Computers & Fluids 35 (2006) 693702 www.elsevier.

com/locate/compuid

An immersed-boundary method for compressible viscous ows

P. De Palma, M.D. de Tullio, G. Pascazio, M. Napolitano
Available online 20 February 2006

DIMeG, Sez. Macchine ed Energetica, and CEMeC, Politecnico di Bari, Via Re David 200, 70125 Bari, Italy

Abstract This paper combines a state-of-the-art method for solving the preconditioned compressible NavierStokes equations accurately and eciently for a wide range of the Mach number with an immersed-boundary approach which allows one to use Cartesian grids for arbitrarily complex geometries. The method is validated versus well documented test problems for a wide range of the Reynolds and Mach numbers. The numerical results demonstrate the eciency and versatility of the proposed approach as well as its accuracy, from incompressible to supersonic ow conditions, for moderate values of the Reynolds number. Further improvements, obtained via local grid renement or non-linear wall functions, can render the proposed approach a formidable tool for studying complex three-dimensional ows of industrial interest. 2006 Elsevier Ltd. All rights reserved.

1. Introduction Many uid dynamic problems of engineering interest exhibit ow regions with very dierent Mach numbers, which render their numerical simulation very dicult. In fact, due to the dierent nature of the physical phenomena associated with ows at dierent Mach numbers, usually a single numerical method performs accurately and eciently within a limited range of the Mach number. Moreover, the presence of complex and/or moving boundaries usually requires time consuming body-tted grid generations. The aim of the present work is to remedy both of the aforementioned diculties, by combining a stateof-the-art method for solving the preconditioned compressible NavierStokes equations accurately and eciently for a wide range of the Mach number with an immersed-boundary (IB) approach which allows to use Cartesian grids for arbitrarily complex geometries. Concerning the preconditioning of the governing equations, the residual of the compressible NavierStokes or Reynolds-averaged NavierStokes (RANS) equations is
Corresponding author. Tel.: +390805963464; fax: +390805963411. E-mail addresses: depalma@poliba.it (P. De Palma), detullio@ imedado.poliba.it (M.D. de Tullio), pascazio@poliba.it (G. Pascazio), napolita@poliba.it (M. Napolitano). 0045-7930/$ - see front matter 2006 Elsevier Ltd. All rights reserved. doi:10.1016/j.compuid.2006.01.004
*

premultiplied by a suitable matrix which uniforms the wave propagation speeds, thus greatly enhancing the accuracy and eciency of the compressible ow solver when applied to low-Mach-number ows. Such a preconditioning technique, originally designed for steady ows [1,2], has been extended to the unsteady ones using a dualtime-stepping (DTS) technique with a three-level backward discretization of the time derivative, in conjunction with a third-order-accurate nite volume method based on ux-vector splitting for the convective terms [3,4]. The IB technique, originally designed for incompressible ows [5,6], allows the body surface to cut the computational cells, so that a simple Cartesian grid can be employed, independently of the complexity of the considered geometry. In this work, the IB technique has been extended to the preconditioned compressible Navier Stokes and RANS equations to provide an accurate, ecient and versatile tool for studying complex threedimensional ows of industrial interest. Here the proposed method is validated versus two-dimensional ows for a very wide range of the Reynolds and Mach numbers. In the following sections, a brief review of the governing equations and their solution technique is given at rst, then, the IB technique is described and, nally, results are provided and compared versus numerical and experimental ones available in the literature.

694

P. De Palma et al. / Computers & Fluids 35 (2006) 693702

2. Governing equations and numerical method The Reynolds-averaged NavierStokes (RANS) equations, written in terms of Favre mass-averaged quantities and using the standard kx turbulence model, can be written as follows:
oq o quj 0; ot oxj oqui o op o^ji s ; quj ui ot oxj oxi oxj   oqU o o ok ui^ij l r lt qj ; s quj H ot oxj oxj oxj   oqk o oui o ok quj k sij ; b qxk l r lt ot oxj oxj oxj oxj   oqx o cx oui o ox 2 . bqx l rlt quj x sij ot oxj k oxj oxj oxj 1 2 3 4 5

where Q is the conservative variable vector, E, F, and Ev, Fv indicate the inviscid and viscous uxes, respectively, D is the vector of the source terms for the turbulence equations, and Qv = (p, u, v, T, k, x)T is the primitive variable vector, which is related to Q by the Jacobian P = oQ/ oQv. Discretizing equation (9) by an Euler implicit scheme in the pseudo-time and approximating the physical-time derivative by second-order-accurate three-point backward dierences, the following equation in delta form is obtained:    3 Ds o o o P Ds Av Rnn Rng C 2 Dt on on og   o o o Bv Rgg Rgn Ds DQv og og on  r  3Q 4Qn Qn1 r R ; Ds 10 2Dt where r and Ds indicate the pseudo-time level and step, n and Dt indicate the physical-time level and step, Av = oE/ oQv, Bv = oF/oQv, Rij are the viscous coecient matrices [8], and the matrix C is evaluated as proposed in [3,4]. The residual is given as Rr oEr Er oF r F r v v Dr ; on og 11

In the equations above, U and H are the total energy and enthalpy comprehensive of the turbulent kinetic energy, k; the eddy viscosity, lt, is dened in terms of k and of the specic dissipation rate, x, according to the kx turbulence model of Wilcox [7], namely: qk . 6 x Moreover, ^ij indicate the sum of the molecular and s Reynolds (sij) stress tensor components. According to the Boussinesq approximation, one has:   oui ouj 2 ouk 2 ^ij l lt s dij qkdij . 7 3 oxj oxi 3 oxk lt c Finally, the heat ux vector components, qj, are given as   l lt oh qj ; 8 Pr Prt oxj where Pr = 0.71 and Prt = 1 are the laminar and turbulent Prandtl numbers, respectively. The Sutherland law is used to compute the molecular viscosity coecient. Finally, the standard coecients of the turbulence kx model are used [7], namely: b 3 ; 40 b 9 ; 100 5 c ; 9 c 1; 1 r r . 2

and the delta unknowns to be annihilated at every pseudotime level are DQv Qr1 Qr . v v 12

The left-hand-side (LHS) of Eq. (10) is modied to improve the eciency of the method, without aecting the residual, that is, the physical solution. Firstly, the non-orthogonal viscous coecient matrices, Rng and Rgn, are neglected, and the remaining ones are approximated by the corresponding spectral radii multiplied times the identity matrix, Rnn = RnI and Rgg = RgI; then, as proposed in [8], the pseudo- and physical-time terms are grouped together into a new term S, SC 3 Ds P; 2 Dt 13

The numerical method employed to solve the RANS equations is described in the following with reference to the twodimensional case, for simplicity. The system of equations is written either in Cartesian or generalized curvilinear coordinates, (n, g); a pseudo-time derivative is added to the lefthand-side in order to use a time marching approach for both steady state and unsteady problems; the preconditioning matrix, C, proposed in [1,2] is nally used to premultiply the pseudo-time derivative in order to improve eciency. The nal system reads C oQv oQ oE oF oEv oF v D; ot on og os on og 9

which is factored out of the LHS in Eq. (10), yielding      o o 1 o 1 o S I DsS A v Rn I B v Rg I DsS on on og og  r  n n1 3Q 4Q Q Rr . DQv Ds 14 2Dt In order to solve the resulting linear system, the diagonalization procedure of Pulliam and Chausee [9] is rstly applied, so that the matrices S1Av and S1Bv can be written as S1 Av Mn Kn M1 ; n S1 Bv Mg Kg M1 ; g 15

where Mn, Mg are the right-eigenvector matrices, M1 , M1 n g are the left-eigenvector matrices; and Kn and Kg are diago-

P. De Palma et al. / Computers & Fluids 35 (2006) 693702

695

nal matrices containing the eigenvalues of S1Av and S1Bv, respectively; then, the LHS of Eq. (14) is factorized,    o o K n Rn I M1 Mg SMn I Ds n on on    o o K g Rg I I Ds M1 DQv g og og  r  3Q 4Qn Qn1 Rr ; Ds 16 2Dt and solved by a standard scalar alternating direction implicit procedure [4]. A cell-centred nite volume space discretization is used on a multi-block structured mesh. A thirdorder-accurate ux vector splitting scheme is employed to discretize the convective terms, the minmod limiter being applied in the presence of shocks, whereas the viscous terms are discretized by second-order-accurate central differences. Further details of the method can be found in Ref. [8], which is available on the web [10], together with the code developed at the Pennsylvania State University. 3. Immersed-boundary technique The immersed boundary (IB) technique used in this work is based on that proposed in [5,6]. In a preliminary step, the geometry under consideration, which is described by a closed polygon in two dimensions (a closed surface in three dimensions), is overlapped onto a Cartesian (nonuniform) grid. Using the ray tracing technique based on the geometrical algorithms reported in [11], the computational cells occupied entirely by the ow are tagged as internal cells; those whose centres lie within the immersed body are tagged as external cells; the remaining ones are nally tagged as interface cells. The main feature of the IB technique is the evaluation of the unknowns at the centres of the interface cells. Here, the direct forcing method proposed by Mohd-Yusof [12] is employed. For each interface cell, the shortest Cartesian distance between the cell centre and the solid wall is determined. Along the corresponding direction, see Fig. 1, the variables at the centre of the interface cell (point P) are linearly interpolated between the values to be imposed at the boundary point (point B) and the computed values at the neighbouring internal-cell centre (point A), except for the pressure, whose value at point P is set equal to that at point A, which amounts to imposing a rst-order-accurate homogeneous Neumann condition for the pressure. In the present work, Dirichlet boundary conditions are imposed to u1 and u2 (the two Cartesian velocity components), T (the temperature) and, for turbulent ows, k and x. As shown in [5], such an approach provides an essentially second-order-accurate solution. It is noteworthy that, for high Reynolds number ows, a nonlinear interpolation procedure which utilizes adaptive wall functions is needed to obtain accurate results. Needless to say, the governing equations are solved at all internal cell centres, whereas all unknowns are set to zero at external ones.

- Interface cell - External cell - Internal cell Computed Interpolated

^ ^
A P B

immersed boundary

Fig. 1. Schematic representation of the interpolation scheme for the ow variable at interface cells.

4. Results The proposed methodology has been applied to compute two-dimensional steady and unsteady ows, for a wide range of the Reynolds and Mach numbers. 4.1. Incompressible ow past a circular cylinder The two-dimensional incompressible ow past a circular cylinder has been considered at rst to test both the preconditioning strategy and the immersed-boundary method versus steady as well as unsteady ows at very low Mach numbers. A single value of the free-stream Mach number, M1 = 0.03, and four values of the Reynolds number, based on the cylinder diameter, D, the free-stream velocity, U1, and kinematic viscosity, m1, namely, 20, 40, 100, and 200, have been considered; the rst two cases correspond to steady ow regimes and the last ones to unsteady ones. A rectangular computational domain has been used with the inlet and outlet vertical boundaries located at

1.8 1.6 1.4 1.2

1 0.8 0.6 0.4 0.2 0 -1.5 -1 -0.5 0 0.5 1 1.5

Fig. 2. Local view of the grid.

696
1.5

P. De Palma et al. / Computers & Fluids 35 (2006) 693702

1.5

0.5

0.5

-0.5

-1

Y
-1 -0.5 0 0.5 1 1.5 2 2.5 3 3.5

-0.5

-1

-1.5

-1.5

-1

-0.5

0.5

1.5

2.5

3.5

Fig. 3. Steady streamlines for Re = 20 (left) and Re = 40 (right).

xi = 10D and xo = 40D and the two horizontal boundaries located at yf = 20D, respectively, the origin coinciding with the centre of the cylinder. Standard characteristic boundary conditions have been imposed at inlet and outlet points, whereas free-shear wall boundary conditions are imposed at the points on the far-eld horizontal boundaries. Computations have been performed using a structured, non-uniform, single-block grid with 244 168 cells. A local view of the mesh is given in Fig. 2. Fig. 3 shows the steady streamlines corresponding to Re = 20 and Re = 40, respectively. Finally, the computed geometrical properties of the symmetrical vortices, as dened in Fig. 4, and the drag coecient, CD, are provided in Tables 1 and 2, together with the corresponding experimental [13,14] and numerical [1517] results available in the literature. The agreement is quite satisfactory. For both computations, the steady solver has been used, which annihilates the physical time derivative and iterates till the steady residual is reduced to 106. Convergence is achieved within about 18,220 iterations, corresponding to 5160 CPU seconds on a single processor Pentium IV (2.6 GHz). Concerning the Re = 100 and Re = 200 unsteady ow computations, the non-dimensional physical time step has been set equal to 0.03, which corresponds to about 200

Table 1 Steady ow past a circular cylinder at Re = 20 L Fornberg [15] Dennis and Chang [16] Coutanceau and Bouard [13] Tritton [14] Linnick and Fasel [17] Present 0.91 0.94 0.93 0.93 0.93 a 0.33 0.36 0.36 b 0.46 0.43 0.43 h 45.7 43.7 45.0 43.5 44.6 CD 2.00 2.05 2.09 2.06 2.05

Table 2 Steady ow past a circular cylinder at Re = 40 L Fornberg [15] Dennis and Chang [16] Coutanceau and Bouard [13] Tritton [14] Linnick and Fasel [17] Present 2.24 2.35 2.13 2.28 2.28 a 0.76 0.72 0.72 b 0.59 0.60 0.60 h 55.6 53.8 53.8 53.6 53.8 CD 1.50 1.52 1.59 1.54 1.55

L a

steps per period. About 200 inner iterations are needed to reduce the unsteady residual to 106 at every physical time step, corresponding to about 63 CPU seconds on the aforementioned processor. Two snapshots of the vorticity contours are given in Fig. 5; in both cases, the lift and drag coecients have regular sinusoidal behavior in time, as shown in Fig. 6 for Re = 200. Finally, the computed Strouhal number based on the shedding frequency, f (St = fD/ U1), as well as the drag and lift coecients are given in Tables 3 and 4 together with the experimental [18] and numerical [17,19,20] results available in the literature. Also for these unsteady ow cases, a very good agreement is obtained. 4.2. Unsteady ow past a heated circular cylinder

Fig. 4. Denitions of the relevant geometrical parameters of the symmetric separation region behind the cylinder.

The unsteady very low-Mach number ow past a heated circular cylinder has been chosen in order to validate the proposed method for a ow in which the energy equation plays a signicant role, insofar as experimental [21,22]

P. De Palma et al. / Computers & Fluids 35 (2006) 693702

5 4 3 2 1 0 -1 -2 -3 -4 -5 -2 5 4 3 2 1 0 -1 -2 -3 -4 -5 -2

697

10

12

10

12

Fig. 5. Snapshot of the vorticity contours for Re = 100 (left) and Re = 200 (right).

0.8 0.6

1.4

1.38 0.4 1.36 0.2

CL

CD
9 10 11 12

1.34

-0.2 1.32 -0.4 1.3 -0.6 -0.8 8 1.28 8

10

11

12

Time (s)

Time (s)

Fig. 6. Time history of the lift and drag coecient for Re = 200.

and numerical [23] investigations indicate that the temperature eld has a signicant inuence on the ow pattern, especially when the ratio between the cylinder wall temperTable 3 Unsteady ow past a circular cylinder at Re = 100 St Berger and Willie [18] Liu et al. [19] Linnick and Fasel [17] Present 0.160.17 0.165 0.166 0.163 CD 1.35 0.012 1.34 0.009 1.32 0.01 CL 0.339 0.333 0.331

ature, Tw, and the free-stream one, T1 (Tw = Tw/T1) exceeds 1.1. In particular, it has been found that, for a given Re1, the vortex shedding frequency, f, i.e., the Strouhal number, St, decreases for increasing values of Tw. Furthermore, by dening an eective Reynolds number, Ree, in terms of the kinematic viscosity corresponding to the maximum temperature in the wake, Te (which happens to be at the centre of the rst vortex shed by the cylinder, the curves St versus Ree, obtained for dierent values of Tw in the range 1 6 Tw 6 2, collapse into a single curve; see Ref. [21], where the following empirical correlation was proposed and employed to compute Te: T eff T 1 0:28T w T 1 . 17 In the present work, the same rectangular domain employed in the previous test-case has been used and discretized with a much ner (non-uniform) multi-block Cartesian grid with 142,200 cells and 12 blocks, see Fig. 7, where a local view of all blocks is provided (only one every ve grid-lines is plotted). Computations have been performed on a cluster with 12 processors for M1 = 0.01 and several values of Re and Tw. The non-dimensional time step has been set equal to 0.012, which corresponds to about 500

Table 4 Unsteady ow past a circular cylinder at Re = 200 St Berger and Willie [18] Belov et al. [20] Rogers and Kwak, reported in [20] Miyake et al., reported in [20] Liu et al. [19] Linnick and Fasel [17] Present 0.180.19 0.193 0.185 0.196 0.192 0.197 0.190 CD 1.19 0.042 1.23 0.050 1.34 0.043 1.31 0.049 1.34 0.044 1.34 0.045 CL 0.64 0.65 0.67 0.69 0.69 0.68

698
0.003

P. De Palma et al. / Computers & Fluids 35 (2006) 693702

0.2 0.19

0.002

0.18 0.17

*
+

0.001
0.16

St

0.15 0.14
*

+ - T*=3.5

-0.001

0.13 0.12

-0.002

0.11 0.1 50 60 70 80 90 100 110 120 130 140

- T*=3.0 - T*=2.5 - T*=2.0 - T*=1.8 - T*=1.6 - T*=1.5 - T*=1.4 - T*=1.1 - T*=1.0

-0.003 -0.002

150

160

-0.001

0.001

0.002

0.003

0.004

0.005

0.006

Reeff

Fig. 7. Local view of the multi-block grid.

Fig. 9. Strouhal number versus eective Reynolds number for ow past a heated cylinder: comparison between experimental data (open circles) for Tw = 1 and numerical results for Tw > 1.

steps per period. About 200 inner iterations are needed to reduce the unsteady residual to 106 at every physical time step, corresponding to about 55 CPU seconds. The computed values of St for Re1 = 80, 100, 120, 140 and Tw = 1, 1.1, 1.5, 1.8 are provided in Fig. 8, together with the experimental results of [21]: a very good agreement is obtained. Then, numerical results have been obtained for a wider range of Tw, for Re1 = 140 and 260. The computed values of St versus Ree are given in Fig. 9, together with those already shown in Fig. 8. A snapshot of the temperature contours for Tw = 2 and Re1 = 260 is provided in Fig. 10, for completeness. Finally, Table 5 provides the values of St and Te computed either at the centre of the rst 1 2 shed vortex, T eff , see Fig. 10 or using Eq. (17), T eff . The present results conrm the experimental data of [21], together with the validity of Eq. (17) in the range 1 6 Tw 6 2; moreover, they show that for Re1 = 140 and Tw = 3.5 the ow becomes steady and for Re1 = 260 and Tw P 2.5 the

values of St do not t the universal curve St(Ree), probably because a change in the physical nature of the unsteady phenomenon occurs at 140 < Re < 260. Notice that for Re1 = 260 the values corresponding toRee > 160 are not plotted insofar as they lie outside the range of laminar periodic wakes (they fall in the A-mode transition region for the case of unheated cylinders). Finally, for the two Re1 = 260 cases with Tw = 1.6 and Tw = 3, the computed values of Te are in good agreement with the experimental data of [22] for the maximum temperature in the wake, namely, Tmax = 343.5 K for Tw = 1.61 and Tmax = 484.1 K for Tw = 2.98. 4.3. Supersonic ow past an NACA0012 airfoil In order to test the proposed methodology versus a well documented viscous ow at high Mach number, the laminar supersonic ow past an NACA0012 airfoil with M1 = 2, a = 10 and Re1 = 1000 has been considered [24]. Three grids with 1252, 2502, and 5002 cells have been used to discretize the computational domain [8c; 9c] [8c; 8c], c being the chord-length of the airfoil, whose

0.2 0.19 0.18 0.17

0.002

0.16

St

0.15
0

0.13 0.12 0.11 0.1 60 70 80 90 100 110 120 130 140

-0.002

- T*=1.0 - T*=1.1 - T*=1.5 - T*=1.8

-0.004

0.14

Teff
0 0.002 0.004 0.006 0.008 0.010 0.012

150

160

Re

Fig. 8. Strouhal number versus Reynolds number for ow past a heated cylinder: comparison between experimental (open symbols) and numerical (solid symbols) data.

Fig. 10. Snapshot of the temperature contours for Tw = 2.0 and Re1 = 260: the point at which Te has been evaluated is indicated.

P. De Palma et al. / Computers & Fluids 35 (2006) 693702 Table 5 Temperatures in K Re = 140 T
w

699

Re = 260
1 T eff 2 T eff

St

(1)

Tw 1.1 1.4 1.6 1.8 2.0 2.5 3.0 3.5

T eff

T eff

St(1) 0.191 0.189 0.188 0.187 0.185 0.183 0.177 0.171

1.1 1.4 1.6 1.8 2.0 2.5 3.0 3.5

303.5 329.6 345.1 364.2 382.2 422.5 476.2

303.3 328.0 344.6 361.1 377.6 418.9 460.2 501.5

0.178 0.170 0.167 0.163 0.161 0.150 0.140

303.7 329.4 346.2 364.4 381.5 423.9 475.0 520.0

303.3 328.0 344.6 361.1 377.6 418.9 460.2 501.5

(1) Present computations, (2) Eq. (17).

0.25
1

0.5

-0.25

Pressure coefficient

-0.5

-0.75

-1

-0.5

-1.25

-1.5
-1 -1 -0.5 0 0.5 1 1.5 2

500 x 500 250 x 250 125 x 125

-1.75

Fig. 11. Local view of the mesh for the ow past an NACA0012 airfoil.

-2

0.25

0.5

0.75

x/c

leading edge is located at the origin; Fig. 11 shows a local view of the nest mesh (only one every ve grid lines is plotted), which is partitioned into 18 blocks, for parallel computing. A benchmark solution has also been obtained for comparison purposes, using the present approach on a very ne body-tted grid (87,500 cells with 12 blocks). The pressure coecient distributions along the prole are given in Fig. 12 where the nest grid solution, which coincides with the aforementioned benchmark one, within plotting accuracy, is clearly seen to be grid-converged. The lift and drag coecients obtained on the three grids are equal to 0.3296, 0.3335, 0.3353, and 0.2448, 0.2485, 0.2514, respectively, which tend towards the values of 0.3400 and 0.2515, obtained from the aforementioned benchmark solution, as the mesh is rened. Finally, the Mach number contours computed on the nest grid are provided in Fig. 13 showing that the shock is computed monotonically. Using the steady solver, reducing the residual to 106 on the nest grid requires about 105 iterations, corresponding to 2 105 CPU seconds on a single processor Pentium IV (2.6 GHz). Two comments are in order. All of the lift and drag coefcients obtained from computation using the IB approach and Cartesian grids are computed performing a momentum balance of the uid comprised within a rectangle sur-

Fig. 12. Pressure coecient distributions along the NACA0012 prole.

rounding the body. It has been veried that by varying the rectangle dimensions, the computed results vary less than one-tenth of 1%. In the heated cylinder case, the block containing the cylinder contains about one-fth of the total number of cells; in the present ow case, the 18 blocks all contain about the

1.5

0.5

-0.5

-1

-1

-0.5

0.5

1.5

2.5

3.5

Fig. 13. Mach number contours (DM = 0.1).

700

P. De Palma et al. / Computers & Fluids 35 (2006) 693702

same number of cells. Accordingly, the CPU time per iteration and grid point required by the parallel computations are ve and 18 times lower than those required by the single processor calculations, respectively. 4.4. Supersonic ow past a circular cylinder Finally, supersonic turbulent steady ows past a circular cylinder at Re1 = 2 105 and M1 = 1.3 and 1.7, have been considered as formidable test-cases for the proposed IB method. The inlet values of the turbulence kinetic energy and specic dissipation rate are k=U 2 0:0009 1 and xD/U1 = 40, respectively. For the considered values of M1, a bow shock is formed upstream of the cylinder; the subsonic ow at the front part close to the wall accelerates along the surface forming a supersonic-ow region, enveloping the subsonic recirculation region behind the cylinder, and two symmetric tail shocks are formed at the end of the separation region. Results have been obtained using a rectangular computational domain [8D; 9D] [8D; 8D], D being the diameter of the cylinder centred at the origin. Two non-uniform Cartesian grids have been employed for the IB computations, the rst one being a single-block grid with 430 200 cells and the second one having 10 blocks and 949,946 cells; the average values of y+ corresponding to the rst grid point away from the cylinder are about 50 and 10, respectively. A partial view of the ner grid is shown in Fig. 14, where only one every 10 grid lines is plotted and the thick horizontal lines delimitate each block. Furthermore, two computations have been performed using body-tted grids having 12 blocks and 99,913 and 188,423 cells (y+ = 0.1 and 0.05), for comparison. Figs. 15 and 16 provide a local view of the Mach number contours around the cylinder obtained using the IB method on the ner grid. It is noteworthy that the length of the recirculation region decreases as the Mach number increases. The computed positions of the separation point, hs, h being the clockwise angle measured from the leading
Y

-4

-2

Fig. 15. Local view of the Mach number contours for M1 = 1.3 (DM = 0.08).

Y
2 0

-4

-2

Fig. 16. Local view of the Mach number contours for M1 = 1.7 (DM = 0.08).

Table 6 Supersonic ow past a circular cylinder: separation-point angle, hs Immersed boundary Coarse-grid M1 = 1.3 M1 = 1.7 107 118 Fine-grid 105 111 Body-tted Coarse-grid 102 111 Fine-grid 103 112 Ref. [25] 103 112

edge, are given in Table 6, together with the corresponding experimental data of [25]; moreover, the computed and experimental drag coecients are reported in Table 7. All numerical results agree reasonably well with the experimental data; in particular, also the coarse-grid results are satisfactory in spite of the inadequate resolution of the boundary layer region. Finally, the computed pressure

-1

Table 7 Supersonic ow past a circular cylinder: drag coecient, CD

-2

Immersed boundary
-3 -2 -1 0 1 2 3

Body-tted Coarse-grid 1.45 1.40 Fine-grid 1.45 1.40

Coarse-grid M1 = 1.3 M1 = 1.7 1.46 1.38

Fine-grid 1.44 1.39

Ref. [25] 1.48 1.43

Fig. 14. Local view of the mesh for the supersonic ow past a cylinder.

P. De Palma et al. / Computers & Fluids 35 (2006) 693702

2 1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 -0.2 -0.4 -0.6 -0.8 0

701

Immersed Boundary (fine grid) Immersed Boundary (coarse grid) Experimental Body Fitted (fine grid) Body Fitted (coarse grid)

rather than to inadequate turbulence modelling, insofar as they are equally important in both the attached and separated ow regions. 5. Conclusions A state-of-the-art method for solving the preconditioned compressible NavierStokes equations accurately and eciently for a wide range of the Mach number is combined with an immersed-boundary approach which allows to use Cartesian grids for arbitrarily complex geometries. The methodology has been applied to compute steady and unsteady ows past circular cylinders and an NACA0012 airfoil for a wide range of the Reynolds and Mach numbers demonstrating its versatility as well as its accuracy for moderate values of the Reynolds number. For high-Reynolds number ows the proposed approach is not yet competitive, insofar as it requires huge numbers of grid points to provide accurate and reliable solutions. A local renement strategy or non-linear wall-laws, which can mimic an accurate resolution of the viscous sub-layer, are needed to obtain a state-of-the-art tool for investigating three-dimensional ows of industrial interest. Both approaches are currently under investigation by the present as well as other researchers. Acknowledgments

Cp

20

40

60

80

100

120

140

160

180

Fig. 17. Pressure coecient distribution along the surface of the cylinder: comparison between experimental and numerical results for M1 = 1.3.

2 1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 -0.2 -0.4 -0.6 -0.8 0

Immersed Boundary (fine grid) Immersed Boundary (coarse grid) Experimental Body Fitted (fine grid) Body Fitted (coarse grid)

Cp

20

40

60

80

100

120

140

160

180

Fig. 18. Pressure coecient distribution along the surface of the cylinder: comparison between experimental and numerical results for M1 = 1.7.

coecient distributions along the surface of the cylinder are provided in Figs. 17 and 18 together with the experimental data of [25]. The two body-tted solutions are grid-converged, insofar as they coincide within plotting accuracy. The coarser-grid IB solution is not satisfactory, especially in the aft-body separation region, whereas the ner-grid solution is smoother and tends towards grid convergence. On the other hand, the IB approach, at its present state, is not yet competitive for computing high Reynolds number ows. In this respect, it is noteworthy that for the ner-grid solution the steady-state solver requires about 150,000 iterations to reduce the residual to 106, corresponding to 137,700 CPU seconds using 10 Pentium IV (2.6 GHz) processors. However, the more accurate coarse-grid body-tted solution requires about 50,000 iterations to reduce the residual to 106, corresponding to 54,000 CPU seconds using a single processor. A nal comment on the experimental results is warranted. In Ref. [25], the authors provide numerical solutions which coincide within plotting accuracy with the present body-tted ones. The minor discrepancies between the numerical solutions and the experimental results are thus believed to be due to three-dimensional or wall eects in the experiments,

This research has been supported by MIUR and Politecnico di Bari, grants Con2003 and Conlab2000. The authors are grateful to their colleagues of Pennsylvania State University, who have developed and made available the basic computer code, as well as to G. Iaccarino and R. Verzicco for valuable suggestions and discussions on the immersed boundary method. References
[1] Venkateswaran S, Weiss S, Merkle CL, Choi YH. Propulsion related ow elds using the preconditioned NavierStokes equations. AIAA Paper 92-3437, 1992. [2] Merkle CL. Preconditioning methods for viscous ow calculations. In: Hafez M, Oshima K, editors. Computational uid dynamics review 1995. John Wiley & Sons; 1995. p. 41936. [3] Venkateswaran S, Merkle CL. Dual time stepping and preconditioning for unsteady computations. AIAA Paper 95-0078, 1995. [4] Buelow PEO, Schwer DA, Feng J-Z, Merkle CL, Choi D. A preconditioned dual time diagonalized ADI scheme for unsteady computations. AIAA Paper 97-2101, 1997. [5] Fadlun EA, Verzicco R, Orlandi P, Mohd-Yosuf J. Combined immersed-boundary nite-dierence methods for three-dimensional complex ow simulations. J Comp Phys 2000;161:3560. [6] Iaccarino G, Verzicco R. Immersed boundary technique for turbulent ow simulations. Appl Mech Rev 2003;56:33147. [7] Wilcox DC. Turbulence models for CFD. second ed. DCW Industries, Inc.; 1998. [8] Schwer DA. Numerical study of unsteadiness in non-reacting and reacting mixing layers. PhD thesis, Department of Mechanical Engineering, The Pennsylvania State University, 1999.

702

P. De Palma et al. / Computers & Fluids 35 (2006) 693702 [18] Berger E, Willie R. Periodic ow phenomena. Ann Rev Fluid Mech 1972;4:31340. [19] Liu C, Zheng X, Sung CH. Preconditioned multigrid methods for unsteady incompressible ows. J Comp Phys 1998;139:3557. [20] Belov A, Martinelli L, Jameson A. A new implicit algorithm with multigrid for unsteady incompressible ows calculations. AIAA Paper 95-0049, 1995. [21] Wang A-B, Travncek Z, Chia K-C. On the relationship of eective Reynolds number and Strouhal number for the laminar vortex shedding of a heated circular cylinder. Phys Fluids 2000;12(6): 14011410. [22] Yahagi Y. Structure of two-dimensional vortex behind a highly heated cylinder. Trans Jpn Soc Mech Eng Ser B 1998;64:209. [23] Sabanca M, Durst F. Flows past a tiny circular cylinder at high temperature ratios and slight compressible eects on the vortex shedding. Phys Fluids 2003;15(7):18219. [24] Bristeau MO, Glowinski R, Periaux J, Viviand A, editorsNumerical simulation of compressible NavierStokes ows. Notes on numerical uid mechanics, vol. 18. Vieweg; 1987. [25] Bashkin VA, Vaganov AV, Egorov IV, Ivanov DV, Ignatova GA. Comparison of calculated and experimental data on supersonic ow past a circular cylinder. Fluid Dyn 2002;37(3):47383.

[9] Pulliam TH, Chaussee DS. A diagonal form of an implicit factorization algorithm. J Comp Phys 1981;39:34763. [10] Available from: http://duns.sourceforge.net. [11] ORourke J. Computational geometry in C. Cambridge University Press; 1998. [12] Mohd-Yosuf J. Combined immersed-boundary/B-spline methods for simulations of ow in complex geometries. Annual Research Briefs, Center for Turbulence Research, 1997. p. 31728. [13] Coutanceau M, Bouard R. Experimental determination of the main features of the viscous ow in the wake of a circular cylinder in uniform translation. Part 1. Steady ow. J Fluid Mech 1977;79: 231256. [14] Tritton DJ. Experiments on the ow past a circular cylinder at low Reynolds number. J Fluid Mech 1959;6:54767. [15] Fornberg B. A numerical study of the steady viscous ow past a circular cylinder. J Fluid Mech 1980;98:81955. [16] Dennis SCR, Chang G-Z. Numerical solutions for steady ow past a circular cylinder at Reynolds number up to 100. J Fluid Mech 1970;42:47189. [17] Linnick MN, Fasel HF. A high-order immersed boundary method for unsteady incompressible ow calculations. AIAA 2003-1124. In: 41st AIAA aerospace sciences meeting and exhibit, Reno, NV, 69 January 2003.