IDENTIFICATION OF SWIRLING FLOW

IN 3-D VECTOR FIELDS

David Sujudi* and Robert Haimes**

Department of Aeronautics and Astronautics
Massachusetts Institute of Technology
Cambridge, MA 02139

Abstract particularly pV3 [8], a scientific visualization software

An algorithm for identifying the center of swirling designed for co-processing and distributed
flow in 3-D discretized vector fields has been computation (i.e., the user can visualize the solution as
developed. The algorithm is based on critical point it is being computed on one or more computers).
theory and has been implemented as a visualization We believe that a tool that identifies the center of
tool within pV3, a package for visualizing 3-D swirling flows satisfies the above need and criteria.
transient data. The scheme works with gridding Investigators have been using swirling flow as one of
supported by pV3: structured meshes as well as the means to locate vortices in 3-D discretized vector
unstructured grids composed of tetrahedra, fields. Swirling flows in a 3-D field are usually
polytetrahedral strips, hexahedra, pyramids, and/or identified by studying vector fields that are mapped
prism cells. The results have been validated using onto planar cuts or by seeding streamlines. These
artificially-generated vector fields and 3-D CFD data. procedures can be very laborious, especially for large
Introduction and complex flows. The scheme presented here allows
automatic identification of the center of swirling flows
This work is motivated by the need to easily locate
in 3-D vector fields.
vortices in large 3-D transient problems. A tool that
will automatically identify such structures is definitely The algorithm for implementing this tool is based
needed to avoid the time-consuming and tedious task on critical-point theory. As will be described below,
of manually examining the data. However, the the scheme works on a cell by cell basis, lending itself
question of what defines a vortex raises considerable to parallel processing, and is flexible enough to work
confusion. As a result, various definitions have been with the various types of grids supported by pV3. By
proposed by investigators, including, among others, employing this method, we have avoided the need for
Moin and Kim [1] [2], Villasenor and Vincent [3], curve integration (which is needed for visualization
Globus, Levit, and Lasinski [4], Banks and Singer [5] tools such as streamlines and particle paths). Curve
[6], and Jeong and Hussain [7]. The reader can refer integration is a serial operation and can not readily
to [6] for a survey of the schemes that have been take advantage of pV3’s distributed computing
developed and [7] for comparisons of some of them capabilities. And, as described in [9], integrations
across a distributed environment involve passing
We see the need for a tool that can help
information between machines and other additional
investigators locate vortices, and yet has a familiar and
complexities which further reduce efficiency.
intuitive interpretation. Due to the confusion on what
constitutes a vortex, we relax the condition that what In the next sections, we will describe the
the tool finds must be vortices. However, the scheme theoretical background of this algorithm and how it is
must be practical, in terms of computational speed and implemented. The results of the scheme on exact
usability. We must also be able to easily integrate it artificially-generated data as well as CFD data will be
into a publicly-available visualization package, shown and its performance will be discussed. Our
algorithm will also be compared (using artificially-
__________________ generated data) against a tool developed by Globus,
* Graduate Research Assistant, Member AIAA Levit,and Lasinski [4].
** Principal Research Engineer, Member AIAA
Theory
Copyright  1995 by the American Institute of Critical points are defined as points where the
Aeronautics and Astronautics, Inc. streamline slope is indeterminate and the velocity is
All rights reserved. zero relative to an appropriate observer [10].

1
According to critical point theory, the eigenvalues and 3-D linear interpolant. More importantly, linear
eigenvectors of the rate-of-deformation tensor, ∂ ui /∂ xj velocity interpolation produces a constant rate-of-
(we’ll call this matrix A), evaluated at a critical point deformation tensor within the entire tetrahedral cell.
defines the flow pattern about that point. Specifically, Consequently, the straight forward algorithm described
if A has one real and a pair of complex-conjugate below can be employed, which otherwise would not
eigenvalues the flow forms a spiral-saddle pattern, as have been possible.
illustrated in Figure 1. The eigenvector corresponding The algorithm proceeds one tetrahedral cell at a
to the real eigenvalue points in the direction about time, and can be summarized as follows (it is assumed
which the flow spirals, and consequently, the plane that a velocity vector is available at each node):
normal to this eigenvector defines the plane on which
the flow spirals. For a complete description of all 1. Linearly interpolate the velocity within the cell.
other possible trajectories the reader can refer to [10] 2. Compute the rate-of-deformation tensor A. Since
or [11]. a linear interpolation of the velocity within the
Since we intuitively identify the pattern in Figure 1 cell can be written as
as swirling flow, we can use the above method to find
∂ ui ∂u ∂u
the center of swirling flows located at critical points. ui = Ci + ∆x + i ∆y + i ∆z (1)
However, there are obviously swirling flows whose ∂x ∂y ∂z
center is not at a critical point. Fortunately, a similar then A can be constructed from the coefficients of
method can be applied in these cases. the linear interpolation function of the velocity
vector.
At a non-critical point with the necessary
eigenvalue combination (i.e., one real and a pair of 3. Find the eigenvalues of A. Processing continues
complex conjugates) the velocity in the direction of only if A has one real (λR) and a pair of complex-
the eigenvector corresponding to the real eigenvalue is conjugate eigenvalues (λC).
subtracted. The invariance of the eigenvectors’
4. At each node of the tetrahedron, subtract the
directions with respect to a Galilean transformation
velocity component in the direction of the
ensures that the resulting flow will have the same
eigenvector corresponding to λR. This is
principal directions. We’ll call the resulting velocity
equivalent to projecting the velocity onto the
vector the reduced velocity. If the reduced velocity is
plane normal to the eigenvector belonging to λR ,
zero, then the point must be at the center of the
and can be expressed as
swirling flow. A similar statement was also made by
Vollmers, Kreplin, and Meier [12]. r r r r r
w = u − (u ⋅ n )n (2)
Therefore, to find a point at the center of a local r
swirling flow, we look for a point whose rate-of- where n is the normalized eigenvector
deformation tensor has one real and a pair of complex- corresponding to λR, and w is the reduced
conjugate eigenvalues and whose reduced velocity is velocity.
zero.
5. Linearly interpolate each component of the
Implementation reduced velocity to obtain
pV3 accepts structured and/or unstructured grids
(containing any combination of tetrahedra, wi = a i + bi x + ci y + d i z (3)
polytetrahedra strips, hexahedra, pyramids, and prism
cells). In the interest of efficiency, we have decided to i = 1, 2 , 3
use only tetrahedral cells, with all other cell types 6. To find the center, we set wi in equation (3) to
reduced to 2 or more tetrahedra. Figure 2 shows how zero. Since the reduced velocity lies in a plane, it
various types of cells are decomposed into tetrahedral has only 2 degrees of freedom. Thus, only 2 of
cells. the 3 equations in equation (3) are independent.
This approach enables us to use a simple linear Any 2 can be chosen as long as their coefficients
interpolation for the velocity, avoiding the more are not all zero. Now we have
complex, and inherently more costly, interpolation
required by other types of cells (such as bilinear 0 = ai + bi x + ci y + d i z (4)
interpolation for hexahedra). A tetrahedron has 4 node
points, sufficient to solve for the four coefficients of a i = 1, 2

2
 which are the equations of 2 planes, whose Table 1 Size of Test Cases
solution (the intersection of 2 planes) is a line.
Case Number of Number Number of Calculation
7. If this line intersects the cell at more than 1 point, Nodes of Cells Tetrahedral Time (sec.)**
then the cell contains a center of a local swirling Cells*
flow. The center is defined by the line segment
Cylinder 131072 123039 738234 34
formed by the 2 intersection points.
F-117 48518 240122 240122 16
Since the 2 intersection points lie on the line found
in step 6, the reduced velocity at those points must be * After decomposition (if needed) of original cells.
zero. This suggests a different (but equivalent) and ** On SGI Indigo2 with MIPS R4400 150 Mhz CPU.
more efficient way to finding the center. This
approach renders steps 5, 6, and 7 unnecessary and Results are shown in Figures 4 and 5. To indicate
replaces them with a new step 5: the existence swirling flow, streamlines have been
5. For each of the tetrahedron’s face, determine spawned near the centers found by the algorithm.
if there is exactly 1 point on the face where the These results indicate that the large coherent structures
reduced velocity is zero. If at the end there are exactly found by the algorithm do indeed correspond to
2 distinct points, then the cell contains a center, which centers of swirling flow. However, the algorithm does
is defined by those 2 points. not find all the swirling flow in the tapered cylinder
data. Missing are a few swirling flow structures
Both approaches have been tried on our test cases
further downstream of the cylinder, which are found
with identical results. Therefore, the second approach
by studying the vector field more closely. We believe
is implemented.
the size of the grid cells might be a factor. The cells
Results and Comparisons are larger away from the cylinder, reducing the
The algorithm is first tested on artificially- accuracy in the calculation of the rate-of-deformation
generated vector fields where the location of the center tensor (and consequently the reduced velocity).
of the swirling flow is known exactly. The field is Another possible cause is the algorithm’s sensitivity to
defined by the strength of the swirl flow. As shown in Figures 4b
and 4c, the structures are very coherent for strong
v = y − y , v = x − x , v = f ( z) (5)
swirls (i.e., the swirl velocity is larger than or
x c y c z comparable to the normal velocity). However, the
Note that this vector field has circular streamlines (in structures start to break up as the swirl weakens, and
the x-y plane) around a central axis whose location is further downstream, where the swirl flows are very
defined by xc and yc. The magnitude of the vector is weak, the algorithm finds no coherent structures.
equal to the distance from the central axis. The field is In the case of the F-117 data, the structures are less
discretized using an 11 x 11 x 11 node structured grid. coherent than in the tapered cylinder. The tetrahedral
The results for various values of x c, yc, and functional grid used in this data is very irregularly sized, and is
forms of vz (including constant, linear, and rather coarse. Comparison between the streamlines in
exponential) are studied and determined to be correct. Figures 4c and 5 also shows that the swirl in the F117
A sample result (with xc = 2.2 ∆x, yc = 1.5 ∆y, and vz = data is noticeably weaker. Both of these factors might
1) is shown in Figure 3. A streamline is also shown in contribute to the incoherency in the structures.
this figure to provide a sense of the swirling vector
field. We have also compared our results with that of
FAST’s vortex-core finder [4] [15]. FAST’s finder
Further tests are done using data from 3-D defines a vortex core by integrating from a critical
calculations of flow past a tapered cylinder [13] and of point in the direction of the eigenvector corresponding
flow over an F-117 fighter at an angle of attack [14]. to the only real eigenvalue of the rate-of-deformation
The tapered-cylinder calculation employs structured tensor. For this comparison, we use 3 artificially-
grid, while the F-117 case uses unstructured grid generated data sets, each bounded by a cube
composed of tetrahedra. The size of these data sets containing 3 randomly-placed vortices. The data
and the time needed to find the swirl flow centers are generator is developed by D. Asimov at NASA Ames
summarized in table 1. Research Center.
The comparisons are shown in Figures 6a to 6f.
Despite the lack of any 3-dimensional cues, the curves

3
in these figures do exist in 3-D space, and each pair of [4] A. Globus, C. Levit, and T. Lasinski, “A Tool for
figures are taken from the same view point. A high Visualizing the Topology of Three-Dimensional
degree of similarities are found in each case except for Vector Fields,” Report RNR-91-017, NAS
the middle curves in each data set, where FAST Applied Research Office, NASA Ames Research
produces curves that are either longer and/or has Center, 1991.
different orientation. Closer inspection of the data [5] D. C. Banks and B. A. Singer, “A Predictor-
shows that pV3’s results are the correct ones, while Corrector Scheme for Vortex Identification,”
FAST’s curve integrations veer away from the core. ICASE Report NO. 94-11, NASA CR-194882,
Conclusion 1994.
An algorithm to automatically locate the center of [6] D. C. Banks and B. A. Singer, “Vortex Tubes in
swirling flow in 3-D vector fields has been developed Turbulent Flows: Identification, Representation ,
and implemented as part of the pV3 visualization Reconstruction,” ICASE Report No. 94-22,
package. By employing cell-by-cell processing and NASA CR-194900, 1994.
using only tetrahedral cells (and transforming other
[7] J. Jeong and F. Hussain, “On the Identification of
cell types to tetrahedra), the scheme can take
a Vortex,” J. Fluid Mech. 285, pp. 69, 1995.
advantage of pV3’s distributed environment and the
simplification from using linear interpolation. [8] Robert Haimes, “pV3: A Distributed System for
Although we believe the strength of the swirl flow and Large-Scale Unsteady CFD Visualization,”
the coarseness of the grid can affect the degree of AIAA Paper 94-0321, 1994.
accuracy and coherency of the results, tests using [9] D. D. Sujudi and R. Haimes,”Integration of
artificially-generated vector fields and 2 different CFD Particle Paths and Streamlines in a Spatially-
data have shown that the coherent structures found by Decomposed Computation,” Parallel CFD
the algorithm are indeed centers of swirling flow. Conference, Pasadena, CA, June 26 - 28, 1995.
Comparisons with FAST vortex-core finder also show
high degrees of similarities in the results. [10] M.. S. Chong, A. E. Perry, and B. J. Cantwell,
“A General Classification of Three-Dimensional
Acknowledgements
Flow Fields,” Phys. Fluids A, vol. 2, pp. 765-777,
The authors would like to thank Al Globus and May 1990.
Creon Levit of NASA Ames for their inputs and
[11] R. H. Abraham and C. D. Shaw, Dynamics: The
suggestions, as well as their eigen solver. Dan Asimov
Geometry of Behavior, parts 1-4, Ariel Press,
of NASA Ames provided the data-set generator used
Santa Cruz, CA., 1984.
for generating the comparison data.
[12] H. Vollmers, H. P. Kreplin, H. U. Meier,
This work was funded by NASA Ames Research
”Separation and Vortical-Type Flow around a
Center (Tom Woodrow and Michael J. Gerald-
Prolate Spheroid. Evaluation of Relevant
Yamasaki, technical monitors) and United
Parameters,” AGARD Conference Proceedings
Technologies Research Center (Dave Edwards,
No. 342.
technical monitor).
[13] Dennis Jespersen and Creon Levit, “Numerical
References
Simulation of Flow Past a Tapered Cylinder,”
[1] P. Moin and J. Kim, “The Structure of the AIAA Paper 91-0751, 1991.
Vorticity Field in Turbulent Channel Flow. Part
1. Analysis of Instantaneous Fields and Statistical [14] S. A. Vermeersch, “Investigation of the F117A
Correlations,” J. Fluid Mech. 155, pp. 441, 1985. Vortical Flow Characteristics,” Master’s Thesis,
Massachussetts Institute of Technology,
[2] J. Kim and P. Moin, “The Structure of the Department of Aeronautics and Astronautics,
Vorticity Field in Turbulent Channel Flow. Part May 1993.
1. Study of Ensemble-Averaged Fields,” J. Fluid
Mech. 162, pp. 339, 1986. [15] G. V. Bancroft, F. J. Merrit, T. C. Plessel, P. G.
Kelaita, R. K. McCabe, and A. Globus,”FAST: A
[3] J. Villasenor and A. Vincent, “An Algorithm for Multi-processing Environment for Visualization
Space Recognition and Time Tracking of of Computational Fluid Dynamics”, Proceedings
Vorticity Tubes in Turbulence,” CVGIP: Image of Visualization ‘90, San Francisco, CA, Oct.
Understanding 55:1, pp. 27, 1992. 1990.

4
 Figure 1 Flow pattern at a critical point
whose rate-of-deformation tensor has
1 real and a pair of complex-conjugate Figure 2a A hexahedron (or structured-grid cell)
eigenvalues. divided into 6 tetrahedra.

Figure 2c A pyramid cell divided into 2 tetrahedra. Figure 2b A prism cell divided into 3 tetrahedra.

5
 center of
swirl flow

streamline

Figure 3 A sample result of an artificially-generated test case.

FLOW

Tapered
cylinder

Figure 4a Flow past a tapered cylinder. Figure 4b flow past a tapered cylinder.
Shown are swirl flow centers found by the Swirl flow centers and streamlines
algorithm. are shown.

6
 Figure 5 Flow over a F117 fighter.
Swirl flow centers and streamlines are shown. Note: The
Figure 4c Blow up of figure 4b. centers are not mirrored.

Integration veers
away from core

Figure 6a Result of FAST vortex core Figure 6b Result of pV3 swirl flow
finder on data set 1. finder on data set 1.

7
 Integration veers
away from core

Figure 6c Result of FAST vortex core Figure 6d Result of pV3 swirl flow
finder on data set 2. finder on data set 2.

Figure 6e Result of FAST vortex core finder Figure 6f Result of pV3 swirl flow
on data set 3. Note that a streamline has also been finder on data set 3.
spawned here to indicate that the twist at the top of
the middle curve is actually outside the vortex.