Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
453468, 1997
0 1997 Elsevier Science Ltd. All rights reserved
Pergamon Printed in Great Britain
PII: SO0457930(97)00004-2 0045-7930/97 $17.00 + 0.00
Ahstract~~alculations for the two-dimensional driven cavity incompressible flow problem are presented.
A p-type Jinite element scheme for the fully coupled stream function-vorticity formulation of the
Navier-Stokes equations is used. Graded meshes are used to resolve vortex flow features and minimize
the impact of comer singularities. Incremental continuation in the Reynolds number allows solutions to
be computed for Re = 12 500. A significant feature of the work is that new tertiary and quaternary corner
vortex features are observed in the flow field. Comparisons are made with other solutions in the literature.
0 1997 Elsevier Science Ltd.
1. INTRODUCTION
This study presents results from a detailed set of calculations done for the steady 2D driven cavity
problem. This problem has been studied extensively in the literature as a validation problem for
2D incompressible flow algorithms [l-20]. Results are available for finite element, finite
difference/volume, exponential approximation and spectral methods. The present study combines
several unique features not found elsewhere in the literature: A p-type finite element formulation
combined with a strongly graded and refined element mesh provide for a highly accurate
calculation. Use of parallel computing resources and incremental continuation allow a range of
Reynolds numblers up to 12,500 to be explored on fine meshes [21]. The study begins with a
summary of the stream function vorticity formulation. This is followed by examination of the
detailed features in the flow field and comparisons to results in the literature. Values of local stream
function extrema and separation points on the cavity walls are presented in tablular form to
conclude the stu.dy. The results appear to be the most accurate available to date for this standard
test problem. As this study deals with the steady problem directly, questions concerning loss of
stability and time varying behavior are not addressed.
453
454 E. Barragy and G. F. Carey
We wish to treat the coupled PDE system for stream function IJ and vorticity w corresponding
to steady, 2D incompressible flow in domain R,
s
n
Vw.Vr dx dy + Re (u.Vw)r dx dy =
sn
sn
(Vt,bVq - oq) dx dy = *q
s as2an
hold for admissible r, q. It is assumed that the boundary aR is divided into two parts, dR,, aR,.
ds (3)
On aR,, let t,,Q= g be given as a Dirichlet condition and therefore r = 0 on Xl,. From (3) it is clear
that unless o is also specified on aR, (resulting in q = 0 there), at,b/an must also be specified. This
is just the tangential component of the velocity. On aR2, II/is not specified. If r # 0, q # 0 one must
specify both at//an and so/an. Alternatively, one could specify o and ac#n. Other combinations
are possible and can be identified by assuming that r, q correspond to variations in $, o. As pointed
out in [24,27], it is clear that the vorticity boundary conditions are given implicitly by equation (3)
evaluated for nonvanishing test functions on the boundary. The approximate problem follows
on introducing a finite element discretization and basis and replacing II/, o, r, q above by &, wh,
rh, qh.
was adequate for the range of Re considered. The reader is referred to [29,30] for more details on
optimal grading and optimal h-p methods.
letter below each reference indicates the formulation: S-stream function vorticity, P-primitive
variables, B-biharmonic, L-lattice Boltzman. Results for the present study are from a p = 8,
h = l/32 graded mesh. Other results are as above from [2] and [7]; [15] where a second-order
accurate method on a uniform mesh with Crank Nicholson time-stepping was used; [19] apply an
8th-order difference scheme on a uniform mesh with Runga Kutta time-stepping and also provide
a mesh refinement study; [12] used a second-order accurate multigrid defect correction method on
a particulary fine but uniform grid; [16] gives data for a uniform mesh lattice Boltzman method.
Results in [13] apply a second-order method (upwind) on various uniform meshes. Additionally,
there are results from [14, 17, 18, 201 for a variety of methods and resolutions; but where no data
on primary vortex location was given. It should be acknowledged that in many of these references
the authors do not claim particularly accurate solutions. Their focus may be on methodology
and/or computational efficiency. However, it is still interesting to include them in the table for the
sake of comparison.
Examining Table 2 at a Reynolds number of 1000, one immediately notices the wide range of
values for the primary vortex strength. There are many aspects which could account for this wide
range: mesh spacing and grading, accuracy/convergence rate of the method employed and issues
such as upwinding, use of primitive variable or stream function-vorticity formulation, convergence
criteria and arithmetic precision. For a Reynolds number of 1000 some comparisons can be made.
The results of [ 181,while extending to very high Reynolds numbers, are for coarse uniform 50 x 50
meshes, and can be discounted. Similarly, in [14] a graded 40 x 40 mesh of bilinear elements is used,
which is probably too coarse, although the results agree to three digits with those for more refined
meshes. The results in [20] are for a second-order method on a uniform 128 x 128 mesh using
primitive variables. The resolution here seems less than or comparable to other results. Similarly,
in [17] the focus is on vorticity boundary conditions and a uniform 128 x 128 mesh is used. These
results appear somewhat over-diffusive. Results for a Lattice Boltzman Method on a uniform
256 x 256 mesh are given in [16]. These results appear to fall in line with the present results
and [2, 71. The results of [13] appear overly diffusive (an upwind method is used), while the results
in [15] are for a second-order method, but with only 128 x 128 mesh points. Wright and
Gaskell[12] give data for a uniform 1024 x 1024 mesh with second-order scheme and should be
quite accurate. Nishida and Satofuka [19] use an 8th-order method on a uniform 128 x 128 mesh.
Unfortunately their results and [12] do not extend to Re = 5000. Finally, Ghia et al. [2] use a
second-order method with 128 x 128 mesh points for this value of Re. These results might then
be considered somewhat under-resolved for Re = 1000. The results in [7] are extrapolated values
for mesh sizes of 100 x 100 to 141 x 141, for a fourth-order method. Of these, comparing the
vortex strength for the present results to those of [7, 12, 191 show all agree to within 6 = 0.0002,
or almost four full digits.
At intermediate Reynolds numbers of 5000 and 7500, the present results agree well with those
for the Lattice Boltzman method [16], but not as well with [2]. None of the other methods agree
well with these values except for [14], although [20] and [2] agree well. Finally, at Re = 10 000, the
agreement between the present results and [7] is quite good 6 = 0.0006, while comparing to [2] the
difference in strength is larger 6 = 0.003. Overall, we believe the present results to be the most
accurate to date, although the methods used may be less efficient than others.
Another aspect of Table 2 which is somewhat disturbing is the behavior of the primary vortex
strength as the Reynolds number increases. For a sequence of three meshes (p = 8; h = l/8,
h = l/l 6, h = l/32) the primary vortex achieved minimum stream function value at Reynolds
numbers of approximately 7500,900O and 9500, respectively. Many of the results in the table show
a trend of increasing stream function value as the Reynolds number increases. For
example, [ 13, 15, 17, 181all show evidence of this, which appears to be characteristic of an under-
resolved solution or boundary layer. Even the results of [2] display this behavior to some degree.
In general, mesh grading and higher order methods appear to substantially improve the quality
of the solution. For example, results in [14] at Re = 1000, 5000 compare well with only a 40 x 40
graded mesh. The present results and those of [19] and [7], all using high order methods, appear
to be quite accurate. This is to be expected as resolving the boundary layer at higher Re is important
to obtaining the correct solution. Finally, it would appear that the use of upwinding is quite
detrimental, as seen by comparing the results of [13] with the present results or [2].
Stream function-vorticity driven cavity solution 457
2
Fig. 1. Stream function and vorticity contours for Re = 5000.
CAF 2615-B
458 E. Barragy and G. F. Carey
3
Fig. 2. Stream function and vorticity contours for Re = 7500.
Stream function-vorticity driven cavity solution 459
6
Fig. 5. Stream function and vorticity contours for Re = 16,000.
Stream function-vorticity driven cavity solution 463
.Q
._
E
.8
i .7
g .8
l2 .5
B
10”
.O .2 .4 .8 .8 1.0 1.2 1.4 1.8
Reynolds Number 4
x10
Fig. 7. Detail of flow separation points as Re varies.
corner of the cavity. The sixth and seventh columns show the attachment points of the tertiary
vortex in the upper left corner. Note that the attachment points of the secondary upper left
vortex bracket those of the tertiary vortex. Figure 6 gives a graphical presentation from a
larger data set.
Table 3. Primaryvortex
Re 0.0001 2500 5000
-
ILnun -O.lOQO758E+OO -0.1214621E+00 -0.1222194E+OO
&?r 0.5OO0OCOE+IXl 0.5188822E+OO 0.5151064E+ 00
Yk 0.76441628+00 0.5434181E+OO 0.5358696E+ 00
z- 7500 10,000 12,500
c - 0.12238038+ 00 -0.1223930E +OO -0.12235848+00
xx 0.51321848+00 0.5113304E+OO 0.5113304E+ 00
Y!+X 0.5320950E+00 0.5302077E+OO 0.52832028+00
464 E. Barragy and G. F. Carey
Table 4. Vortices
in lowerrightcorner
Re 0.0001 2500 5000
0.2227267&-05 0.2662249E-02 0.30735158-02
0.9622189E+OO 0.83423248+00 0.8041016E+ 00
0.3778106E-01 0.907512lE-01 0.72486528-01
-0.61415878-10 -O.l226317E-06 -O.l427913E-05
0.99771448+ 00 0.9903702E+00 0.97860178+OO
0.2285573E-02 0.9321324E-02 O.l881959E-01
O.l534814E-I4 0.3366884BII 0.38872898-10
0.9998708E+ 00 0.9994164E+ 00 0.9987615E+OO
O.l292307E-03 0.5836428E-03 0.11726798-02
- 0,5945803E-15 -0,2272706E-14
0,9999354E+ 00 0.9999354E+00
0.645819lE-04 0.6458191E-04
Re 7500 10.000 12.500
0,3226968E-02 0.3191218E-02 0.3099803E-02
0,7900250E+ 00 0.7746636E+00 0.7603326E+OO
0.6483485E-01 0.58780138-01 0,5407320E-01
- 0.32790118-04 - O.l404461E-03 -0.2558322B03
0.9517405E+00 0.9351651E+00 0.92751358+ 00
0.42189WE-01 0.67528338-01 0.81219448-01
0.8899567E-09 0.39496268-08 0.77503508-08
0.99734218+00 0.9960358E+OO 0.9952875E+OO
0.2657867E-02 0.3964201E-02 0.4899706B02
- 0.2126772E-13 - 0,1044644E-12 -0,2077842E-12
0.999806lE+OO 0.9997413E+OO 0.99967648+OO
O.l292307E-03 0.2587290E-03 0.25872908-03
Table 5. Vortices
in lowerleft
corner
RC? 0.0001 2500 5000
0.2227263E-05 0.9310542E-03 0.13765048-02
0.3778106E-01 0.8439557E-01 0.72486528-01
0.3778106E-01 O.l109646E+ 00 0.13702978+OO
-0.6141578E-IO - 0.28094618-07 -0.66651278-07
0.2285573&02 0.6023922E-02 0.78985718-02
0.22855738-02 0.62113898-02 0.7898571E-02
O.l534812E-14 0.75958178-12 O.l814952E-11
0.12923078-03 0.3884944E-03 0.5185267E-03
O.l292307E-03 0.38849448-03 0.4534772E-03
7500 10,000 12,500
0.15364668-02 O.l619575E-02 0.36677528-02
0.6416178E-01 0.5878013E-01 0.55418028-01
0.15258898+OO 0.1622930E+OO O.l680841E+OO
- 0.2043063E-06 -0.11334178-05 -0.6787536E-05
O.l117058E-01 O.l733886E-01 0.2655169E-01
0.11786448-01 0.2010944E-01 0.32X2161&01
0.5591540E-II 0.3074254E-IO 0.1828415E-09
0.6488255E-03 0.11069658-02 0.17670X0&02
0.7140747E-03 0.11069658-02 O.l767080E-02
-O.l141646E-I4 -0.22480008-14 - 0,4384933E-I4
0.64581918-04 0.645X1918-04 O.l292307E-03
0.6458191B04 0.645819lB04 0.645819lE-04
Table 6. Vortices
in upper left
comer
Re 0.0001 2500 5000
3.4. Accuracy
The question of the accuracy of the computed flow solutions presented in the previous section
is an important one. Results in the literature rarely present systematic grid refinement studies due
to the computational cost. Instead, comparisons to other results in the literature are made along
Stream function-vorticity driven cavity solution 465
with discussions of truncation error. Here an attempt to assess solution accuracy will be made based
on mesh refinement studies. Figure 8 compares the location of the lower flow separation point for
the primary vortex, on the left wall of the cavity for three successive h refinements of the mesh.
Values of the primary separation point were computed for h = l/8, l/16, l/32 (p = 8). The figure
shows the absolute values of the difference of the location for h = l/S, l/32 and h = l/16, l/32.
This says nothing specific about the error in the finite element solution. It does give some confidence
that the solution is converging as the mesh is refined and that the differences between successive
solutions are decreasing. As indicated, the average difference in the location of the separation point
is roughly 1.e-3 for the h = l/S solution and 3.e-5 for the h = l/l 6 solution. The coarse mesh
problem shows particularly large differences as the Reynolds number increases, as would be
expected.
This effect for the coarse mesh solution is shown more dramatically in Fig. 9. That figure
compares the st.ream function value at the center of the primary vortex for successive mesh
refinements. Stream function values are compared for h = l/S, l/16, l/32 (p = 8). The figure shows
the absolute values of the difference in tj for h = l/S, l/32 and h = l/16, l/32. As indicated, the
difference in $ decreases as the mesh is refined. However, the figure clearly indicates the problem
with coarse mesh solutions: the value for h = l/8 diverges radically from that for h = l/32 as the
Reynolds number increases, whereas up to Re = 14 000, the difference between h = l/16, l/32 is
roughly fixed at 1 x 10 - 5. As indicated, for the coarse mesh the calculations begin to diverge
around a Reynolds number of 5000.
On a qualitative level, Fig. 5 showing the flow contours at a Reynolds number of 16 000 is
interesting. The ,vorticity contours clearly display oscillations in three areas: in the contour lines
1o-3
lo4
!I
f
B 10”
los
+ h=lIiB, h-l/32
-6- h=1/8. hd32
lo-’
.u .2 A .6 .8 1.0 1.2 1.4 1.6
Reynolds Number x10 4
Fig. 8. Difference in successive approximations of flow separation points.
466 E. Barragy and G. F. Carey
on the upper right corner, on the right side of the cavity in the area just inside the primary vortex
boundary and in the contour lines towards the upper left corner. In the computations presented
here, this pattern was typical of an under-resolved mesh. All of these oscillations appear related
to boundary layer resolution issues.
In the tabular data for vortex location and strength it would appear that up to five nested vortices
have been detected in the flow field. Based on the location of sign changes in the vorticity, these
finest vortices appear at the grid frequency and grow with Re. The strength of these vortices shows
odd behavior; the value of stream function is very small for a large range of Re. At some critical
value it begins to grow appreciably. The question arises as to when these vortices can be considered
numerically significant features of the flow field.
A Newton tolerance of 10 -6 was applied to the RMS residual norms to assess convergence. One
might conclude that any vortex of strength less than 10m6is not numerically significant. However,
even for vortices of strength 0( 10 - I’), the data show remarkably consistent trends with respect to
the Reynolds number and vortex movement within the domain. To obtain a better estimate of what
might constitute a numerically significant value of the stream function, consider a tolerance of 10m6
and note that we have roughly 130,000 degrees of freedom in the finest mesh. The average error
at a given grid point can then be estimated from 10m6= (130,000~*)“*, where E represents the
average error. This yields a value of 10m9 for the average grid point error. Maximum observed
nodal residuals averaged roughly 10 - ‘Ofor vorticities and 10 -I4 for the stream function equation.
Given that the row norms of the Jacobian matrix are much greater than 1, then stream function
values of O(l0 -I”) would appear to be numerically significant. Data below this cutoff tolerance
may not be numerically significant. However, it exhibits a high degree of organization, which seems
unusual. All data, including that which may not be numerically significant, has been presented in
the tabular results.
4. CONCLUSION
High resolution calculations have been presented for the driven cavity problem. The calculations
have several unique features distinguishing them from other results in the literature. These include
o.jcK),
k...........................
i
i............................
::::::::::::::::::
____..___..___...____ i__..........................
::::::: j ::::::::::::::::::::::::::::
.______.____..______........
_____.____.______..........
<__......___________________
j..................._.........
j
<.............................
__._________
&:. ...........
;:_: ::::::::::::::::
._.._.__..____
.............................
:::::::::
+$i..__...____._.._.........
............................._..______.._..............
j..........................
j :::::::::::::: ::::::::::::::
j........_.._____..___........
A .j ::::::__
i...,Y’:::::::::::::::::::I::::::::::::::::::::::::::::;::::::::::::::::::::::::::
> )___________________.........
:
[ ._......................._.
t .-..........,....-.-..-.-.~....---.-------..-..........~..-.............-.--..---...~.................~....--..-~----..-.-.............
.-...~-.-.....-.-............-...~-.-..-
....-.-.....--._-..
le-06 1 I I I 1 I I 1
0 2ooo 4000 6ooo Nu12 10000 12000 14000
Reynolds
the use of a high ,u finite element approximation (p = 8) and a graded element mesh (30: 1) element
size ratio. These factors significantly improve solution accuracy and help to minimize the impact
of corner singularities. Several unique flow features have been observed which are not reported
elsewhere. This includes tertiary and quaternary corner vortices and an enclosed tertiary corner
vortex in the upper (upstream) corner. This vortex is attached to the cavity wall and enclosed by
the secondary corner vortex. The present results include comparison to other results in the literature
(to assess accuracy) as well as grid refinement studies of the location of flow separation points and
the stream function value for the primary vortex to assess solution accuracy. Detailed comparisons
are given as the Reynolds number varies.
AcknoM,ledgemenrs--This research has been supported by the Department of Energy CLS fellowship program and by
NASA.
REFERENCES
1. Cortes, A. B. and Miller, J. D., Numerical experiments with the lid driven cavity flow problem. Compurers & Fluids,
1994, 23, 1005-1027.
2. Ghia, U., Ghia, ;K. and Shin, C., High Re solutions for incompressible flow using the Navier-Stokes equations and
a multigrid method. Journal of Computarional Physics, 1982, 48, 387411.
3. Goodrich, J., Gustafson, K. and Halasi, K., Hopf bifurcation in the driven cavity. Journal of Computational Physics,
1990, 90, 219-261.
4. Gustafson, K. and Halasi, K., Cavity flow dynamics at higher Reynolds numbers and higher aspect ratios.
Journal of Computational Physics, 1987, 70, 271-283.
5. Huser, A. and Biringen, S., Calculation of two dimensional shear driven cavity flows at high Reynolds numbers.
International Journal of Numerical Methods in Engineering, 1992, 14, 1087-I 109.
6. Kim, J. and Moin, P., Application of a fractional step method to incompressible Navier-Stokes equations.
Journal of Computational Physics, 1985, 59, 308-323.
7. Schreiber, R. and Keller, H. B., Driven cavity flows by efficient numerical techniques. Journal of Computational Physics,
1983, 49, 31&333.
8. Schreiber, R. and Keller, H. B., Spurious solutions in driven cavity calculations. Journal of Computational Physics, 1983,
49, 165-172.
9. Shen, J., Hopf bifurcation in the unsteady regularized driven cavity flow. Journal of Computational Physics, 1991,
95, 228-245.
10. Shen, C. Y. and Reed, H. L., Application of a solution adaptive method to fluid flow: line and arclength approach.
Computers & Fluids, 1994, 23, 373-395.
Il. Sohn, J., Evaluation of FIDAP on some classical laminar and turbulent benchmarks. Infernational Journal of Numerical
Methods for Fluids, 1988, 8, 1469-14907.
12. Wright, N. G. and Gaskell, P. H., An efficient multigrid approach to solving highly recirculating flows. Computers &
Fluids, 1995, 24, 63-79.
13. Bruneau, C. H. and Jouron, C., An efficient scheme for solving steady incompressible Navier Stokes equations.
Journal of Computational Physics, 1990, 89, 389413.
14. Comini, G., Ma.nzan, M. and Nonino, C., Finite element solution of the streamfunction vorticity equations
for incompressible two dimensional flows. International Journal of Numerical Methods for Fluids, 1994, 19,
513-525.
15. Goyon, O., High Reynolds number solutions of Navier Stokes equations using incremental unknowns. Computer
Methods in Applred Mechanics and Engineering, 1996, 130, 319-335.
16. Hou, S., Zou, Q,, Chen, S., Doolen, G. and Cogley, A., Simulation of cavity flow by the lattice Boltzman method.
Journal of Computational Physics, 1995, 118, 329-347.
17. Huang, H. and Seymour, B., The no slip boundary condition in finite difference approximations. Infernational
Journal of Numerical Methods for Fluids, 1996. 22, 713-729.
18. Nallasamy, M. and Prasad, K., On cavity flow at high Reynolds numbers. Journal of Fluid Mechanics, 1977, 79,
391414.
19. Nishida, H. and Satofuka, N., Higher order solutions of square driven cavity flow using a variable order multigrid
method. International Journal of Numerical Methods in Engineering, 1992, 34, 6376531.
20. Thompson, M. C. and Ferziger, J. H., An adaptive multigrid technique for the incompressible Navier Stokes
equations. Journal of Computational Physics, 1989, 82, 94121.
21. Barragy, E., Carey, G. F. and Van de Geijn, R., Parallel performance and scalability for block preconditioned
finite element (p) solution of viscous flow. International Journal of Numerical Methods in Engineering, 1995, 38,
1535-1554.
22. Carey, G. F. and Oden, J. T., Finite Elements: Fluid Mechanics. Prentice Hall, Englewood Cliffs, NJ, 1986.
23. Roache, P. J., C~smputational Fluid Dynamics. Hermosa, Albuquerque, NM, 1972.
24. Campion-Renson, A. and Crochet, M. J., On the stream function-vorticity finite element solutions of Navier-Stokes
equations. International Journal of Numerical Methods in Engineering, 1978, 12, 1809-1818.
25. Hughes, T. J. R., Levit, 1. and Winget, J., Element-by-element solution algorithm for problems of structural and solid
mechanics. Computer Methods in Applied Mechanics and Engineering, 1983, 36, 241-254.
26. Barragy, L. and Carey, G. F., Stream function vorticity solution using high degree finite elements and element by
element techniques. Communications in Numerical Methods in Engineering, 1993, 9, 387-395.
27. Dennis, S. C. R.. and Quartapelle, L., Some uses of Green’s theorem in solving the Navier-Stokes equations.
International Journal of Numerical Methods for Fluids, 1989, 9, 871-890.
28. McLay, R. and Carey, G. F., Local pressure oscillations and boundary treatment. International Journal of Numerical
Methods for FlubIs, 1986, 6, 155-172.
468 E. Barragy and G. F. Carey
29. Babuska, I. and Guo, B., The h, p. h-p version of FEM for I-d problem: I p error analysis, 11 h, h-p error analysis,
III adaptive h-p. Numerical Mathematics, 1986, 49, 577-683.
30. Oden, J. T., Wu, W. and Legat, V., An hp adaptive strategy for finite element approximations of the Navier Stokes
equations. International Journal of Numerical Methoris ,for Fluids, 1995, 20, 83 I-851.