Computers & Fluids 34 (2005) 1096–1114

www.elsevier.com/locate/compfluid

Direct numerical simulation of flow separation around a

NACA 0012 airfoil
Hua Shan, Li Jiang, Chaoqun Liu *

Department of Mathematics, University of Texas at Arlington, P.O. Box 19408,

411 S. Nedderman, Arlington, TX, 76019-0408, USA

Received 6 February 2003; received in revised form 20 July 2004; accepted 16 September 2004
Available online 8 December 2004

Abstract

Direct numerical simulation (DNS) for the flow separation and transition around a NACA 0012 airfoil
with an attack angle of 4
and Reynolds number of 105 based on free-stream velocity and chord length is
presented. The details of the flow separation, detached shear layer, vortex shedding, breakdown to turbu-
lence, and re-attachment of the boundary layer are captured in the simulation. Though no external distur-
bances are introduced, the self-excited vortex shedding and self-sustained turbulent flow may be related to
the backward effect of the disturbed flow on the separation region. The vortex shedding from the separated
free shear layer is attributed to the Kelvin–Helmholtz instability.
 2004 Elsevier Ltd. All rights reserved.

1. Introduction

Flow transition in separation bubbles is a classic topic and has been studied for many years [3].
A laminar boundary layer over a solid surface will separate as a result of curvature changes or
adverse pressure gradient. The separated shear layer is inviscidly unstable and vortices are formed
due to the Kelvin–Helmholtz mechanism. The detached shear layer may also undergo rapid tran-
sition to turbulence and the resulting turbulent flow may reattach to the wall surface and form an

*
Corresponding author. Tel.: +1 817 272 5151/3261; fax: +1 817 272 5802.
E-mail addresses: hshan@uta.edu (H. Shan), ljiang@uta.edu (L. Jiang), cliu@uta.edu (C. Liu).

0045-7930/$ - see front matter  2004 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compfluid.2004.09.003
 H. Shan et al. / Computers & Fluids 34 (2005) 1096–1114 1097

attached turbulent boundary layer [1]. Obviously, the length of the separation bubble depends on
the location of transition. On the other hand, the size of the separation bubble could directly affect
the flight characteristics of the airfoil and the efficiency of the turbine machine. The flow separa-
tion over a wing in flight results in the loss of lift and the increase of drag as well as the generation
of aerodynamic noise and threatens the stability and efficiency of the aircraft. Understanding the
mechanism of the separation and transition is of great importance in improving the design of air-
craft and turbo-machinery.
The mechanism of the transition in separation bubble could be very different from that in the
attached boundary layer. The latter is usually characterized by the linear and nonlinear amplifi-
cation of small disturbances, such as the cross-flow and the Tollmien–Schlichting (TS) waves. The
growth of the instability can be predicted by linear stability theory (LST; see [6]) for the linear
stage, and parabolized stability equations (PSE; see [2,7]) method for both linear and nonlinear
stage. The LST is mainly a local analysis with the assumption of parallel base flow. The PSE
incorporates inhomogeneous initial and boundary conditions but still assumes a steady base flow.
Neither LST nor PSE may work for the cases involving flow separations. The alternative ap-
proaches include direct numerical simulation (DNS) and large eddy simulation (LES), which have
become more feasible with the development of computer resources and efficient numerical meth-
ods. DNS and LES are considered as ideal tools to study the flow transition in separation bubbles.
To date, most DNS studies are limited to flow separation over flat surfaces. The transition in a
‘‘short’’ laminar separation bubble was simulated in a DNS conducted by Alam and Sandham [1].
In their simulations, the adverse pressure gradient was reproduced by a method of suction
through the upper boundary, and the three-dimensional disturbances were imposed on the lower
boundary at a location upstream of the separation point. Their DNS results showed that the tran-
sition region was characterized by the staggered large-scale K-vortices, similar to those found in a
typical transition scenario for a boundary layer flow. In the DNS study of Spalart and Strelets
[15], the laminar boundary on a flat surface was made to separate by way of aspiration through
an opposite boundary. Their results showed the wavering of a separated shear layer followed by a
sudden transition to turbulence without vortex pairing. They suggested that transition occurs as a
result of the amplification of disturbance due to the Kelvin–Helmholtz mechanism and the expo-
sure to pressure signals from the downstream flow. The LES of the separation and transition of a
boundary layer over a blunt leading edge was carried out by Yang and Voke [17]. Their results
indicated that the free shear layer formed in the separation bubble is inviscidly unstable via the
Kelvin–Helmholtz mechanism and the initial two-dimensional instability waves grow downstream
linearly, with slow development of three-dimensional motions via a secondary instability mecha-
nism. Then the distorted spanwise vortices roll up accompanied by significant three-dimensional
motions as a result of nonlinear interactions. Finally, the breakdown to turbulence occurs around
the mean re-attachment point, and the flow develops into turbulent boundary layer after the re-
attachment. Although the effect of the change of curvature near the leading edge that evokes
airfoils is considered in their LES, other important issues, e.g. the influence of the wake, are still
absent. As a matter of fact, the DNS or LES results of flow separation and transition around an
airfoil with nonzero angle of attack is rarely found in the literatures.
In the present work, an attempt is made to conduct a DNS study with high-accuracy and high-
resolution numerical methods to investigate the details of the flow separation and transition
over a NACA 0012 airfoil with an attack angle of 4
and a Reynolds number of 105 based on
1098 H. Shan et al. / Computers & Fluids 34 (2005) 1096–1114

the free-stream velocity and the airfoil chord length. The simulation includes a two-dimensional
case and a three-dimensional case. The two-dimensional case is designed to study the flow sepa-
ration and unsteady flow structures within the separated shear layer, while the three-dimensional
case is used to simulate both flow separation and transition to turbulence. The organization of the
paper is as follows. In Section 2, the governing equations and numerical methods will be intro-
duced briefly. In Section 3, a description of the flow configuration and the computational set
up will be given. In Section 4, the computational results will be presented and analyzed. Conclud-
ing remarks can be found in Section 5.

2. Governing equations and numerical methods

2.1. Governing equations

The three-dimensional compressible Navier–Stokes equations in generalized curvilinear coordi-

nates (n, g, f) are written in conservative form as
1 oQ oðE
Ev Þ oðF
F v Þ oðF
F v Þ
þ þ þ ¼ 0: ð1Þ
J ot on og of
The vector of conserved quantities Q, inviscid flux vector (E, F, G), and viscous flux vector
(Ev, Fv, Gv) are defined by
0 1 0 1 0 1 0 1
q qU qV qW
B qu C B qUu þ pnx C B qVu þ pg C B qWu þ pfx C
B C 1B C 1B
x C
1B C
B C B C B C B C
Q ¼ B qv C; E ¼ B qUv þ pny C; F ¼ B qVv þ pgy C; G ¼ B qWv þ pfy C;
B C JB C JB C JB C
@ qw A @ qUw þ pnz A @ qVw þ pgz A @ qWw þ pfz A
Et U ðEt þ pÞ V ðEt þ pÞ W ðEt þ pÞ
0 1 0 1 0 1
0 0 0
B sxx nx þ syx ny þ szx nz C B sxx gx þ syx gy þ szx gz C B sxx fx þ syx fy þ szx fz C
1B
B
C
C 1 B
B
C
C 1 B
B
C
C
Ev ¼ B sxy nx þ syy ny þ szy nz C; F v ¼ B sxy gx þ syy gy þ szy gz C; Gv ¼ B sxy fx þ syy fy þ szy fz C;
JB C JB C JB C
@ sxz nx þ syz ny þ szz nz A @ sxz gx þ syz gy þ szz gz A @ sxz fx þ syz fy þ szz fz A
Qx nx þ Qy ny þ Qz nz Qx gx þ Qy gy þ Qz gz Qx fx þ Qy fy þ Qz fz

where J ¼ oðn;g;fÞ
oðx;y;zÞ
is the Jacobian of the coordinate transformation between the curvilinear (n, g, f)
and Cartesian (x, y, z) frames, and nx, ny, nz, gx, gy, gz, fx, fy, fz are coordinate transformation met-
rics. q is the density. The three components of velocity are denoted by u, v, and w. The total energy
is Et and is given by
p 1
Et ¼ þ qðu2 þ v2 þ w2 Þ:
c
1 2
The contravariant velocity components U, V, W are defined as
U  unx þ vny þ wnz ; V  ugx þ vgy þ wgz ; W  ufx þ vfy þ wfz :
 H. Shan et al. / Computers & Fluids 34 (2005) 1096–1114 1099

The terms Qx, Qy, Qz in the energy equation are defined as

Qx ¼
qx þ usxx þ vsxy þ wsxz ;
Qy ¼
qy þ usxy þ vsyy þ wsyz ;
Qz ¼
qz þ usxz þ vsyz þ wszz :
The components of the viscous stress tensor and heat flux are denoted by sxx, syy, szz, sxy, sxz, syz,
and qx, qy, qz, respectively. The components of the viscous stress tensor and the heat flux in their
nondimensional form are as follows:

 
l oui ouj 2 ouk
sij ¼ þ
dij ;
Re oxj oxi 3 oxk
l oT
qi ¼
2
:
ox
ðc
1ÞM Re Pr i
The reference values for length, density, velocities, temperature, pressure and time used to nondi-
mensionalize the flow quantities are Lr, qr, Ur, Tr, qr U 2r , and Lr/Ur, respectively, where the sub-
script ‘‘r’’ denotes the reference state. The dimensionless parameters, including the Mach
number M, the Reynolds number Re, the Prandtl number Pr, and the ratio of specific heats c,
are defined as follows:
Ur qr U r Lr C p lr Cp
M ¼ pffiffiffiffiffiffiffiffiffiffi ; Re ¼ ; Pr ¼ ; c¼ ;
cRg T lr kr Cv

where Rg is the ideal gas constant, Cp and Cv are specific heats at constant pressure and constant
volume, respectively. Throughout this work, Pr = 0.7 and c = 1.4. Viscosity is determined accord-
ing to SutherlandÕs law in dimensionless form,
T 3=2 ð1 þ SÞ 110:3 K
l¼ and S ¼ :
T þS T1
The system of governing equations is completed by the equation of state,

cM 2 p ¼ qT :

2.2. Numerical methods

In Eq. (1), a second-order Euler Backward scheme is used for time derivatives, and the equation
is written in the fully implicit form:
3Qnþ1
4Qn þ Qn
1 oðEnþ1
Enþ1
v Þ oðF nþ1
F nþ1
v Þ oðGnþ1
Gnþ1
v Þ
þ þ þ ¼ 0: ð2Þ
2J Dt on og of
The term Qn+1 is estimated iteratively as
Qnþ1 ¼ Qp þ dQp ;
1100 H. Shan et al. / Computers & Fluids 34 (2005) 1096–1114

where
dQp ¼ Qpþ1
Qp :
At step p = 0, Qp = Qn; as dQp is driven to zero, Qp approaches Qn+1. The flux vectors are linear-
ized by
Enþ1  Ep þ Ap dQp ;
F nþ1  F p þ Bp dQp ;
Gnþ1  Gp þ C p dQp :
Then, Eq. (2) can be written as
 
3
I þ DtJðDn A þ Dg B þ Df CÞ dQp ¼ R; ð3Þ
2
where R is the residual,


3 p 1 n
1
R ¼
Q
2Q þ Q n

DtJ ½Dn ðE
Ev Þ þ Dg ðF
F v Þ þ Df ðG
Gv Þp ; ð4Þ
2 2
and Dn, Dg, Df represent partial differential operators along the curvilinear coordinate direction
indicated by the subscript, and A, B, C are the Jacobian matrices of the flux vectors:
oE oF oG
A¼ ; B¼ ; G¼ :
oQ oQ oQ
The right-hand side of Eq. (3) is discretized using a sixth-order compact scheme [9] for spatial
derivatives, and the left-hand side of the equation is discretized following the LU-SGS method
[18]. In this method, the Jacobian matrices of flux vectors are split as
A ¼ Aþ þ A
; B ¼ Bþ þ B
; C ¼ Cþ þ C
;
where
A ¼ 12½A  rA I; B ¼ 12½B  rB I; C  ¼ 12½C  rC I;
and
rA ¼ j maxðjkA jÞ þ ~m; rB ¼ j maxðjkB jÞ þ ~m; rC ¼ j maxðjkC jÞ þ ~m;
with kA, kB, and kC as eigenvalues of A, B, and C, respectively, j is a constant greater than 1. The
effect of viscous terms are taken into account by ~m which is expressed by
 
l 4 l
~m ¼ max ; :
ðc
1ÞM 2 Re Pr 3 Re
The first-order upwind finite difference scheme is used for the split flux terms on the left-hand side
of Eq. (3). This does not affect the accuracy of the scheme when solutions are converged. The finite
difference representation of Eq. (3) can be written as
 
3
I þ DtJðrA þ rB þ rC ÞI dQpi;j;k þ DtJ ½A
dQpiþ1;j;k
Aþ dQpi
1;j;k
2
þ B
dQpi;jþ1;k
Bþ dQpi;j
1;k þ C
dQpi;j;kþ1
C þ dQpi;j;k
1  ¼ Rpi;j;k : ð5Þ
 H. Shan et al. / Computers & Fluids 34 (2005) 1096–1114 1101

In the LU-SGS method, Eq. (5) is usually solved with three steps. First, initialize dQ0 by
 
1
0 3
dQi;j;k ¼ I þ DtJ ðrA þ rB þ rC ÞI Rni;j;k :
2
In the second step, the following relation is used.
 
1
3
dQi;j;k ¼ dQi;j;k þ I þ DtJ ðrA þ rB þ rC ÞI DtJ ½Aþ dQi
1;j;k þ Bþ dQi;j
1;k þ C þ dQi;j;k
1 :
 0
2
In the last step, dQp is obtained by
 
1
3
dQi;j;k ¼ dQi;j;k
I þ DtJ ðrA þ rB þ rC ÞI DtJ ½A
dQpiþ1;j;k þ B
dQpi;jþ1;k þ C
dQpi;j;kþ1 :
p 
2
The spatial derivatives appearing in the residual term as shown in Eq. (4) are computed using a
sixth-order centered compact difference scheme in the interior of the domain. The compact scheme
is reduced to fourth- and third-order near the boundary. To suppress the numerical oscillation
associated with the high-order compact scheme, an eighth-order compact filter is applied at reg-
ular intervals in the simulation. The coordinate transformation metrics are computed using the
compact scheme by an approach that satisfies the geometric conservation law numerically [8].
The numerical verification of the computer code can be found in the previous work of Jiang
et al. [10]. Parallel computation based on the Message Passing Interface (MPI) has been utilized
to improve the performance of the code. The parallel computation is combined with the domain
decomposition method. The computational domain is divided into n equal-sized subdomains
along the n direction. Readers may refer to early work of Shan et al. [13,14] for more details of
the parallel algorithm.

3. Problem definitions and boundary conditions

Numerical simulations are performed on a NACA 0012 airfoil at an angle of attack (AOA) of
4
. The free-stream velocity U1, the free-stream pressure p1, the free-stream temperature T1, and
the chord length of the airfoil C, are selected as the reference parameters for nondimensionaliza-
tion. The computational domain viewed from the spanwise direction is shown in Fig. 1. The up-
stream boundary is three chord lengths away from the leading edge of the airfoil. The upper and
lower boundaries are about four chord lengths from the airfoil surface. The outflow boundary is
located at two chord lengths downstream of the trailing edge. The airfoil is regarded as infinitely
long in the spanwise direction. In our simulations, spanwise length is set to 0.1C and a periodic
boundary condition is imposed at the spanwise boundaries.
The flow conditions and grid parameters are summarized in Table 1. The Reynolds number
based on the free-stream velocity and the chord length is 105. The free-stream Mach number is
0.2. The numbers of grid points in the n, g, and f directions are 1200, 32, and 180, respectively.
The grid distributions in the (x, z) plane and on the airfoil surface are shown in Fig. 2. The grid
sizes in wall units shown in Table 1 are estimated based on a turbulent boundary layer flow with
the same Reynolds number. The C-type grid is produced in the (x, z) plane using the elliptic grid
1102 H. Shan et al. / Computers & Fluids 34 (2005) 1096–1114

Fig. 1. Computational domain.

Table 1
Computational parameters
Re = U1C/t1 M Nn · Ng · Nf Dx+ Dy+ Dz+
5
10 0.2 1200 · 32 · 180 <13 <15 <1

Fig. 2. Grid distribution (one out of every three grid point is shown).

generation method [16] and the grids are uniformly distributed in the spanwise direction and clus-
tered near the airfoil surface in the wall-normal direction. Parallel computation based on the do-
main decomposition is implemented on a SGI Origin 2000 platform with the Message Passing
Interface (MPI) libraries. The computational domain is divided evenly into 24 subdomains along
the n direction and 24 processors are used in the parallel computation.
 H. Shan et al. / Computers & Fluids 34 (2005) 1096–1114 1103

The nonreflecting boundary conditions proposed by Poinsot and Lele [12] and used in Jiang
et al. [11] are imposed at the upstream, far field, and outflow boundaries. Due to the subsonic nat-
ure of the flow, the uniform free-stream velocities and temperature are prescribed at upstream
boundary, while density is determined by the nonreflecting boundary conditions. At the outflow
boundary, the perfectly nonreflecting boundary condition [12] is used. The no-slip and the adia-
batic boundary conditions are used on the surface of the airfoil.

4. Results and analysis

In this study, numerical simulations of a two-dimensional flow and a three-dimensional flow

around the NACA 0012 airfoil at 4
angle of attack have been carried out. The time step size
is set to 8.35 · 10
5C/U1. For the two-dimensional simulation, the uniform flow is used as the
initial condition. For the three-dimensional simulation, the two-dimensional solutions are used
as the initial field. The following subsections will present the results from both two-dimensional
and three-dimensional simulations.

4.1. Two-dimensional instability waves and vortex shedding

Although the transition and turbulent flow around a two-dimensional airfoil is inherently three-
dimensional, the two-dimensional simulation still helps us gain a better understanding of the
development of the flow separation, instability, and vortex shedding. The two-dimensional simu-
lation starts initially from a uniform flow field, which is not the solution of the governing equa-
tions and may bring in some disturbances, namely the initial perturbation resulting from the
residual of the numerical solution. If such a perturbation does not dissipate quickly in the simu-
lation, it may trigger the most unstable instability waves in the separation bubble and in the near
wake. This explains the unsteadiness in the results of the two-dimensional simulation even though
all the specified boundary conditions are steady and no external disturbances are enforced.
The contours of the instantaneous spanwise vorticity from the two-dimensional simulation are
shown in Fig. 3, where the presence of the flow separation and vortex shedding is clearly visible on
the upper surface of the airfoil. The separation that starts near the leading edge of the airfoil
forms a separated shear layer, which becomes unstable near the mid-chord and leads to the
quasi-periodic shedding of large-scale vortical structures. It is believed that the vortices are gen-
erated through streamwise growth of the disturbance in the separated shear layer [3]. Then, the
two-dimensional vortices are carried downstream by the mean flow along the airfoil surface,

Fig. 3. Contours of instantaneous spanwise vorticity from the two-dimensional simulation.

1104 H. Shan et al. / Computers & Fluids 34 (2005) 1096–1114

where the vortex pairing can also be observed from Fig. 3. There is no evidence of small-scale vor-
tical structures or flow transition in the two-dimensional simulation.
The separation zone can also be seen clearly from the time-averaged velocity vectors shown in
Fig. 4, where a very strong reversed flow region can be found. The separation of the mean flow
starts near x/C = 0.2. The re-attachment of the mean flow occurs near x/C = 0.7.
Since no external disturbance is imposed in the two-dimensional simulation, except for the ini-
tial residual disturbance, a question is raised as to the origin of the instability of the separated
shear layer and vortex shedding. Fig. 5 shows the time histories of the pressure fluctuation at

Fig. 4. Time-averaged velocity vectors from the two-dimensional simulation (one out of every five streamwise locations
is shown).

Fig. 5. Time history of pressure fluctuation at different locations.

 H. Shan et al. / Computers & Fluids 34 (2005) 1096–1114 1105

Fig. 6. Time history of streamwise velocity at different locations.

various streamwise locations. The time histories of streamwise velocity at the same locations are
shown in Fig. 6. After the flow is established (t > 10.0C/U1), the amplitude of oscillations stays at
a certain level without significant change in time, indicating a saturation of disturbance in the
flow. However, the process of the oscillation development before saturation is worth more de-
tailed studies to find out the origin of these disturbances in the computation. A closer look at
Fig. 5 shows that the first appearance of pressure oscillation occurs at t  5.4C/U1 in a location
near the wake at x/C = 1.048 corresponding to Fig. 5(f). At an upstream station of x/C = 0.874
shown in Fig. 5(e), the disturbance starts to grow after t  6.0C/U1. Fig. 5(d) shows a further
upstream station at x = 0.645C and the visible pressure disturbance emerges at t  7.2C/U1.
Therefore, the instability first appears in the near wake just behind the trailing edge of the airfoil.
This is also confirmed by Fig. 6, where the streamwise velocity fluctuation in the near wake at
x/C = 1.048 reaches its saturated amplitude at t  7.2C/U1, as shown in Fig. 6(f), meanwhile
the small oscillation just starts to emerge at x/C = 0.645 in Fig. 6(d). Assume that a backward
effect [3] exists, the disturbance at the leading edge of the airfoil is influenced by the perturbation
amplified earlier on in the near wake. It is also interesting to note that the first sign of the pressure
oscillation at the leading edge appears at t  5.6C/U1, compared to t  5.4C/U1 in the near
wake, the time lag indicates that the amplified perturbation in the near wake may propagate
against the flow direction in the form of acoustic waves. Therefore, the early stage of the devel-
opment of instability can possibly be interpreted as follows: The flow in the near wake contains
strong shear layers as a result of the two flows with opposite vorticity sweeping off the upper and
lower surface of the airfoil. Since the shear layer in the near wake is very receptive to background
fluctuations, probably more unstable than the free shear layer in the separation zone over the
1106 H. Shan et al. / Computers & Fluids 34 (2005) 1096–1114

Fig. 7. Mean velocity profiles at different locations.

upper surface of the airfoil, the instability first emerges in the near wake as the shear layer re-
sponds to the aforementioned residual disturbance. Then the disturbances may propagate up-
stream in the form of acoustic waves and introduce disturbance to the separated free shear
layer near the leading edge. The streamwise growth of the separated layer disturbance results
in the oscillation of the free shear layer and the shedding of vortices.
Fig. 7 shows the mean profiles of the streamwise velocity as a function of the dimensionless
wall-normal coordinate at different streamwise locations. The mean flow separation is visible
on the velocity profiles from x/C = 0.2 through 0.7. The mean velocity profiles in the separation
zone where strong reversed flow occurs have inflection points, which also appear on instantaneous
streamwise velocity profiles. Therefore, the separated flow is subjected to an inviscid instability
with an amplification rate larger than its viscous counterpart [17]. It has been demonstrated by
some authors from their DNS results that the separated shear layer has an absolute instability
when the maximum reversed flow reaches 15–20% of the external free-stream velocity [17]. The
results from our two-dimensional simulation show that the maximum value of the mean reversed
velocity reaches only 8% of the free-stream velocity. However, in the two-dimensional simulation,
the disturbances are amplified in the separated shear layer and result in vortex shedding. Many
authors have related this instability to the Kelvin–Helmholtz mechanism [5,15,17], which arises
from the crinkling of the interface by the shear.

4.2. Three-dimensional mean flow

As expected, compared to the two-dimensional simulation, the three-dimensional simulation re-

sults are more complicated. The mean flow characteristics are presented first. Fig. 8 shows the
maximum mean reversed flow distribution along the upper surface of the airfoil. The mean sep-
aration zone lies between x/C = 0.19 and 0.68, corresponding to the mean separation point, and
the mean re-attachment point, respectively. Due to the large-scale oscillations in the three-dimen-
sional flow, the shape and the length of the separation bubble are functions of time. The time-
averaged length of the bubble is about lb/C = 0.49, as estimated from Fig. 8. The reversed flow
is about 8% of the free-stream velocity at the mid-chord position of x/C = 0.5, followed by a sharp
 H. Shan et al. / Computers & Fluids 34 (2005) 1096–1114 1107

Fig. 8. Mean reversed flow distribution on the suction side, U rev ¼ minð

uÞ.

Fig. 9. Mean velocity vector (one out of every five streamwise location is shown near the surface).

increase in its amplitude and reaches a peak of about 26% at x/C = 0.6. The separation zone can
also be identified clearly from the velocity vectors of the mean flow shown in Fig. 9.
Fig. 10 shows the mean pressure coefficient of the airfoil obtained from the three-dimensional
simulation. Near the leading edge, the strong adverse pressure gradient on the upper surface
causes the separation of the boundary layer. The flattened region that follows the adverse pressure
gradient is corresponding to the separation. On the upper surface near x/C = 0.6, the adverse pres-
sure gradient forms again and causes the rapid increase of reversed flow corresponding to the peak
shown in Fig. 8.
Fig. 11 shows the skin friction coefficient obtained from the mean flow on the upper surface of
the airfoil. The simulation result showing a negative surge of Cf agrees qualitatively with the DNS
result of Spalart and Strelets [15]. In the majority part of the mean separation zone between x/
C = 0.2 and 0.7, the skin friction coefficient is negative, indicating the near wall reversed flow
in the separation bubble. The small positive value of Cf in the middle of the separation zone cor-
responds to a thin layer of the mean flow in the streamwise direction underneath the reversed flow,
1108 H. Shan et al. / Computers & Fluids 34 (2005) 1096–1114

Fig. 10. Mean pressure coefficient.

Fig. 11. Mean skin friction coefficient.

as shown in Fig. 12. The sudden decrease of Cf before reaching its negative peak at x/C = 0.63
indicates the transition. The abrupt recovery of Cf from its negative peak to a positive value near
x/C = 0.68 corresponds to the re-attachment of the separated flow.
The profiles of the mean streamwise velocity and its r.m.s fluctuation as a function of the
dimensionless wall-normal coordinate at various streamwise locations are shown in Fig. 13(a)
and (b). In Fig. 13(a), the mean velocity profiles show the strong reversed flow in the separation
zone, where inflection points can be clearly seen on the profiles. After re-attachment, the mean
velocity profiles, though not fully re-established, resemble the velocity profile of the turbulent
flow. It has been shown in both numerical simulation [1] and experiment [4] that the turbulent
log law may re-establish after several separation bubble lengths in the boundary layer behind
 H. Shan et al. / Computers & Fluids 34 (2005) 1096–1114 1109

Fig. 12. An enlarged view of the thin layer of flow in the streamwise direction underneath the reversed mean flow.

Fig. 13. The profiles of mean streamwise velocity and its r.m.s. fluctuation at several streamwise locations.

the re-attachment from a separation. In our case, the separation zone occupies half of the chord
over the airfoil upper surface. Therefore, before the log law is re-established in the reattached
boundary layer, the flow already sweeps off the airfoil into the wake.
In Fig. 13(b), as we can see from the profiles in the upstream part of the separation zone, e.g., at
x/C = 0.2, 0.3, 0.4, and 0.5, the location of the peak fluctuation along the wall-normal direction
almost coincides with the inflection point on the mean velocity profile, indicating the dominance
of the inviscid instability in the separated shear layer. The fluctuation near the wall overtakes the
inviscid instability and becomes dominant at x/C = 0.6 showing strong turbulent activities in the
near wall region.
1110 H. Shan et al. / Computers & Fluids 34 (2005) 1096–1114

4.3. Three-dimensional instantaneous flow

Fig. 14 shows the contours of instantaneous spanwise vorticity at the mid-span location of the
three-dimensional computational domain at selected instants during the simulation. The evolution
of vortex shedding can be clearly seen from these snapshots. Since the three-dimensional simula-
tion starts from the results of the two-dimensional simulation, the first three frames (a), (b), and
(c) in this figure resemble the two-dimensional results in terms of the large-scale vortices shedding
from the separated shear layer and rolling downstream, except some small-scale structures, which
are absent from the two-dimensional results, appear upstream of the trailing edge. After the three-
dimensional flow is established in the simulation, as we can see from frames (e), (f), (g), and (h) in
Fig. 14, the large-scale vortices have almost disappeared. This is because the two-dimensional
structure cannot maintain its shape due to the development of three-dimensional motion during
the transition to turbulence, as it will be shown later on. Fig. 14 also shows that the size of the
separation zone changes with time in the three-dimensional simulation.
Fig. 15 shows the three-dimensional isosurfaces of the spanwise component of the instanta-
neous vorticity over the upper surface of the airfoil. The separation of the free shear layer from
the airfoil surface, the vortex shedding from the separated shear layer, and the breakdown to tur-
bulence can be clearly recognized in this figure. Near the leading edge of the airfoil, a two-dimen-
sional shear layer is seen detached from the surface. Although there exists streamwise growth of

Fig. 14. Instantaneous spanwise vorticity at various times.

 H. Shan et al. / Computers & Fluids 34 (2005) 1096–1114 1111

Fig. 15. Three-dimensional isosurface of instantaneous spanwise vorticity (dark color: xy = 50; light color: xy =
50).

the separated-layer disturbance, the free shear layer seems to be two-dimensional and the flow is
laminar before x/C = 0.5, where some slight distortions become visible on the isosurface and the
three-dimensional fluctuations start to grow. The three-dimensional disturbance grows so rapidly
that the two-dimensional vortical structure attached to the separated shear layer breaks into
small-scale three-dimensional structures in the immediate downstream region. Fig. 15 also shows
that the flow becomes turbulent around the re-attachment point, indicating the correlation be-
tween transition and re-attachment.
Fig. 16 shows the streamwise distribution of the extreme values in the three components of
instantaneous vorticity, where xx, xy, and xz denote the vorticity components in the streamwise,
spanwise, and wall-normal directions, respectively. The solid lines in Fig. 16 represent the maxi-
mum values, while the dashed lines the minimum values. As it can be seen from Fig. 16(a) and (c),
xx and xz remain almost zero from the leading edge to x/C = 0.4, indicating that the flow is two-
dimensional in this region. These results confirm our earlier observations in Fig. 15. In Fig. 16(b),
the nonzero value of xy before x/C = 0.4 corresponds to the separated shear layer. The growth of
three-dimensional disturbance starts near x/C = 0.55 in Fig. 16, corresponding to the small distor-
tion that appears on the isosurface of spanwise vorticity in Fig. 15, followed by a significant
growth near x/C = 0.6 where the large-scale vortical structures break into smaller structures, as
shown in Fig. 15.
The three-dimensional simulation starts from the results of the two-dimensional simulation and
there is no three-dimensional disturbance applied to the computation in the form of external
forces, or initial and boundary conditions. Nevertheless, the three-dimensional distortion of the
separated shear layer, the generation of small-scale structures, and the transition to turbulence,
as observed in the results of the simulation, are all associated with a three-dimensional distur-
bance. Again, the question that needs to be answered is what is the original source of the
three-dimensional disturbance? The answer may be found in Fig. 14. As we can see in Fig. 14
(b), the first sign of small-scale structures, which are three-dimensional and similar to those shown
by Fig. 15, appear in the region near the trailing edge of the airfoil, where a strong interaction
occurs between the large-scale vortex and the wake. The three-dimensional perturbation, from
which the small-scale structures develop in the near wake region, could be the numerical residual
fluctuation, which is also three-dimension in this case. At a later time, the small-scale structures
can be seen at other upstream locations, as shown in Fig. 14(e)–(g). Without externally enforced
1112 H. Shan et al. / Computers & Fluids 34 (2005) 1096–1114

Fig. 16. Streamwise distribution of the extreme values of the instantaneous vorticity.

perturbation, the three-dimensional instability and transition to turbulence seem to be self-excited

and self-sustained. This might be related to the backward effect [3] as the separation bubble re-
sponds to the high-amplitude three-dimensional perturbations of the downstream reattaching
flow. The upstream propagation of the disturbance in the form of pressure signal is also reported
by Spalart and Strelets [15] in their DNS results.

5. Conclusions

Two-dimensional and three-dimensional direct numerical simulation (DNS) has been carried
out to study flow separation around a NACA 0012 airfoil with an angle of attack of 4
and a Rey-
nolds number of 105. The full compressible Navier–Stokes equations in the generalized curvilinear
coordinate system are solved by a LU-SGS implicit scheme together with the high-order compact
central difference scheme and nonreflecting boundary conditions.
In the two-dimensional simulation starting initially from the uniform flow field, the first sign of
the instability appears in the near wake region as the free shear layer responds to the background
perturbation. The disturbances in the near wake may propagate upstream in the form of acoustic
waves and introduce a disturbance to the separated shear layer over the upper surface of the air-
foil. Since the free shear layer is inviscidly unstable via the Kelvin–Helmholtz mechanism, the
 H. Shan et al. / Computers & Fluids 34 (2005) 1096–1114 1113

streamwise growth of the disturbance results in oscillations of the separated layer and eventually
leads to vortex shedding.
The initial flow of the three-dimensional simulation is obtained from the two-dimensional sim-
ulation results. No three-dimensional disturbance is enforced in the simulation. The origin of
three-dimensional instability that emerges in the simulation comes from the near wake region,
where intensive interactions between the large-scale vortical structure and the wake occur. The
three-dimensional instability seems to be self-sustained and leads to transition to turbulence.
The three-dimensional simulation results, showing the correlation between the re-attachment
and the transition, may shed light on the separation control for an airfoil. For example, the dis-
turbance introduced by unsteady blowing will excite the inherent local instability wave and lead to
early transition to turbulence, which will reduce the size of separation zone by an early re-
attachment.

Acknowledgments

This work was sponsored by the Air Force Office of Scientific Research (AFOSR) and
monitored by Dr. Leonidas Sakell and then Dr. Tomas Beutner under the grant number
F49620-01-1-0028. The views and conclusions contained here should not be interpreted as neces-
sarily representing the official policies or endorsements of the AFOSR or the US Government.
The authors would also like to thank the High Performance Computing Center of the US
Department of Defense for providing valuable computer hours.

References

[1] Alam M, Sandham ND. Direct numerical simulation of ÔshortÕ laminar separation bubbles with turbulent
reattachment. J Fluid Mech 2000;410:1–28.
[2] Bertolotti FP, Herbert T, Spallart PR. Linear and nonlinear stability of the Blasius boundary layer. J Fluid Mech
1992;242:441–74.
[3] Boiko AV, Grek GR, Dovgal AV, Kozlov VV. The origin of turbulence in near-wall flows. Springer; 2002.
[4] Castro I, Epik E. Boundary development after a separated region. J Fluid Mech 1999;374:91–116.
[5] Craik ADD. Wave interaction and fluid flow. Cambridge University Press; 1982.
[6] Drazin PG, Reid WH. Hydrodynamic stability. Cambridge University Press; 1981.
[7] Herbert T. Parabolized stability equations. Annu Rev Fluid Mech 1997;29:245–83.
[8] Gaitonde DV, Visbal MR. Further development of a Navier–Stokes solution procedure based on higher-order
formulas. AIAA Paper 99-0557, 1999.
[9] Lele SK. Compact finite difference schemes with spectral-like resolution. J Comput Phys 1992;103:16–42.
[10] Jiang L, Shan H, Liu C. Direct numerical simulation of boundary-layer receptivity for subsonic flow around airfoil.
In: Recent Advances in DNS and LES, Proceedings of the Second AFOSR (Air Force Office of Scientific Research)
International Conference, Rutgers, New Jersey, June 7–9, 1999.
[11] Jiang L, Shan H, Liu C. Nonreflecting boundary conditions for DNS in curvilinear coordinates. In: Recent
Advances in DNS and LES, Proceedings of the Second AFOSR (Air Force Office of Scientific Research)
International Conference, Rutgers, New Jersey, June 7–9, 1999.
[12] Poinsot TJ, Lele SK. Boundary conditions for direct simulations of compressible viscous flows. J Comput Phys
1992;101:104–29.
[13] Shan H, Jiang L, Liu C. Numerical simulation of complex flow around a 85
delta wing. DNS/LES Progress and
Challenges, in: Proceedings of the Third AFOSR (Air Force Office of Scientific Office) International Conference on
DNS/LES, Arlington, Texas, August 5–9, 2001.
1114 H. Shan et al. / Computers & Fluids 34 (2005) 1096–1114

[14] Shan H, Jiang L, Liu C. Direct numerical simulation of three-dimensional flow around a delta wing. AIAA Paper
2000-0402, 2000.
[15] Spalart PR, Strelets MK. Mechanisms of transition and heat transfer in a separation bubble. J Fluid Mech
2000;403:329–49.
[16] Spekreuse SP. Elliptic grid generation based on Laplace equation and algebraic transformations. J Comput Phys
1995;118:38–61.
[17] Yang ZY, Voke PR. Large-eddy simulation of boundary-layer separation and transition at a change of surface
curvature. J Fluid Mech 2001;439:305–33.
[18] Yoon S, Kwak D. Implicit Navier–Stokes solver for three-dimensional compressible flows. AIAA J 1992;30:
2653–9.