Computers & Fluids 80 (2013) 408–422

Contents lists available at SciVerse 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

Modeling of turbulent separated flows using OpenFOAM

Dmitry A. Lysenko a,⇑, Ivar S. Ertesvåg a, Kjell E. Rian b
a
Norwegian University of Science and Technology, Kolbjørn Hejes vei 1B, 7491-NO Trondheim, Norway
b
Computational Industry Technologies AS, 7462-NO Trondheim, Norway

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

Article history: Turbulent separated planar bluff-body flows were numerically analyzed using the state-of-the-art Open-
Received 13 September 2011 FOAM and ANSYS FLUENT technologies, based on the conventional URANS approach. Several popular in
Received in revised form 3 January 2012 fluid dynamics test problems such as laminar and turbulent flows over a circular cylinder and turbulent
Accepted 22 January 2012
fully developed flows over a triangular cylinder in a channel were numerically replicated with the goal of
Available online 3 February 2012
validation of the selected numerical methods. The detailed, face-to-face comparison between OpenFOAM,
FLUENT and experimental data was discussed. Parallel performance in the terms of a strong and weak
Keywords:
scalability was assessed up to 1024 cores and compared as well. In general, the present results demon-
Strong and weak scalability
Compressible URANS
strated minimum deviations between OpenFOAM and FLUENT and agreed fairly well with the experi-
Turbulent separated flows mental data and other numerical solutions.
k  e turbulence models Ó 2012 Elsevier Ltd. All rights reserved.

1. Introduction simulation) models for high Reynolds number flows of practical

interest with further adaptation for turbulent combustion model-
The long-term goal of the present work is to develop URANS ing. The aim of this particular study was to validate a numerical
(unsteady Reynolds averaged Navier–Stokes) and LES (large-eddy method based on the conventional compressible URANS approach
for modeling of turbulent separated flows. Since the flows around
. 
Abbreviations: DP, static pressure drop; DP, 2DP qU 21 ; Dt, time step (s); X, bluff-bodies are a common test case for URANS due to the large
mean vorticity magnitude; e, dissipation rate of turbulence kinetic energy (m2/s3); eddies formed in the wake, several well referenced benchmarks
c, NVD GAMMA differencing scheme parameter; j, Von Kármán constant; j, were selected keeping in mind their further adaption for the com-
0.4187; l, dynamic molecular viscosity (kg/m/s); ls, turbulent viscosity (kg/m/s); bustion and LES simulations.
x, specific rate of turbulence energy dissipation (Hz); /, ratio of specific heats; q,
mass density (kg/m3); h, circumferential coordinate (°); hsep, mean separation angle
The core numerical method was implemented in the open
(°); t2, velocity scale; A,B,C, constants; Al ; A ; C l , constants; Cl, constant, Cl = 0.09; source OpenFOAM toolbox (hereafter OF). This code was chosen
Cd, mean drag coefficient, C d ¼ 0:5F x =q1 U 21 Z x ; Cl, mean lift  coefficient,  because of several reasons. OF was originally developed as a hi-
C l ¼ 0:5F y =q1 U 21 Z y ; Cp, mean pressure coefficient, Cp ¼ 2ðP  P 1 Þ= qU 21 ; Ce1, end C++ classes library (Field Operation and Manipulation) for a
Ce3, constants in the production and sink terms of the e equation; Cp,b, mean base
broad range of fluid dynamics applications and quickly became
suction coefficient; CFL, courant number; D, bluff-body diameter or base (m); E(p),
parallel efficiency, E(p) = Ts/(Tpp) = S(p)/p; F, force in stream-wise direction, very popular in industrial engineering as well as in academic re-
consisting of the friction force Ff and the pressure force Fp (Pa); H, channel height
pffiffiffiffiffiffiffiffiffiffiffi
search. OF is open source and there are no limitations for parallel
(m); K, normalized turbulence kinetic energy, K ¼ 4=3k=U 1 ; L, channel length computing and no black boxes compare to commercial solvers like
(m); Lr, recirculation zone length (m); M, Mach number, M = U1/c1; N, number of ANSYS FLUENT (hereafter AF). From respect to combustion OF has
pffiffiffi Re, Reynolds number, Re = q1U1D/
cells; P, static pressure (Pa); Pr, Prandtl number;
no limitations for detailed-chemistry modeling (for example, AF
l; Rey, turbulent Reynolds number, Re ¼ qy k=l; S(p), strong scalability, S(p) = Ts/
Tp; St, Strouhal number, St = f D/U1; T, temperature (K); Tp, execution time using has the limit for maximum number of species transport equations).
multiple processor system (s); Ts, execution time using single processor system (s); From respect to programing OF is more effective than methods like
U, velocity magnitude (m/s); W(p), weak scalability, W(p) = Tp/Ts; Z, projected bluff- user-defined functions in AF. The wide list of the numerical
body surface in stream-wise or transverse direction (m); c, speed of sound (m/s); f,
schemes and mathematical models, implemented in OF, provides
Von Kármán vortex shedding frequency (Hz); f2, function in the sink term of the e
equation; k, turbulence kinetic energy, k ¼ 34 ðhu02 i þ hv 02 iÞ (m2/s2); ll,le, length robustness and efficiency of this technology for a wide spectrum
scales; p, number of processors (cores); t, time (s); u,v, stream-wise and transverse of fluid dynamics problems.
velocity components (m/s); x,y, stream-wise and transverse directions (m); y⁄, wall- Nevertheless, in spite of many attractive features, OF has some
normal distance (m); Subscripts: 1, value in an incoming flow; hi, mean or time- disadvantages as well. The most crucial among them is the absence
averaged value; o, stagnation value; x,y, stream-wise and transverse components of
a vector; 0 , fluctuation component; min, minimum value.
of the quality certification and as a consequence – the lack of high-
⇑ Corresponding author. quality documentation and references. Thus, the problem of OF val-
E-mail address: [email protected] (D.A. Lysenko). idation and verification becomes more principal and fundamental

0045-7930/$ - see front matter Ó 2012 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compfluid.2012.01.015
 D.A. Lysenko et al. / Computers & Fluids 80 (2013) 408–422 409

compared to other commercial computational fluid dynamics (CFD) Table 2

codes. The stretched goal of this study is to provide high-quality ver- OF/AF numerical methods, schemes and models.

ification and documentation of the selected numerical methods. AF OF

With this purpose we performed face-to-face results comparison Numerical method
with the ‘best-in-the-class’ commercial solver AF. Besides an accu- Algorithm URANS
rate prediction of the turbulent separated flows with quantitative Method Unstructured FVM
and qualitative representation of fluid mechanics complexity and Solver Pressure-based rhoPisoFoam
Pressure–velocity PISO PISO
behavior, the parallel performance of the numerical method is coupling
important as well (specially for unsteady, combustion and LES Linear algebra and GS/ILU, 107 ICCG, 107
applications). Therefore a parallel performance was assessed and accuracy
compared for both codes in terms of a weak and strong scalability Multigrid AMG –
Under-relaxation factors Default 0.3 – pressure, 0.7 –
up to 1024 cores.
others
The paper is organized as follows. Test case descriptions are gi-
Spatial discretization
ven in Section 2. In Section 3 the main features of the employed
Convective terms QUICK NVD, c = 0.1
numerical methods are summarized. The computational results Pressure Standard 4th order
are presented and analyzed in Section 4 and concluding remarks
Temporal discretization
are given in Section 5. Scheme Implicit dual- BDF2
stepping
Time step Constant Dt, CFL < 1, Dynamic D t, CFL < 1
2. Test cases description
Thermodynamics
Compressibility Ideal gas law
The selected numerical methods were validated against test Dynamic viscosity, l Constant, 1.7894  105
problems listed and referenced in Table 1. Any of these bench- Prandtl number, Pr Constant, 0.75
marks has been replicated numerically with different assumptions Turbulence model
and approaches in dozens of papers during several decades. The Modified k   RKE LSKE
overall accumulated experience in mathematical and experimental Near-the-wall treatment Low-Reynolds-number formulation
modeling dedicated to these data yields a high level of compliance Boundary conditions
for any of them. Inlet Fixed profiles
Outlet Pressure-far-field Wave-transmissive
Walls Isothermal non-slip
3. Brief description of numerics

The main emphasis of this work was put on the problem of

validation and verification of a numerical method implemented model [11]. A side-by-side comparison of the OF and AF numerical
in the OpenFOAM toolbox [10]. However, to achieve a more consis- methods, models and assumptions are given in Table 2 with the
tency and validity of the results ANSYS FLUENT [11] was used as description in the next several sections.
well. The quite similar numerical methods were chosen for the
present calculations. In both codes, a so-called pressure-density 3.1. OpenFOAM
solver, based on the projection method [12] for a solution of com-
pressible URANS equations, was used. One should notice that both The OpenFOAM code [10] v.1.7.1 was used for the numerical
codes have the broad lists of turbulence models for the closure simulations. The standard solver rhoPisoFOAM was utilized for
problem including one-equation Spalart–Allmaras model, the the compressible URANS modeling based on the finite-volume
family of k   models, k  x SST (Shear Stress Transport) model, (FVM) factorized method [15] and the predictor–corrector PISO
etc. In this study the family of k   models was chosen based on (pressure implicit with splitting of operators) algorithm [16].
the strong previous author’s experience. Another reason for this Two and one iterations were set for a PISO loop and for non-
pragmatic choice is the fact that k   models are still dominant orthogonal corrections, respectively. The generalized fully sec-
for combustion applications using RANS/URANS approach. The ond-order setup (in space and time) was used for all simulations.
modified low-Reynolds-number k   turbulence model of Launder The normalized variable diagram (NVD) type differencing scheme
and Sharma (hereafter LSKE) [13] was chosen as the ‘baseline’ – GAMMA [17] with c = 0.1 was applied for all convective terms
model for OF. This model has been implemented in AF as an undoc- approximation. All other inviscid terms and the pressure gradient
umented feature, however, we stayed at the Realizable k   model were approximated with a fourth order accuracy. A second order
(hereafter RKE) [14] for our simulations as ‘baseline’ for AF. It implicit Euler method (backward differentiation formula, BDF2
should be noticed that RKE model is implemented in OF, but in [15]) was used for the time integration together with the dynamic
its original high-Reynolds-number formulation only. Special atten- adjustable time stepping technique to guarantee the local Courant
tion was drawn to the treatment of a near-the-wall region. In all number less then CFL < 1. Preconditioned (bi-) conjugate gradient
cases the low-Reynolds-number formulation was used. With this method [18] with incomplete-Cholesky preconditioner (ICCG) by
purpose so-called ‘damping functions’ were developed in the LSKE Jacobs [19] was used for solving linear systems with a local accu-
model, and a two-layer approach was implemented for the RKE racy of 107 for all dependent variables at each time step.
The modified low-Reynolds-number k   turbulence model of
Launder and Sharma was chosen [13] for the Navier–Stokes equa-
Table 1
tions closure. Such approach does not require the specification of
Test matrix of the selected plane bluff-body flows.
the ‘wall-functions’ as used in a high-Reynolds-number formula-
Test case description Re M Refs. tions to describe a near-the-wall region treatment. So-called
Laminar flow over a circular cylinder 140 0.2 [1] ‘damping functions’ were introduced and incorporated in k  
Turbulent flow over a circular cylinder 3900 0.2 [2–6] model by Jones and Launder [50]. It was later re-optimized by
Turbulent flow over a triangular rod 1.75  104 0.03 [7,8]
Launder and Sharma [13], and demonstrated satisfactory results
Turbulent flow over a triangular rod 4.5  104 0.05 [9]
in many applications. The wide acceptance of such formulation
410 D.A. Lysenko et al. / Computers & Fluids 80 (2013) 408–422

gradually granted the status of the benchmark for low-Reynolds- planes. The reason for the inlet buffer domain was to avoid setting
number k   turbulence model [20]. In this work the model was of the turbulent velocity and temperature boundary profiles at a
applied with the following differences in the sink term of the  channel inlet since it was not measured in lab tests and to avoid
equation compared to the ‘original’ LSKE model [13]: investigating the influence of the boundary layer width on the flow
development.
1. Low-Reynolds-number function f2 was slightly changed: The turbulence intensity at the inlet was set equal to 3–4%
  
f2 ¼ 1  0:3 exp  min Re2y ; 50 . which is common for the typical wind tunnels. The characteristic
scale of the turbulence was set equal to the bluff-body diameter.
2. The model constant C3 was updated from the ‘original’ value
  The molecular viscosity and the thermal conductivity were taken
C3 = 2 to C 3 ¼ 0:33  23 C 1 .
to be constant. The Prandtl number was assumed to be Pr = 0.75,
and the ratio of specific heats was / = 1.4. The compressibility
3.2. ANSYS FLUENT was treated with the ideal gas law. The initial conditions, at the
moment of time t = 0, corresponded to the conditions of sudden
Using the factorized FVM [15] we solved compressible URANS stopping of a bluff-body in a flow, i.e., the input conditions were
equations with a second order accuracy in space and time. The lin- extended to the whole computational region.
ear system of equations was solved with Gauss–Seidel smoother or
the incomplete lower upper (ILU) decomposition smoothers, which
3.4. Data sampling
was accelerated by an algebraic multi-grid (AMG) technique, based
on the additive-correction strategy [21]. The convective terms
The turbulent wake has been considered as fully established
were represented according to the Leonard quadratic upwind
after a duration of t = 150D/U1. Statistics were collected over 40–
scheme (QUICK) [11]. It is worth to noticing, that due to its imple-
50 shedding cycles after that. Since the compressible formulation
mentation in AF, it is possible to use the QUICK scheme for the
with the stability condition CFL < 1 and the grids with high resolu-
unstructured non-hexahedral grids. However in this case the
tion of boundary layers were used, the typical values of Dt did not
method is switched to a second order upwind discretization
exceed 1  105 s. Thus the number of time steps per one shedding
scheme. The velocity and pressure fields were matched with a cen-
cycle varied in a range of 1000–2000. The main integral flow
tered computational template within the spirit of Rhie and Chou
parameters were: the mean pressure coefficient, mean drag coeffi-
[22]. The PISO [16] algorithm with a fixed time-stepping for phys-
cient, mean base suction coefficient, mean separation angle (deter-
ical-time integration was used.
mined from the condition of vanishing wall shear stress), mean
To close the system of equations we used the Realizable k  
recirculation length (corresponded to the distance between the
model of Shih [14]. Originally this model was developed in a high-
base of a bluff-body and the sign change of the centerline mean
Reynolds-number formulation. However, a two-layer approach
stream-wise velocity) and Strouhal number.
has been employed in FLUENT to specify both the  and the turbulent
viscosity in near-the-wall cells. The main idea of this approach is to
3.5. Mesh independence
subdivide a fluid domain in a vicinity of a wall into viscosity-affected
and fully-turbulent regions. For this purpose, turbulent Reynolds
It is common knowledge that the disagreement between experi-
number, Rey, is introduced. The boundary between these two
mental data and numerical results is determined by two groups of
regions is defined at Rey  200. In the fully-turbulent region
errors (apart from experimental errors): (1) ‘model’ errors due to
(Rey > Rey ) Realizable k   model is used. In the viscosity-affected
the inadequate assumptions made in selecting one turbulence mod-
near-the-wall region ðRey < Rey Þ the one equation model of
el or another and (2) ‘discretization’ errors caused by the inadequate
Wolfshtein [23] is employed for the turbulent
pffiffiffi viscosity calculation,
resolution of the employed finite-element grids and computational
which is computed from: ls ¼ qC l ll k. The  field in the viscosity-
methods. Whereas the errors of the first group are assumed to be
affected region is defined from  = k3/2/l. The length scales (ll and l)
‘systematic’ under certain assumptions, e.g., for a fixed computa-
are calculated according to Chen and Patel [24]:
tional methodology, ‘discretization’ errors are controlled by the
ll ¼ y C l ð1  expðRey =Al ÞÞ; method of adaptation (increase in the resolution) of a computational
l ¼ y C l ð1  expðRey =A ÞÞ; grid. The grid independence of the present results was confirmed by:

were the dimensionless constants are: C l ¼ jC 3=4 l ; Al ¼ 70;  the application of the two different CFD technologies (AF and
A ¼ 2C l . The blending functions for a smooth transition between OF);
the near-the-wall algebraically predicted  and ls and their values  the use of the projection method [12] for a solution of the com-
obtained from a solution of the Realizable k   model transport pressible URANS equations (or so-called pressure-based solver)
equations in the outer region are implemented according to Jongen with the quite similar implementation in both codes;
[25].  the use of different grid topologies (structured vs. unstructured)
and cell types (quadrilateral vs. triangular). This aspect was not
3.3. Boundary and initial conditions discussed there, since the investigated difference between the
results obtained for the grids with a different topology could
The following boundary conditions were applied. Inlet: fixed be considered as non-significant (less than 5%) and was dis-
values for velocity, temperature, turbulence kinetic energy and cussed in details earlier [27];
its dissipation rate; pressure – zero gradient. At the outlet wave-  some results of the grid convergence study are presented in
transmissive [26] or characteristic conditions were applied. Bluff- Fig. 1. The recirculation zone length was chosen as the most
body and channel walls were treated as isothermal no-slip condi- important integral parameter since the quality of its prediction
tions. For the problems where a bluff-body was located in a chan- may be considered as the deciding factor about the agreement
nel (assuming that the flow is fully-developed and turbulent), between experimental and numerical results [6]. One can see
additional inlet buffer domains were applied. In such cases the clearly that a deviation for Lr as a function of a cells number
upper and lower buffer domain boundaries were used as symmetry was bounded by ±2.5%.
 D.A. Lysenko et al. / Computers & Fluids 80 (2013) 408–422 411

est (Fig. 2b) strong scalability with the average efficiency of 40%.
It should be noted that OF data were collected including all input/
output (I/O) operations and that special network-hardware tuning
was not maintained. Concluding this paragraph, one can see clearly
from Fig. 2b that the parallel performance of the state-of-the-art
implicit pressure-based solvers varies from 40–70% of an ideal.

4. Results

Fig. 1. Grid convergence study for the recirculation zone length: 1 – the flow over a 4.1. Laminar unsteady flow over circular cylinder
circular cylinder at Re = 3900; 2 – the flow over a triangular rod at Re = 17500; 3 –
the flow over a triangular rod at Re = 45000. Methodical investigation of the laminar flow around a circular
cylinder at Re = 140 was carried out with the goal of validation,
verification and understanding of the numerical methods and their
3.6. HPC and scalability analysis capabilities implemented in the OF. Both incompressible and com-
pressible Ma = 0.2 formulations of the URANS approach were
Parallel scalability analysis is required to understand and opti- applied.
mize the existing software for high-performance computing Two types of the grids were used:
(HPC). Both, AF and OF are massive-parallel solvers, based on mes-
sage-passing (MPI) interface. In general, compressible flow solvers  The computational domain for the O-type grid (S1) had the form
(or so-called density-based) have a higher parallel performance of a circle (Fig. 3a). A cylinder of diameter, D = 0.1 m was located
compared to incompressible (and compressible) flow algorithms in the center of the computational domain. The size of the inte-
that are based on projection methods (pressure-based), i.e., they gration domain was 20  D. S1 grid was used for the incom-
require a solution of the pressure Poisson equation. However, pressible flow simulation. The grid points were clustered in
many state-of-the-art numerical algorithms are implemented in the vicinity of the cylinder (Fig. 3b). The obstacle as well as
an implicit manner, which significantly affects the parallel perfor- the outer boundary profiles were divided into 325 equal inter-
mance irrespective of the method for solving the Navier–Stokes vals. Radial states were divided into 325 intervals with an
(NS) equations. expansion factor in the radial direction of 1.020. For the com-
For the assessment of the parallel performance, such parame- pressible flow simulation the O-type grid (S2) with the size of
ters as strong scalability, weak scalability and system efficiency the computational domain of 50  D was used consisting
were used with their standard definition according to Wilkinson 600  600 cells. The expansion factor for radial states was the
and Allen [28]. same as for S1.
The most of the present calculations were carried out using  Unstructured triangular grid (U1) had the rectangle computa-
Stallo [29] HPC facility. This system (HP BL 460c cluster) was in- tional region of a size L  H = 6.5 m  2 m. For a consistency
stalled at the end of 2007 and has a total of 704 nodes with 5632 with the previous results, the computational domain replicated
cores. Each computer node has a 2.66 GHz quad-core Intel Xeon solutions discussed by Isaev et al. [31]. A cylinder of diameter,
processor with 16 GB of memory and 120 GB disk. The facility is D = 0.1 m was located at a distance of 17.5  D from the inlet
provided with 128 TB central storage. AF was not tested since it and symmetrically relative to the upper and bottom boundaries
was not installed at the Stallo. Instead, we used the official AF data (Fig. 3c). The obstacle was surrounded by a structured ring grid
available in the literature [30]. The best reported scalability for the or so-called a viscous boundary layer (BL hereafter) with a min-
pressure-based implicit solver implemented in AF demonstrated at imum near-the-wall step size 5  106 m (Fig. 3d). The cylin-
the 2009 Parallel CFD conference (ParCFD 2009) was about 70% up der’s surface was divided by 100 equal intervals. The cell size
to 512 cores. OF parallel performance, both in terms of strong at the outer boundaries was fixed and equal to 0.015 m, leading
(starting from p = 1) and weak scalability (from p = 8), was tested to smooth grid refinement in the vicinity of a cylinder.
at the Stallo up to p = 1024 cores. The code demonstrated quite
good weak scalability, which was achieved for a size of It should be noted that the present results were obtained for OF
35  35  35 grid points (43 K) per one core (Fig. 2a) and the mod- only. The same data for AF were reported earlier [31]. Figs. 4–6

20 103
a b
ideal
OF, N = 25
OF, N = 35

15
Strong scaling
Weak scaling

102

10
ideal
AF
101 AF
AF
5 AF, averaged
OF, 150x150x150
OF, 200x200x200
OF, 250x250x250
OF, averaged
0 100
100 101 102 103 100 101 102 103
Number of Cores Number of Cores

Fig. 2. OF parallel performance. Weak scalability tests (a) for the 3D laminar lid-driven cavity flow using 25  25  25 (h) and 35  35  35 ð}Þ grid points per one core and
an ideal case (–). Strong scalability (b) for the 3D laminar lid-driven cavity flow with different grid’s sizes: 150  150  150 ð}Þ; 200  200  200 () and 250  250  250
(h) with an ideal case (–) and AF data (+, , ) [30]. Note that OF data include I/O operations.
412 D.A. Lysenko et al. / Computers & Fluids 80 (2013) 408–422

Fig. 3. Description of the grids for the laminar flow over a circular cylinder at Re = 140: S1 (a) and U1 (c) grids and zoom of the near-the-wall regions (b and d), respectively
and a general scheme (e).

Fig. 4. Time evolution of the lift coefficient for the laminar flow over a circular cylinder at Re = 140 (1 – compressible flow, S2; 2, 3 – incompressible flow, S1 and U1
respectively).

Fig. 5. Time evolution of the drag coefficient for the laminar flow over a circular cylinder at Re = 140. For details, see the caption for Fig. 4.

show some results. The deviations between the lift coefficients, ob- grids (1–2) are quite close to each other in opposite to another
tained for the different grids and the incompressible/compressible curve (3) related with the unstructured triangular mesh. With re-
formulation, may be considered as non-significant (Fig. 4). How- spect to AF, the same trend was confirmed in the previous study
ever, the discrepancies for the drag coefficient may be associated [31] as well. The mean drag coefficient was determined in the
with distinct mesh topologies resulted in the slightly different range Cd = 1.34–1.36. These values agree with the direct numerical
pressure field prediction (Fig. 5). The signals obtained at the O-type simulation (DNS) results reported by Inoue and Hatakeyama [32]
 D.A. Lysenko et al. / Computers & Fluids 80 (2013) 408–422 413

Fig. 6. Mean pressure coefficient distribution over the cylinder’s surface for the laminar flow over a circular cylinder at Re = 140: 1–3 – present numerical results (for details,
see the caption for Fig. 4); 4 – DNS by Inoue and Hatakeyama [32]; 5 – Experiment by Grove et al. [36].

and with the data obtained by Müller [33]. The amplitudes of the Williamson [1] provided the plot of the mean base suction coef-
lift and drag coefficients are C 0l ¼ 0:48—0:53 and ficient over wide range of Reynolds numbers, as well. For Re = 140,
0
C d ¼ 0:023—0:026. These non-dimensional force amplitudes are the experimental value of Cp,b = 0.84 can be determined and cor-
in a good agreement with the values by Inoue and Hatakeyama relates well with the numerically predicted values of
[32], who report C 0l ¼ 0:52 and C 0d ¼ 0:026, respectively, for the Cp,b = 0.842–1.060. The distribution of the time-averaged pres-
inlet free stream Mach number M = 0.2. Müller [33] using a high- sure coefficient over the cylinder’s surface is presented in Fig. 6.
order finite difference method (HOFDM), got C 0l ¼ 0:5203 and The numerical results obtained for the compressible flow (1)
C 0d ¼ 0:02614, respectively. The calculated recirculation zone matched well with the DNS data by Inoue and Hatakeyama [32].
length Lr/D = 1.08–1.14 are close to the value obtained by Franke Data from incompressible flow calculations obtained at different
et al. [34] (Lr/D = 1.14). grids (2–3) are very close to each other on the one hand, and on
The computed mean separation angle hsep = 114° was in a well the other hand – close to the experimental results by Grove et al.
agreement with the hsep = 112° obtained by Franke et al. [34]. The [36]. The gap between the distributions of mean pressure coeffi-
detailed analysis of the existed experimental data for the hsep in cient (1–3) can be explained by the difference in the implementa-
the range of 10 < Re < 200 were carried out by Wu et al. [35] tion of NS equations solution algorithms in compressible and
who obtained a linear empirical equation for the hsep–Re incompressible formulations. The DNS as well as the current com-
relationship pressible simulations were carried out for the Mach number
M = 0.2. Experimentally measured values by Grove et al. [36] are
hsep ¼ 101:5 þ 155:2Re0:5 : ð1Þ close to the results obtained with the incompressible flow assump-
tion. Thus, we can expect that the last one was obtained in ambient
Thus, for Re = 140 the experimental mean separation angle was
conditions with the weak compressibility (M < 0.1). It should be
determined to hsep = 114.6°.
noticed that there is some underprediction of integral parameters
The Strouhal number was computed to St = 0.18–0.184. Inoue
and fluctuations values in the force coefficients, predicted by AF
and Hatakeyama [32] found the value of St = 0.183 in their DNS.
(Table 3). The main reason for such lower values may be treated
Müller [33] predicted the value of St 0.1831 in the similar condi-
as a result of excessive numerical scheme dissipation when the dy-
tions. Williamson [1] discussed in details the Strouhal–Reynolds
namic mesh adaptation algorithm is applied.
number relationship for the laminar shedding regime
(47 < Re < 200). Williamson [1] demonstrated the single St–Re
function with an agreement to the 1% level of St–Re relationship 4.2. Turbulent flow over a circular cylinder at Re = 3900
for laminar parallel shedding using different techniques, such as
a wind tunnel facility and a water facility known as a towing tank. The turbulent flow over a circular cylinder at Re = 3900 is prob-
The generalized St–Re curve proposed by Williamson [1] has the ably the more documented one in the literature and can be viewed
following equation: as a generic benchmark for the sub-critical regime [6]. Available
St ¼ A=Re þ B þ C
Re; ð2Þ experimental data for this particular test case cover practically
most integral (such as forces, wake dynamics, separation angle
where A = 3.3265, B = 0.1816, C = 1.6  104, which leads to the and recirculation bubble) and local (velocity, vorticity and Rey-
experimental Strouhal number, St = 0.18. nolds stresses) features of the flow allowing to assess a numerical

Table 3
Integral characteristics for the laminar unsteady flow over a circular cylinder at Re = 140.

Contributors Method M Cd C 0d C 0l St Cp,b Lr/D hsep (°)

Williamson [1] HWA, PIV 0.180 0.84

Müller [33] HOFDM 0.2 1.340 0.026 0.520 0.183
Inoue and Hatakeyama [32] DNS 0.2 0.026 0.520 0.183
Franke et al. [34] URANS 1.31 0.65 0.194 1.14 112
Isaev et al. [31] URANS 1.270 0.011 0.400 0.172
Present
U1 URANS 1.360 0.026 0.533 0.184 0.842 1.05 115
S1 URANS 1.342 0.024 0.518 0.184 0.860 1.08 114
S2 URANS 0.2 1.340 0.023 0.480 0.180 1.060 1.14 114
414 D.A. Lysenko et al. / Computers & Fluids 80 (2013) 408–422

Fig. 7. Mean pressure coefficient distribution on the cylinder’s surface at Re = 3900: h = 0° is the stagnation point, 1 – experiment by Norberg [4], 2, 3 – present results (OF and
AF, respectively).

method qualitatively and quantitatively. In spite of some experi- and AF data. The detailed analysis of the flow-field in the near
mental difficulties (the presence of the recirculation zone and high wake calculated by AF revealed the existence of two small coun-
instantaneous flow angles), there are several hot-wire anemometry ter-rotating vortices (beside the main recirculation bubble) at-
(HWA) and particle image velocimetry (PIV) measurements avail- tached to the backward side of the cylinder. This finding means
able in the near wake of the circular cylinder. The pioneer work that several separation angles (beside the primary hsep) existed in
was done by Ong and Wallace [5], who managed to accurately the AF solution. Breuer [39] got the same results in LES study, using
measure velocity and vorticity vectors in the near wake outside the HYBRID (a combination of the upwind and the central differ-
the recirculation bubble and proposed turbulence statistics and encing approximations) and HLPA (hybrid linear/parabolic approx-
power spectra of the stream-wise and normal velocity components imation) schemes. From the experimental point of view, the
at several locations. To avoid the restrictions associated with the existence of the small vortices at the backward side of the cylinder
presence of back flow, the techniques of particle image (PIV) or la- was confirmed by Son and Hanratty [41]. In Fig. 8 the mean vortic-
ser Doppler (LDV) velocimetry are more appropriate [6]. One of the ity magnitude
pffiffiffiffiffi
ffi (which was non-dimensionalized with the factor
first PIV works for the flow over a circular cylinder at Re = 3900 2 Re, according to Ma et al. [42]) is plotted along with experimen-
was the study of Lourence and Shih [2], who performed time re- tal data. The agreement between the numerical and experimental
solved measurements in the recirculation region. Statistical quan- results was fairy well starting from the stagnation point (hsep = 0°)
tities were assessed even though this PIV experiment was not til the prime separation angle (hsep 90°). As known from mea-
designed for this purpose. Nevertheless, these results are often surements [2], the separation should take place at hsep = 86°. This
used as reference for validation of numerical simulations in the lit- finding was confirmed by Son and Hanratty [41], however it is
erature (e.g., [37–39], etc.). Dong et al. [40] investigated the near worth to notice that their experiment was conducted for
wake with PIV and direct numerical simulation (DNS) at Re = 5000. The calculated hsep = 88.8° by OF and hsep = 89° by AF
Re = 3900/4000 and 10000. The main emphasis was drawn on an were slight higher the experimental one. At the backward side of
investigation of the shear-layer instability using DNS data, but a the cylinder the OF solution was quite similar to the experimental
PIV/DNS comparison of mean and turbulent isocontours PIV/DNS data [41], however it is did not represent secondary vorticity struc-
maps was discussed also. Recently, the flow over a circular cylinder tures besides the main recirculation zone. The length of the sec-
was studied by Parnaudeau et al. [6] in the near wake at Re = 3900 ondary separation zone predicted by AF was h 18° while the
both numerically (LES) and experimentally with PIV and HWA measured value by Son and Hanratty [41] was h 35°.
methods. The computed mean drag coefficient Cd = 0.98–1.07 and the
The present results were obtained using O-type mesh (S2 from back-pressure coefficient Cp,b = 0.99–1.16 were not much too
the previous subsection) with the compressible flow assumption high ( 10%) compared with experimental measurements of Nor-
using both OF and AF. Two supplementary O-type grids with the berg [4], who measured Cd = 0.98 and Cp,b = 0.90 respectively.
spatial resolution 200  200 and 400  400 were built to confirm The computed values of the Strouhal number of the vortex shed-
the grid convergence. ding frequency St = 0.215–0.216 were found to be within the
The distribution of the mean pressure coefficient on the cylin- experimental range of St = 0.20–0.22 (Table 4).
der’s surface is plotted in Fig. 7. The calculated data by OF was in Fig. 9 compares mean stream-wise velocity hui/U1 in the wake
a good agreement with the measured values provided by Norberg centerline with the experiments of Lourenco and Shih [2], Parnau-
[4]. However there are some deviations between the experimental deau et al. [6] and Ong and Wallace [5]. The mean stream-wise

Fig. 8. Mean vorticity magnitude distribution on the cylinder surface at Re = 3900: h = 0° is the stagnation point, 1 – experiment of Son and Hanratty [41]; 2, 3 – present
results (OF and AF, respectively).
 D.A. Lysenko et al. / Computers & Fluids 80 (2013) 408–422 415

Table 4
Overview of the experimental works and present URANS results of the circular cylinder flow at Re = 3900.

Contributors Method M Cd St Cp,b Lr/D humini/U1 hsep (°)

Cardell [3] HWA 0.215 1.33 0.25
Lourenco and Shih [2] PIV 0.99 0.22 1.19 0.24 86
Norberg [4] HWA 0.98 0.90
Ong and Wallace [5] HWA 0.21
Ma et al., Case II [42] DNS 0.84 0.22 1.59
Dong et al. [40] PIV 0.20 1.36–1.47 0.252
Dong et al. [40] DNS 0.20 1.41–1.59 0.291
Parnaudeau et al. [6] PIV 0.21 1.51 0.34
Present
S2 URANS-LSKE 0.2 1.07 0.22 1.16 1.10 0.19 89
S2 URANS-RKE 0.96 0.19 0.97 1.11 0.20 90
S2 URANS-RKE 0.2 0.98 0.20 0.99 1.03 0.20 89

Fig. 9. Mean stream-wise velocity in the wake centerline for the flow over a circular cylinder at Re = 3900 (Experiment: 1 – Lourenco and Shih [2], 2 – Parnaudeau et al. [6], 3
– Ong and Wallace [5]; 4, 5 – present results (OF and AF, respectively)).

Fig. 10. Mean stream-wise velocity hui/U1 at the different locations (x/D = 1.06, 1.54, 2.02, 4.00, 7.00) in the wake of a circular cylinder at Re = 3900. For details, see the
caption for Fig. 9.

velocity is zero at the base of the cylinder due to no-slip condition. AF, but in general our calculations agrees well with the experiments
It reaches a negative minimum humini in the recirculation zone and by Lourenco and Shih [2] and Ong and Wallace [5].
converges asymptotically toward U1. The present results predicted In Fig. 11 the Reynolds normal stresses hu0 u0 i=U 21 are shown. At
humini/U1 = 0.19 to 0.2, which are in reasonable agreement with x/D = 1.06, the hu0 u0 i-profile presents two strong peaks mainly due
the experimental data by Cardell [3], Lourenco and Shih [2] and to the transitional state of the shear layers, which show a flapping
Dong et al. [40] who reported humini/U1 = 0.24 to 0.251. The cal- behavior due to primary vortex formation [6]. The position of these
culated recirculation zone length Lr/D = 1.03–1.1 are close to the two peaks agrees with the experiment of Lourenco and Shih [2]. At
experimental value of Lourence and Shih [2] (Lr/D = 1.19) and x/D > 1.06, the two peaks of the shear layers are overlapped by two
DNS (Lr/D = 1.12), performed by Ma et al. [42]. Thus, the present re- larger peaks due to primary vortex formation. Overall, the good
sults are closer to the PIV data by Lourenco and Shih [2] and the agreement is observed between the current calculations both for
HWA data by Ong and Wallace [5]. OF and AF and the experimental data of Lourenco and Shih [2]
Fig. 10 shows velocity profiles of hui/U1. A strong velocity deficit and Ong and Wallace [5]. However, it should be noticed that
occurs in the region of the recirculation bubble. The mean velocity hu0 u0 i-profiles by AF is slightly overshoot the ones by OF, as well
profiles shows a U-shape close to the cylinder which involves to- as the experimental data.
wards a V-shape further downstream. One can observe some non- Table 4 summaries the main integral flow parameters from the
significant discrepancies between the present results for OF and experimental and DNS data and the present results.
416 D.A. Lysenko et al. / Computers & Fluids 80 (2013) 408–422

Fig. 11. Variation of the mean stream-wise velocity hu0 u0 i=U 21 at the different locations (x/D = 1.06, 1.54, 2.02, 4.00, 7.00) in the wake of a circular cylinder at Re = 3900. For
details, see the caption for Fig. 9.

4.3. Turbulent flow over a triangular rig in a channel – Fujii lab test triangular rod. All these features provided good near-the-wall
mesh resolution and allowed to apply the low-Reynolds-num-
LDV measurements by Fujii et al. [8,7] were carried out in an ber k   turbulence model.
open circuit, forced flow type of wind tunnel at ambient conditions  Baseline unstructured quadrilateral mesh (Fig. 12d) was
(P1 100 kPa, T1 = 280 K, Re = 17500 and M 0.03). A sketch is designed in the same manner (UQ1).
presented in Fig. 12a. An equilateral (D = 0.025 m) triangular rod  To check solution mesh independence, two additional grids
was placed inside the channel passage of 0.05 m-square cross- were built, both with the quad cells (UQ0 and UQ2). The type
section. of the attached viscous BLs was the same as in the baseline quad
The two-dimensional computational domain (Fig. 12b) was de- mesh and domain was meshing with sizes of 0.0015 m and
signed to replicate lab test conditions. The channel length and 0.0005 m, respectively.
height were set to L = 0.305 m and H = 0.05 m, respectively. The
bluff-body was located at x = 0.117 m from the channel inlet. Time-averaged measured and numerically predicted stream-
Two additional buffer domains with length 0.05 m and height lines are presented in Fig. 13. The results of the CFD analysis pro-
0.1 m were attached to the channel. Note, that this problem was vide a more extended length of the reversed zone, Lr/D = 2.67
calculated by OF only. Two finite-element baseline grids with dif- compared to the measured one, Lr/D = 2.2. Time-averaged pressure
ferent cell types were used (the detailed description is provided coefficient distribution downstream the bluff-body inside the
in [27]): recirculation bubble is shown in Fig. 14. Numerical results demon-
strate quite similar behavior with the experimental data. Minimum
 Unstructured triangular mesh (Fig. 12c). Viscous BLs were values were also in a good agreement between numerical
attached to the obstacle with the following parameters: the first (Cp,min = 2.7) and measured (Cp,min = 2.73) data.
row size 5  105 m, the growth factor 1.25 and the number of The normalized turbulence kinetic energy in the wake is pro-
rows 7. The same (except the growth factor) viscous BL was also vided in Fig. 15. Overall, the good qualitative and quantitative
applied for the channel walls. Each edge of the bluff-body was agreement for the K between lab test and numerical modeling data
divided into 80 equal intervals. The computational domain was achieved.
was meshed with the size of 0.001 m, which guarantees smooth The time history of the instantaneous transverse velocity are
triangular element distribution from the inlet and outlet to the presented in Fig. 16. As expected, periodic sinusoidal type signals

Fig. 12. A general view of the experimental test rig (a) taken from [7], the computational domain (b) and the fragments of the designed unstructured viscous grids (c and d) at
the vicinity of the bluff-body.
 D.A. Lysenko et al. / Computers & Fluids 80 (2013) 408–422 417

4.4. Turbulent flow over a triangular rig in a channel – Volvo test rig

Fig. 17a shows a schematic drawing of the test section. The test
set-up consisted of a straight channel with a rectangular cross-sec-
tion, divided into an inlet section length 0.5 m and a channel pas-
sage section length L = 1 m and 0.12 m  0.24 m cross-section. The
inlet section was used for flow straightening and turbulence con-
trol. The air entering the inlet section was distributed over the
cross-section by a critical plate that, at the same time, isolated
Fig. 13. Time-averaged streamlines for the flow past a triangular rod in a channel at the channel acoustically from the air supply system. The channel
Re = 17500: a – experiment by Fujii et al. [7,8], b – present results (UQ1). passage section ended in a circular duct with a large diameter.
The triangular bluff-body (with base diameter, D = 0.04 m) was
that were obtained clearly illustrate self-regular vortex shedding mounted with its reference position 0.681 m upstream of the chan-
behavior of the wake. The averaged calculated Strouhal number nel exit.
was, St = 0.44, with the corresponding main frequency, The cold flow measurements were conducted in the ambient
f = 177 Hz. This is in a quite good agreement with the experimental conditions (T1 = 288 K, P1 = 100 kPa, Re = 45000, and M = 0.05).
values, where a pronounced frequency, f = 160 Hz, and the corre- Honeycombs and screens controlled the approximate inlet turbu-
sponding Strouhal number St = 0.4, was detected. lence level of 3–4%. A two-component LDA system was used for
Table 5 summaries the integral flow features for this problem. the stream-wise and transverse velocity components and its

Fig. 14. Mean pressure coefficient in the wake centerline for the flow over a triangular rod in a channel at Re = 17500: 1 – experiment by Fujii et al. [7,8], 2 – present results
(UQ1).

Fig. 15. Normalized turbulence kinetic energy in the wake centerline for the flow over a triangular rod in a channel at Re = 17500: 1 – experiment by Fujii et al. [7,8], 2 –
present numerical data (UQ1).

Fig. 16. Time history of the instantaneous transverse velocity at the designed grids UQ0–UQ2 (1–3, respectively). The reference point was located at the same position as in
the lab test of Fujii et al. [7,8] with the coordinates: x/D = 1.2, y/D = 0.6.
418 D.A. Lysenko et al. / Computers & Fluids 80 (2013) 408–422

Table 5
Integral parameters for Fujii lab test simulations.

Contributors Method M Cp,min Lr/D St

Fujii et al. [8,7] LDV 2.73 2.2 0.40
Present
UQ1 URANS-LSKE 2.7 2.67 0.45
UQ1 URANS-LSKE 0.03 2.65 2.87 0.45

Fig. 17. The sketch of the Volvo test rig (a) taken from Ref. [9] and the general view of the computational domain (b).

Fig. 18. Mean stream-wise velocity (1,3,4) and normalized turbulence kinetic energy (2,5,6) in the wake centerline for the flow over a triangular rod in a channel at
Re = 45000: 1, 2 – experiment by Sjunnesson et al. [9]; 3, 5 – present results, OF; 4,6 – present results, AF.

Fig. 19. Mean stream-wise velocity hui/U1 at the different locations (x/D = 2.5, 0.375, 1.525, 3.75, 9.4) in the channel flow with a triangular rod at Re = 45000: 1 –
experiment by Sjunnesson et al. [9]; 2, 3 – present results (OF and AF, respectively).

fluctuations measurements. Further detailed description of the (with a detailed description provided in [27]): quadrilateral, par-
LDA system, experimental procedure and sampling technique can tially triangular UT1 (Fig. 17b) and two supplementary partially
be found in [9]. triangular grids (UT0, UT2) to check grid convergence. The results
The computational grids were built in the same spirit as for the presented in this section were obtained using UT1 for both OF
previous test problem. Four unstructured grids were designed and AF.
 D.A. Lysenko et al. / Computers & Fluids 80 (2013) 408–422 419

Fig. 20. Normalized turbulence kinetic energy K at the different locations (x/D = 2.5, 0.375, 1.525, 3.75, 9.4) in the channel flow with a triangular rod at Re = 45000: 1 –
experiment by Sjunnesson et al. [9]; 2, 3 – present results (OF and AF, respectively).

Fig. 21. Flow visualization of the vortex shedding behind a triangular rod: a – the experimental interferogram taken from Nakagawa [46]; b – numerical interferogram (OF).

Fig. 22. Mean surface pressure coefficient distribution on the sides of a triangular rod: 1 – experiment by Tatsuno et al. [48], 2, 3 – present results (OF and AF, respectively).

Figs. 18–23 represent some results. As was reported by Sjunnes- the central-line behind the obstacle. It should be noticed a 18%
son et al. [9], the experimental profile of stream-wise velocity com- undershoot compared to the experimental data for the humini/U1
ponent was slightly skewed due to small misalignment in the in the recirculation zone independently of the grid type and
flange between the inlet section and the channel passage. How- numerical methods. But, overall, there is a good match between
ever, fully symmetrical inlet velocity profile was implicitly formed numerical and experimentally measured data. For example, the
during flow development in the numerical simulations. In both same level of undershoot was observed by Hasse et al. [43] for UR-
cases one could observe the outer regions of BLs at the channel ANS (with k  x SST) data. Present numerical results showed good
walls, which corresponded very well with the ‘1/7 power law pro- prediction for the recirculation lengths: calculated Lr/D = 1.36 (OF)
file’ typical for fully developed turbulent channel flows. and Lr/D = 1.32 (AF) in comparison with the measured [9] value of
Fig. 18 shows the measured and predicted mean stream-wise Lr/D = 1.33 and with the numerically predicted value of Lr/D = 1.3
velocity hui/U1 and normalized turbulence kinetic energy K along by Durbin [44] and Johansson et al. [45]. The recirculation zone
420 D.A. Lysenko et al. / Computers & Fluids 80 (2013) 408–422

Fig. 23. Time history of the instantaneous transverse velocity: 1–3 – OF (UT0-UT2), respectively; 4 – AF (UT1). The reference point was located at the same position as in the
lab test of Sjunnesson et al. [9] with the following coordinates: x/D = 1, y/D = 0.

lengths obtained by Hasse et al. [43] for URANS-SST and detached Table 6
eddy simulation (DES) were only 0.94 and 1.18, respectively. In Integral parameters for Volvo lab test simulations.
Fig. 18 the comparison between measured and numerically pre-
Contributors Method M Lr/D St DP
dicted normalized turbulence kinetic energy in the wake is shown
Sjunnesson et al. [9] LDA 1.33 0.250 0.6
as well. One can observe the same trend between experimentally
Johansson et al. [45] URANS-SKE 1.30 0.270
measured values and the data, obtained by numerical modeling, Durbin [44] URANS-KEV2 1.30 0.285
since the latter significantly underpredict the level of fluctuations Strelets [49] URANS-SA 0.90
inside the recirculation zone. This indicates that the vortex shed- Hasse et al. [43] URANS-SST 0.94 0.296
ding was much stronger in the physical experiment. Present
UT1 URANS-LSKE 1.57 0.28 0.7
Figs. 19 and 20 show the mean stream-wise velocity and nor-
UT1 URANS-LSKE 0.05 1.36 0.28 0.7
malized turbulence kinetic energy at several locations in the chan- UT1 URANS-RKE 0.05 1.3 0.28 0.7
nel. Overall, there are minimal discrepancies between the
numerical data predicted by OF and AF, as well as the experimental
data by Sjunnesson et al. [9]. At the same time, as was mentioned
sure distribution at the surface of a prism yet. Here, the experiment
already, a strong underprediction of numerically predicted turbu-
of Tatsuno et al. [48] was adopted (Fig. 22). One should notice sat-
lence kinetic energy was observed inside the recirculation zone.
isfied agreement between the present numerical and experimental
Meanwhile, the same level of turbulence kinetic energy was
results although the measurements were carried out in the slight
contained both in the downstream of the bluff-body and in the up-
different conditions (Re = 9  104).
stream part of channel, after the separation bubble was vanished.
Finally, Fig. 23 shows the time history of the transverse velocity
A numerically predicted frequency of von Kármán vortex
measured at the same probe location using OF and AF. The varia-
shedding matched quite well to experimental data: f = 117 Hz
tion between the curves are quite small (about ±5% both for the
(OF) and f = 118 Hz (AF) vs. f = 105 Hz in experimental data, and
frequencies and amplitudes). Table 6 summaries the main integral
St = 0.28 vs. St = 0.25, respectively, which are differences of about
flow features available in the literature for this problem and the
10%. These results correspond well also with those obtained by
present results.
Hasse et al. [43]: f = 122 Hz and f = 117 Hz peak frequencies for
URANS and DES, respectively.
Fig. 21 shows the experimental [46] and numerical interfero- 4.5. Discussion
grams of the flow. The visualization of the instantaneous density
field was done using numerical interferogram technique according Several well referenced turbulent separated bluff-body prob-
to Hadjadj and Kudryavtsev [47]. In spite of the slight different lems were analyzed numerically with the conventional URANS ap-
flow conditions between the experiment by Nakagawa [46] proach based on two state-of-the art CFD technologies (OpenFOAM
(M = 0.377 and Re = 1.73  105) and the present calculations, the and ANSYS FLUENT).
agreement between them was fairly well. The shear layers sepa- The influence of ‘discretized’ errors on the solutions were vali-
rated from the upper and lower trailing edges delineate the bound- dated and confirmed by the independence from grids topology
ary of the vortex formation region behind the triangular cylinder. and applied numerical schemes and factors (CFD codes). For all
The main vortices are generated alternately at the upper and lower present results carried out with OF and AF, the same computa-
trailing edges, and subsequently shed downstream. It is worthy to tional grids were used for the each test problem. Preliminary re-
note that no secondary vortices are generated around a triangular sults [27] showed that the influence of the grids topology could
rod. be considered as non-significant (the difference between results
Although the accuracy of the pressure loss measurements was were less than 5%). To check the mesh independence the grid con-
not too high, as it was mentioned in [9], it is still of interest in vergence study was performed for the most principal flow param-
many engineering applications. The experimental value for the eter such as the recirculation zone bubble behind a bluff-body. As
normalized pressure drop was measured (between axial states was discussed, the test cases were calculated using the compress-
x/D = 5 and x/D = 10) to DP = 0.6 while the numerically predicted ible formulation. Since a limited number of the problem-related
value was DP 0.7 (both for OF and AF), which correspond quite articles were found in the literature, and all cited numerical anal-
well between each other. Despite the fact that many researchers ysis were carried out with incompressible flow assumption, we
have used this test for validation, nobody has not compared pres- can regard that these data can be considered as the first one where
 D.A. Lysenko et al. / Computers & Fluids 80 (2013) 408–422 421

a compressible URANS approach has been applied for the selected assumption strongly depends from M, e.g., the Mach number should
test cases. not exceed 0.14 to achieve 1% error. However, the present results
The influence of the ‘temporal discretization’ errors was not indicated that the maximum local Mach number values were 0.3–
analyzed in the present study. Usually the spatial discretization er- 0.4. These high Mach number areas were located just behind a
ror effect is larger than the error arising from time integration [15]. bluff-body and delineated from a vortex formation region by sepa-
It may be shown that for fully developed turbulent flows the impli- rated sheared layers. Thus, a recirculation zone region was cor-
cit schemes appear less efficient compared to explicit methods rected by the compressibility effect.
[15]. However, the main advantage of the adopted implicit BDF2 To assess local compressibility effects we compared the local
scheme is a larger stability region compared to the explicit meth- mean stream-wise velocity (Fig. 24) and normalized turbulence ki-
ods, which is important in case of combustion applications. More- netic energy (Fig. 25) for the compressible and incompressible
over, in case of fully developed turbulent flows existed small time flows over a triangular rod in a channel at Re = 45000 (Volvo test
and space scales are simply advected by the most energetic eddies case). One can observe some expected deviations between solu-
[15]. This argument yields an accuracy time-scale similar to the tions in the reverse zone area for the mean stream-wise velocity
CFL criterion. Thus in all present calculations the stability condition due to the difference in the predictions of the recirculation bubble.
CFL < 1 was employed which had to guarantee that actual time- The normalized turbulence kinetic energy had the same trends
step was close to accuracy time step. however the incompressible solution under-shoot the compress-
To check solutions sensitivity to the compressible effects ible one, which in its turn under-shoot the experimental one. This
several additional runs were conducted within the incompressible is interesting finding since the inlet values of the turbulence inten-
URANS approach. Preliminary results [27] indicated first of all, that sity were the same as in the experiment for both numerical cases.
there were not significant deviations for the selected test cases Two different implementations of the k   turbulence model
between the compressible and incompressible runs for the main were used to check the ‘modeling’ error influence. As it was shown
integral flow features except the recirculation zone length. The both versions of this turbulence model were capable to predict the
reverse zone length could be varied significantly in a range of flow physics reasonably well. It is worth to notice, that no wall
±25% for all selected benchmarks, that can be observed from Tables functions were used in any of simulations. The difference between
3–6. The minimum deviation ( 5%) is observed for the laminar the principal flow characteristics obtained with RKE and SHKE
flow over a circular cylinder at Re = 140. models was quite acceptable and did not exceed 10%. The only
One of the possible explanations can be the following. From the one issue was identified concerning the formation of the small vor-
isentropic theory the local and total densities can be related tices at the backward side of the cylinder’s surface besides the pri-
mary recirculation bubble for the flow over a circular cylinder at
 /1
1
/1 2 Re = 3900. However, as was mentioned by Breuer [39], the struc-
qo =q ¼ 1 þ M : ð3Þ
ture and the length of the recirculation bubbles behind the cylinder
2
were strongly influenced by the numerical scheme. In our case the
It may be shown that for the subsonic flows (M < 1): q/q0 = 1  OF was not able to predict the secondary vortices contrary to the
M2/2. The deviation in a solution using the incompressible AF.

Fig. 24. Mean stream-wise velocity hui/U1 at the different locations (x/D = 2.5, 0.375, 1.525, 3.75, 9.4) in the channel flow with a triangular rod at Re = 45000: 1 –
experiment by Sjunnesson et al. [9]; 2, 3 – present compressible and incompressible results (OF, UT1), respectively.

Fig. 25. Normalized turbulence kinetic energy K at the different locations (x/D = 2.5, 0.375, 1.525, 3.75, 9.4) in the channel flow with a triangular rod at Re = 45000: 1 –
experiment by Sjunnesson et al. [9]; 2, 3 – present compressible and incompressible results (OF, UT1), respectively.
422 D.A. Lysenko et al. / Computers & Fluids 80 (2013) 408–422

5. Concluding remarks [18] Hestens M, Steifel E. Methods of conjugate gradients for solving systems of
algebraic equations. J Res Nat Bur Stand 1952;29:409–36.
[19] Jacobs D. Preconditioned conjugate gradient methods for solving systems of
The objective of the present work was an extensive investiga- algebraic equations. Tech. rep., Central Electricity Research Laboratories,
tion on the numerical aspects influencing the quality of compress- Leatherhead, Surrey, England; 1980.
[20] Cotton M, Kirwin P. A variant of the low-Reynolds-number two-equation
ible URANS solutions, implemented in the OpenFOAM toolbox.
turbulence model applied to variable property mixed convection flows. Int J
With this purpose several well-referenced turbulent massively Heat Fluid Flow 1995;16(6):486–92.
separated bluff-body benchmarks were selected and analyzed [21] Hutchinson B, Raithby G. A multigrid method based on the additive correction
strategy. J Numer Heat Transfer 1986;9:511–37.
numerically. The commercial code ANSYS FLUENT was used to pro-
[22] Rhie C, Chow W. Numerical study of the turbulent flow past an airfoil with
vide detailed face-to-face results analysis with OpenFOAM. Com- trailing edge separation. AIAA J 1983;21(11):1525–32.
parison of the numerical estimates using the OpenFOAM and [23] Wolfshtein M. The velocity and temperature distribution of one-dimensional
ANSYS FLUENT technologies and the most frequently applied in flow with turbulence augmentation and pressure gradient. Int J Heat Mass
Transfer 1969;12:301–18.
engineering semi-empirical differential models of turbulence, has [24] Chen H, Patel V. Near-wall turbulence models for complex flows including
shown that they are close and are practically independent of the separation. AIAA J 1988;26(6):641–8.
grid used. The parallel performance of OpenFOAM was analyzed [25] Jongen T. Simulation and modeling of turbulent incompressible flows. Ph.D.
thesis, EPT Lausanne, Lausanne, Switzerland; 1992.
up to 1024 cores and shown the modest parallel efficiency of [26] Poinsot T, Lele S. Boundary conditions for direct simulations of compressible
40%. In general, the URANS results agreed fairly well with the viscous flows. J Comput Phys 1992;101:104–29.
experimental data and other numerical solutions. All these indi- [27] Lysenko D, Ertesvag I, Rian K. Turbulent bluff body flows modeling using
OpenFOAM technology. In: Skallerud B, Andersson HI, editors. Computational
cated on the adequacy and accuracy of the established numerical mechanics, Trondheim; 2011. p. 189–208.
method based on the OpenFOAM and its readiness for further com- [28] Wilkinson B, Allen M. Parallel programming: techniques and applications
bustion application development. using networked workstations and parallel computers. 2-nd ed. Prentice Hall;
2004.
[29] <http://www.notur.no/hardware/stallo/>.
Acknowledgment [30] <http://www.ansys.com/Support/Platform+Support/Benchmarks+Overview>.
[31] Isaev S, Baranov PA, Kudryavtsev NA, Lysenko DA, Usachev AE. Comparative
analysis of the calculation data on an unsteady flow around a circular cylinder
This work was conducted as a part of the CenBio Center for
obtained using the VP2/3 and FLUENT packages and the Spalart-Allmaras and
environmentally-friendly energy. We are very appreciated to the Menter turbulence models. J Eng Phys Thermophys 2005;78(6):1199–213.
Norwegian Meta center for Computational Science (NOTUR) for [32] Inoue O, Hatakeyama N. Sound generation by a two-dimensional circular in a
providing the uninterrupted HPC computational resources and uniform flow. J Fluid Mech 2002;471:285–314.
[33] Müller B. High order numerical simulation of aeolian tones. J Comput Fluids
the useful technical support. 2008;37:450–62.
[34] Franke R, Rodi W, Schönung B. Numerical calculation of laminar vortex
References shedding flow past cylinders. J Wind Eng Ind Aerodyn 1990;35:237–57.
[35] Wu M-H, Wen C-Y, Yen R-H, Weng M-C, Wang A-B. Experimental and
numerical study of the separation angle for flow around a circular cylinder at
[1] Williamson C. Vortex dynamics in the cylinder wake. Ann Rev Fluid Mech
low Reynolds number. J Fluid Mech 2004;515:233–60.
1996;28:477–539.
[36] Grove A, Shair F, Petersen E, Acrivoss A. An experimental investigation of the
[2] Lourenco L, Shih C. Characteristics of the plane turbulent near wake of a
steady separated flow past a circular cylinder. J Fluid Mech 1964;19:60–80.
circular cylinder, a particle image velocimetry study. Tech. rep., Published in
[37] Beaudan P, Moin P. Numerical experiments on the flow past a circular cylinder
[37]; 1993.
at sub-critical Reynolds number. Tech. rep., Technical report TF-62, CTR
[3] Cardell G. Flow past a circular cylinder with permeable splitter plate. Tech.
Annual Research Briefs, NASA Ames/Stanford University; 1994.
rep., Data taken from Mittal [38]; 1993.
[38] Mittal R. Progress on les of flow past a circular cylinder. in: Annual research
[4] Norberg C. Experimental investigation of the flow around a circular cylinder:
briefs. Tech. rep., Center of Turbulence Research, Stanford University; 1996.
influence of aspect ratio. J Fluid Mech 1994;258:287–316.
[39] Breuer M. Large eddy simulation of the sub-critical flow past a circular
[5] Ong L, Wallace J. The velocity field of the turbulent very near wake of a circular
cylinder: numerical and modeling aspects. Int J Numer Methods Fluids
cylinder. Exp Fluids 1996;20:441–53.
1998;28:1281–302.
[6] Parnaudeau P, Carlier J, Heitz D, Lamballais E. Experimental and numerical
[40] Dong S, Karniadakis G, Ekmekci A, Rockwell D. A combined direct numerical
studies of the flow over a circular cylinder at Reynolds number 3900. Phys
simulation particle image velocimetry study of the turbulent air wake. J Fluid
Fluids 2008;20(8):085101.
Mech 2006;569:185–207.
[7] Fujii S, Gomi M, Eguchi K. Cold flow tests of a bluff body flame stabilizer. ASME
[41] Son J, Hanratty T. Velocity gradients at the wall for flow around a cylinder at
J 1978;100:323–32.
Reynolds numbers from 5  103 to 105. J Fluid Mech 1969;35(2):353–68.
[8] Fujii S, Eguchi K. A comparison of cold and reacting flows around a bluff body
[42] Ma X, Karamanos G-S, Karniadakis G. Dynamics and low-dimensionality of a
flame stabilizer. ASME J 1981;103:328–34.
turbulent near wake. J Fluid Mech 2000;410:29–65.
[9] Sjunnesson A, Nelsson C, Erland M. LDA measurements of velocities and
[43] Hasse C, Sohm V, Wetzel M, Durst B. Hybrid URANS/LES turbulence simulation
turbulence in a bluff body stabilized flame. Tech. rep., Volvo Flygmotor AB,
of vortex shedding behind a triangular flameholder. J Flow Turbul Combust
Trollhättan, Sweden; 1991.
2009;83:1–20.
[10] Weller H, Tabor G, Jasak H, Fureby C. A tensorial approach to computational
[44] Durbin P. Separated flow computations with the k    v2 model. AIAA J
continuum mechanics using object-oriented techniques. J Comput Phys
1995;33:659–64.
1998;12(6):620–31.
[45] Johansson S, Davidson L, Olsson E. Numerical simulation of vortex shedding
[11] ANSYS FLUENT R12. Theory guide. Tech. rep., Ansys Inc.; 2009.
past triangular cylinders at high Reynolds number using a k   turbulence
[12] Chorin A. Numerical solution of Navier–Stokes equations. J Math Comput
model. Int J Numer Methods Fluids 1993;16:859–78.
1968;22:745–62.
[46] Nakagawa T. Vortex shedding mechanism from a triangular prism in a
[13] Launder B, Sharma B. Application of the energy-dissipation model of
subsonic flow. J Fluid Dyn Res 1988;5:69–81.
turbulence to the calculation of flow near a spinning disc. J Lett Heat Mass
[47] Hadjadj A, Kudryavtsev A. Computation and flow visualization in high-speed
Transfer 1974;1:131–8.
aerodynamics. J Turbul 2005;6(16):33–81.
[14] Shih T-H, Liou W, Shabbir A, Yang Z, Zhu J. A new k   eddy-viscosity model
[48] Tatsuno M, Takayama T, Amamoto H, Ishi-i K. On the stable posture of a
for high Reynolds number turbulent flows – model development and
triangular or a square cylinder about its central axis in a uniform flow. J Fluid
validation. J Comput Fluids 1995;24(3):227–38.
Dyn Res 1990;6:201–7.
[15] Geurts B. Elements of direct and large-eddy simulation. R.T.Edwards; 2004.
[49] Strelets M. Detached eddy simulation of massively separated flows. In: AIAA
[16] Issa R. Solution of the implicitly discretized fluid flow equations by operator
aerospace sciences meeting and exhibit, January 8–11; 2001/Reno, NV; 2001.
splitting. J Comput Phys 1986;62:40–65.
[50] Jones W, Launder B. The prediction of laminarization with a two-equation
[17] Jasak H, Weller H, Gosman A. High resolution NVD differencing scheme for
model of turbulence. Int J Heat Mass Transfer 1972;15:301–14.
arbitrarily unstructured meshes. Int J Numer Methods Fluids 1999;31:431–49.