pressures and skin friction with experimental data. (Case 9, grid: 1280  256). (a) Surface pressures, (b) Surface skin friction.

(a) 0 RK3/Implicit, SA Model, RAE 2822: Case 9


1 (b) 4 RKI-SGS, SA Model, RAE 2822: Case 9
M ∞ = 0.73, α = 2.79 , Re = 6.5 x 10 M ∞ = 0.73, α = 2.79 , Re - 6.5 x 10
o 6 o 6

-2 2
0.8
0
-4 jl/4
Log (||Restur||2)
Log (||Res|| 2)

jl/2
jl/4 -2 jl
-6 CL 0.6
jl/2
CL

CL -4
-8 jl
CL 0.4 -6
-10
-8

-12 0.2 -10

-14 -12
0 20 40 60 80 100 120 0 20 40 60 80 100 120
Cycles Cycles

Fig. 15. Effect of number of points in line solves on convergence (Case 9, grid: 640  128). (a) Flow equations, (b) SA equation.

coefficients for turbulent viscous flows at both subsonic and tran- [13] Roe PL. Approximate Riemann solvers, parameter vectors and difference
schemes. J Comput Phys 1981;43:357–72.
sonic speeds. Moreover, in just three cycles on each refinement le-
[14] Van Leer B, Tai C-H, Powell KG. Design of optimally smoothing multi-stage
vel of the FMG, the lift and drag coefficents for both the subsonic schemes for the Euler equations. AIAA paper 89-1933; June 1989.
and transonic cases have an error less than 1.0%. [15] Jameson A. The evolution of computational methods in aerodynamics. J Appl
Mech 1983;50(4b):1052–70.
[16] Swanson RC, Rossow C-C. An initial investigation of the effects of turbulence
References models on the convergence of the RK/Implicit scheme. NASATM-2008-
215342; 2008.
[1] Dick E. A flux-vector splitting method for steady Navier–Stokes equations. Int J [17] Jameson A. Numerical solution of the Euler equation for compressible inviscid
Numer Methods Fluids 1988;8:317–26. fluids. MAE report 1643, Princeton University, Princeton, New Jersey;
[2] Wackers J, Koren B. Multigrid solution method for the steady RANS equations. J December 1983.
Comput Phys 2007;226:1784–807. [18] Swanson RC, Turkel E. Multistage schemes with multigrid for Euler and
[3] Mavriplis DJ, Venkatakrishnan V. Agglomeration multigrid for two Navier–Stokes equations. NASA TP 3631; 1997.
dimensional viscous flows. Comput Fluids 1995;24(5):553–70. [19] Trottenberg U, Oosterlee CW, Schüller A. Multigrid. Academic Press; 2001.
[4] Pierce NA, Giles MB. Preconditioned multigrid methods for compressible flow [20] Roberts TW, Swanson RC. A study of multigrid preconditioners using
calculations on stretched meshes. J Comput Phys 1997;136:425–45. eigensystem analysis. AIAA paper 2005-5229; June 2005.
[5] Jespersen DC, Pulliam TH, Buning PG. Recent enhancements to overflow. AIAA [21] Adams M, Brezina M, Hu J, Tuminaro R. Parallel multigrid smoothing:
paper 97-0644; January 1997. polynomial versus Gauss-Seidel. J Comput Phys 2003;188(2):593–610.
[6] Venkatakrishnan V. Improved convergence of compressible Navier–Stokes [22] Chow E, Falgout RD, Hu JJ, Tuminaro RS, Yang UM. A survey of parallelization
solvers. AIAA paper 98-2967; June 1998. techniques for multigrid solvers. In: Heroux MA, Raghaven P, Simon HD,
[7] Rossow C-C. Efficient computation of compressible and incompressible flows. J editors. Parallel Processing for Scientific Computing. SIAM; 2006. p. 179–201
Comput Phys 2007;220:879–99. [chapter 10].
[8] Swanson RC, Turkel E, Rossow C-C. Convergence acceleration of Runge-Kutta [23] Klopfer GH, Van der Wijngaart RF, Hung CM, Onufer JT. A diagonalized
schemes for solving the Navier–Stokes equations. J Comput Phys diagonal dominant alternating direction implicit (D3ADI) scheme and
2007;224:365–88. subiteration correction. AIAA paper 98-2824; June 1998.
[9] Baldwin BS, Lomax H. Thin layer approximation and algebraic model for [24] MacCormack RW, Pulliam TH. Assessment of a new numerical procedure for
separated flows. AIAA paper 78-257; January 1978. fluid dynamics. AIAA paper 98-2821; June 1998.
[10] Spalart PR, Allmaras SR. A one-equation turbulence model for aerodynamic [25] Walsh PC, Pulliam T. The effect of turbulence model solution on viscous flow
flows. La Recherche Aerospatiale 1994;1:5–21. problems. AIAA paper 2001-1018; January 2001.
[11] Krist SL, Biedron RT, Rumsey CL. CFL3D user’s manual. NASA TM 1998-208444; [26] Pulliam TH, MacCormack RW, Venkateswaran S. Convergence characteristics
June 1998. of approximate factorization methods. In: Sixteenth international conference
[12] Rumsey CL. Apparent transition behavior of widely-used turbulence models. on numerical methods in fluid dynamics. Lecture notes in physics, vol.
Int J Heat Fluid Flow 2007;28:1460–71. 515. Berlin/Heidelberg: Springer; 1998. p. 409–14.
R.C. Swanson, C.-C. Rossow / Computers & Fluids 42 (2011) 13–25 25

[27] Allmaras SR. Multigrid for the 2-D compressible Navier–Stokes equations. [29] Faßbender J. Improved robustness of numerical simulation of turbulent flows
AIAA paper 99-3336; July 1999. around civil transport aircraft at flight Reynolds numbers. Ph.D. thesis,
[28] Cook PH, McDonald MA, Firmin MCP. Aerofoil RAE 2822 pressure Technical University of Braunschweig, Braunschweig, Germany; 2004.
distributions and boundary layer and wake measurements. AGARD-AR- [30] Feng J, Merkle CL. Evaluation of preconditioning methods for time-marching
138; 1979. systems. AIAA paper 90-0016; January 1990.

Das könnte Ihnen auch gefallen