Professional Documents
Culture Documents
Numerical Analysis of The Flow Around A Circular
Numerical Analysis of The Flow Around A Circular
To cite this article: Jie Shao & Chao Zhang (2006) Numerical analysis of the flow around a circular
cylinder using RANS and LES, International Journal of Computational Fluid Dynamics, 20:5,
301-307, DOI: 10.1080/10618560600898437
Department of Mechanical and Materials Engineering, The University of Western Ontario, London, Ont., Canada N6A 5B9
The present study is to simulate the flow past a circular cylinder at a Reynolds (Re) number of 5800,
which is based on free-stream velocity and the cylinder diameter. The cylinder is slightly heated and the
amount of heat is small enough to be considered as a passive scalar. Due to its complexity, the flow
around a circular cylinder is considered as a challenging problem for computational fluid dynamics
(CFD) simulation. Re-averaged Navier –Stokes (RANS) equations and large eddy simulation (LES) are
two commonly used approaches in turbulent flow simulation. In this study, these two methods are both
investigated by employing a CFD software called FLUENT. For two-dimensional (2D) simulation, the
renormalization group k– 1 model is used with enhanced wall treatment. Moreover, 2D LES is also
tested, which reveals the necessity for three-dimensional (3D) LES computations. For 3D simulations,
computations with the Smagorinsky – Lilly subgrid-scale (SGS) model and dynamic SGS model are
used. A phase-averaging technique is employed to study turbulence structure in the circular cylinder
wake. An instantaneous quantity is decomposed into a time-mean component, a coherent component
and an incoherent component (Reynolds and Hussain 1972). After the triple decomposition and
structural averaging, the coherent contributions to the Reynolds stresses and temperature variance can
be analyzed. The reference phase for phase averaging is calculated for the time history of the lift
coefficient CL. Both velocity field and temperature field are investigated and compared with the
experimental measurements.
Smagorinsky – Lilly model and dynamic model, were (RANS) equations and LES are currently two commonly
investigated, respectively. In the near wake, it showed that used approaches in turbulent flow simulations.
the dynamic model predicted mean velocities and
Reynolds stresses better than the other two cases, i.e.
2.1 RANS
without a SGS model and with the Smagorinsky – Lilly
model. Beyond this region, in the intermediate wake, the The RANS equations solve the Reynolds equation for
differences between the results computed with or without a mean flow quantities only. Thus, the computational
SGS model became smaller. Ong and Wallace (1996) expense is reduced. RANS is widely used for practical
conducted an experiment at the same Re to validate the engineering calculations. The key point to close the RANS
LES simulation carried out by Beaudan and Moin (1994). equations appropriately is the modeling of the Reynolds
The comparison revealed that the LES simulation over- stress, 2rui uj . The Boussinesq hypothesis is adopted in
predicted the velocity deficit at the centerline. As regards k –1 models,
the velocity statistic, the LES simulation had a good
agreement with experimental result in the near wake 2
tij ¼ 2nt Sij 2 kdij
(x/D ¼ 4). However, in the intermediate wake, for x/D ¼ 7 3
and 10, LES simulation underpredicted the Reynolds
stresses. Breuer (1998) studied the numerical and nt ¼ C m k 2 =1
modeling aspects of the LES for the circular cylinder
wake at Re ¼ 3900. Influences of SGS modeling, where nt, is the turbulent kinematic eddy viscosity; k, is
discretization scheme and grid resolution were investi- the turbulence kinetic energy; 1 is the turbulent kinetic
gated. The results showed that the simulation using energy dissipation rate; and Sij, is the mean strain-rate
dynamic model and central differencing scheme gave the tensor. For a standard k –1 model, two additional transport
best agreement with the experimental data. Circular equations need to be solved, and nt is computed as a
cylinder wake with a high Re ¼ 140,000, was simulated by function of k and 1. In the current study, RNG k– 1 model
Breuer (2000). It revealed that the deviations between the is employed. This model is derived using renormalization
results from different SGS models increased with higher group theory (Fluent 2005). It has additional terms in
spanwise resolution. Nevertheless, doubling the domain in the transportation equations for k and 1, and the closure
the spanwise while maintaining the same spanwise coefficients in these equations are different from those of
resolution would not affect the results much (Breuer 2000). the standard k– 1 model. The standard k –1 model is a
As mentioned previously, the velocity field of the high-Re-number model, while the RNG k – 1 model is
circular cylinder wake, especially in the near wake, has more appropriate for flows with a low Re-number.
been investigated numerically by many researchers.
However, most of these numerical simulations were 2.2 LES
focused on isothermal flows. In the current study, the
main objective is to simulate the flow past a slightly heated LES is capable of solving three-dimensional (3D) and
cylinder and investigate both the velocity field and transient turbulent flow problems. The large scales are
temperature field in the intermediate wake. A commercial solved directly, whereas the small scales are modeled
CFD software, FLUENT, is employed for the numerical by a SGS model. Considering the computational cost,
simulation. Meanwhile, the phase-averaging technique is LES is between the DNS and RANS methods. However,
applied to examine the contributions of coherent structures LES is computationally much more expensive than
to global quantities, using a different reference phase. The RANS methods.
numerical results are compared with the experimental data In FLUENT (2005), the finite-volume discretization
performed by Matsumura and Antonia (1993) and Zhou provides the filter operation:
et al. (2002). ð
1
FðxÞ ¼ Fðx0 Þdx0 ; x0 [ v
V
v
2. Mathematical model
where V is the volume of a computational cell. The
Navier– Stokes equations are able to describe turbulent governing equations for LES are derived by filtering the
flow without any turbulence modeling. For direct time-dependent incompressible Navier– Stokes equations,
numerical simulation (DNS), Navier –Stokes equations
are solved directly. However, DNS is quite demanding ›u i › ›p › ›u i ›u j ›tij ›u i
computationally since all the lengthscales and timescales þ ðuj u i Þ ¼ 2 þ n þ 2 ; ¼ 0;
›t ›xj ›xi ›x j ›x j ›x i ›x j ›x i
should be resolved without any modeling. It is estimated
that the number of mesh cells required for incompressible
free turbulence is about Re 9/4. The application of DNS is where tij is the SGS stress tensor defined by
limited to flows at moderate Re due to the prohibitive
tij ¼ ui uj 2 u i u j :
computational cost. The Re-averaged Navier – Stokes
Numerical analysis using RANS and LES 303
A SGS model is required to model the SGS eddy A constant inlet velocity of U1 ¼ 7 m/s is applied in
viscosity. For the Smagorinsky–Lilly model, all simulations. For the RANS case, the inlet turbulence
qffiffiffiffiffiffiffiffiffiffiffiffi intensity is set to 0.4%, the same as the value measured
mt ¼ rL2s 2S ij S ij in the experiment (Zhou et al. 2002). One cycle
sinusoidal lateral velocity is imposed to the flow inlet
to break the symmetry of the calculation domain. For the
Ls ¼ min kd; Cs ðDxDyDzÞ1=3 bigÞ LES case, no perturbation is added since they will be
highly damped since the grid is clustered near the
where Ls is the mixing length for the subgrid scales; d is cylinder surface (Breuer 1998). A symmetry boundary
the distance to the closest wall; k is the von Karman condition is imposed to the lateral upper and lower
constant; and Cs is the Smagorinsky coefficient. It has boundaries, and a pressure outlet boundary condition is
been found that Cs ¼ 0.1 yielded the best results for a set for the flow outlet. In the spanwise direction of the 3D
wide range of flows (FLUENT 2005). Dynamic LES case, periodicity of the flow is assumed. No-slip
Smagorinsky – Lilly model is also available in FLUENT, boundary condition and constant temperature is applied
where the Smagorinsky coefficient Cs is dynamically to the circular cylinder wall.
calculated from the resolved scales (FLUENT 2005).
higher velocity in the center point, whereas Run C2 figures 7 and 8, RANS approach yields much lower time-
obtains lower value in the center point. In the lateral sides, averaged Reynolds stresses.
all simulations yield little higher velocity in comparison to The normalized temperature variance u * is illustrated in
the experimental measurements. Figure 6 illustrates the figure 9. Run B shows two peaks whereas there is only one
time-averaged temperature normalized by the maximum peak in the experimental data. Run D1 obtains the highest
temperature difference between the heated flow and the peak value among all these simulations. The experimental
ambient flow. Fair results are obtained by all simulations data of Matsumura and Antonia (1993) showed the peak
expect for Run B. value of u * at about 0.45, which was higher than that of
The normalized streamwise Reynolds stress at Zhou et al. (2002). If this is considered then the result of
x/D ¼ 10 is shown in figure 7. Again, Run B fails to Run D1 remains reasonable.
reproduce the streamwise Reynolds stress. As regards the The difficulty in predicting the Reynolds stresses in
3D LES, the results calculated with 10 spanwise grid near and intermediate wakes has been reported by many
points (Run C1, C2) are in better agreement with the researchers: Beaudan and Moin (1994), Breuer (2000),
experimental data than the results calculated with 20 Mittal (1996) and Franke and Frank (2002), etc. The
spanwise grid points. deviation between the simulation result and the exper-
Figure 8 shows the cross-stream distribution of lateral imental data increases when the position is moved further
Reynolds stress. Among all these results, the results downstream. The relatively coarse grid downstream and
computed with higher spanwise resolution are in good the diffusive numerical schemes contribute to the under-
agreement with the experimental data. As shown in prediction of the Reynolds stress.
Figure 5. Normalized mean streamwise velocity. Figure 7. Normalized streamwise Reynolds stress.
306 J. Shao and C. Zhang