Topology- and Feature-based

Flow Visualization:
Methods and Applications

1 Introduction
The ongoing strive for improved efficiency, performance, and safety is at the core
of technical flow design. Typical examples of such flows can be found at all scales,
ranging from the gas-burning chamber in household heating appliances that depends
on an optimal mixture of oxygen and gas, to high-performance aircraft design. In
the recent past, computers have become powerful enough to be truly instrumental in
this field, primarily through the widespread use of Computational Fluid Dynamics
(CFD). This methodology offers new means to rapidly simulate and evaluate new
designs. Moreover, it allows for unprecedented insight into complex fluid dynamics
phenomena observed in practical experiments by generating very high resolution
data sets that accurately reproduce the entirety of the flow. This evolution em-
phasizes the need for analysis tools that are both effective and efficient. Scientific
Visualization has become essential in this context.
The dedicated research effort is called Flow Visualization. Its major task is
to provide tools that allow the user to visually explore and assess the properties of
the ever-increasing amount of numerical information that results from CFD compu-
tations. One classical approach is to focus the visualization on features of interest
that engineers and fluid dynamicists consider essential for both scientific and indus-
trial applications. Prominent examples are vortices, shock waves, and separation
or attachment lines. The corresponding visualization techniques are very useful in
practice because they yield a simplified representation of involved flow phenomena
made of patterns that directly match the intuition of the observer. Their limitation,
however, follows from the loose notion of feature that, in most cases, is essentially
application specific. In general, different methods rely on different (if not
contranictory) definitions of the same feature and therefore yield heterogeneous
results. Vortices constitute a typical illustration of that problem: many definitions
have been proposed over the years but none of them is able to properly
characterize vortices in all types of flows [8]. Following a different approach
topology-based methods ex- tract global flow structures defined with respect to the
limit sets of streamlines. The corresponding techniques are built upon the rigorous
mathematical formalism of the qualitative theory of dynamical systems, which
guarantees objective results. Unfor- tunately, the connection between topological
structures and the practical properties of the flow is sometimes unclear and the
resulting pictures tend to lack the intuitive appeal of feature-based
representations.
The objective of the paper is two-fold. First, it provides an introduction to
state-of-the-art feature and topology-based flow visualization methods. Beyond the
techniques traditionally used in practice it describes recent contributions made by
the authors, following an approach aimed at combining the strengths of both topo-
logical and feature-based techniques to yield more effective visualizations. Second,
it demonstrates the use of these algorithms for practical applications. In particu-
lar it discusses their ability to meet the needs raised by the analysis of large-scale
multi-field CFD data sets.

The contents of this paper are organized as follows. Our description starts with on
overview of feature-based flow visualization techniques in section 2. We focus our
presentation on the feature types that are most prominent in practice. Topology-
based methods are introduced in section 3. Basic notions are defined along with
algorithms relevant for visualization purposes. After these initial considerations
we adopt a more practical viewpoint and consider successively two different visu-
alization applications. One concerns vortex breakdown analysis, as discussed in
section 4. The other is dedicated to flow analysis and optimization in engine com-
ponents, section 5.

2 Feature-based Visualization
The goal of feature-based visualization methods is to generate images that restrict
the depiction of complex flow data to a limited set of points, lines, and volumes
representing features of particular interest for the considered application. This
yields fairly abstract pictures that convey significant flow properties in a concise
and compact form. The most prominent examples of features in CFD applications
include vortices, separation and attachment lines, shock waves and recirculation
zones.
The loose, empiric nature of the definition of those feature explains the vari-
ety of algorithms available to locate, identify, and visualize them and requires the
user to determine experimentally which method is best suited for the needs of his
particular application. Further restrictions on the type of method can be imposed
by the size or the structure of the data.

In this section we present visualization methods dedicated to vortices on one hand

Figure 1. Samples of feature-based visualization methods. Left: Volume

rendering of the λ2-criterion to illustrate vortices above a delta wing. Middle: Vortices as
extracted by the Sujudi-Haimes algorithm. Note the false positives and the generally bad
quality of the results. Right: Separation and attachment lines of the shear flow on the
wing surface, computed using the scheme from[19] (cf. 4 for a detailed discussion of this
dataset).

and separation and attachment lines on the other hand. This choice is motivated
by the major practical significance in practice as explained below and illustrated
in sections 4 and 5. For a more general introduction to the topic of feature-based
visualization, refer to [11].

2.1 Vortices
The extraction of vortical structures has been a major topic in visualization for quite
some time. Although a vortex is most intuitively conceived as the superposition of
a flow along an axis and a flow around this axis, there is no satisfying and objective
definition that exists for this flow pattern. As a result, vortex extraction methods
are essentially characterized by the type of vortex criteria they are built on. This
is either a region-based criterion (identifying regions of vortical flow behavior) or
a line-type description (focusing on the vortical axis or vortex core line). Region
definitions include high vorticity, helicity, low pressure. Most often used in engi-
neering is the λ2 definition by Jeong and Hussain [5]. The physical meaning behind
this method is a similarity measure of the local flow structure to that induced by a
pressure valley line. The major limitation of λ2, however, lies in its incapacity to
isolate individual structures.
Among the line-type definitions, the approach of Sujudi and Haimes [14] is
most widely used. The idea here is to perform on a cell-wise basis the pattern
matching of a rotation motion on the vector field and to extract locally sections
of the rotation axis that can be patched together to approximate the vortex core
line. Because of the linear nature of the sought pattern, the method has issues with
vortex core lines that are strongly curved. Roth and Peikert proposed a higher-
order scheme that can extract curved core lines reliably [12]. They also showed
in a subsequent paper that this and other similar methods can be formulated in a
unified framework involving their parallel operator

Figure 2. An illustration of the vortex extraction algorithm from [3]. Left:

Manually seeded stream surface showing vortical behavior. Middle: Stream surface medial
axis as a smooth approximation to the vortex core line. Right: Vortex surface grown
outward from the medial axis.

Additionally, some approaches try to identify a vortical or swirling flow be-

havior by examining the evolution of particles [6]. Recently, Garth et al. presented
a more general stream surface1-based approach [3] that allows for the extraction of
a vortex core line approximated as the medial axis of a stream surface that exhibits
vortical behavior. The stream surface can be seeded along a closed curve surround-
ing the center of a two-dimensional rotation identified on a probing plane that the
user can position arbitrarily in the flow. The starting curve can also be adjusted
to the vortex core line approximation provided by another scheme. The authors
combine this technique with a region-growing algorithm that relies on a common
vortex model from fluid dynamics to identify a vortex region. Overall this method
proves significantly more robust than alternative solution when applied to slowly
swirling vortices. An overview of the successive steps of the method is proposed in
Fig. 2.1.

While most of the methods mentioned above provide satisfactory visualizations

for simple datasets, the extraction of vortices in modern CFD datasets remains
essentially challenging and a significant amount of user interaction is required to
obtain useful results.

2.2 Separation and Attachment Lines

Separation and attachment lines are another major feature type. They are defined as
the lines along which the flow attaches or separates from the surface of an embedded
body (e.g. an aircraft). This phenomenon is induced by viscous effects that take
place in direct proximity of the object. In that setting, the flow has so-called no-slip
boundary condition which means that the velocity magnitude goes to zero as one
approaches the surface along a normal direction. Therefore, the analysis is focused
on the non-zero, tangential shear-stress vector field defined over the surface that
exhibits the same flow patterns as nearby located streamlines. In particular flow
separation and attachment induce the creation of curves of asymptotic

convergence which are visible in the shear stress vector field. The corresponding
three-dimensional flow pattern is characterized by the presence of a stream surface
starting or ending along the feature line that, on the other hand, swirls around a
nearby located vortex. As a matter of fact, flow separation and vortex genesis are
two closely related phenomena.
Following the original idea of Sujudi and Haimes for vortex core lines, Ken-
wright et al. proposed a simple and fast method for the extraction of separation and
attachment lines [7]. Their basic observation is that these feature lines are present
in two linear patterns, namely saddle points and nodes (see section 3), where they
are aligned with an eigenvector of the Jacobian. The original method works on a
cell-wise basis and extract these pattern within each triangle. Hence it results in dis-
connected line segments, caused to the discontinuity of the Jacobian. Yet, applying
the parallel operator leads to the reformulation of the features in terms of lines of
zero curvature and yields connected lines. However this definition is quite restric-
tive because it assumes that separation resp. attachment lines always have zero
curvature. Moreover, since it requires derivative computation it is very sensitive to
noise. Consequently strong pre-smoothing of the data is often necessary which in
turn can deform and shift the features. Another approach was proposed earlier by
Okada and Kao [9] who extend the classical Line Integral Convolution (LIC) algo-
rithm [1] by color coding the flow direction so as to highlight the fast changes in
flow direction that occur as streamlines approach separation resp. attachment lines.
The weakness of this approach lies in the heavy computation associated with LIC
on one hand, and in the fact that the geometry of the feature lines is not extracted.
Instead, the method computes a density function that indicates the proximity /
likelihood of these feature lines.
Using a different approach, Tricoche et al. recently proposed a scheme [19]
designed to overcome the restrictions imposed by the purely local analysis used in
the algorithms mentioned previously. Their method can be decomposed in three
stages. The first one applies a local criterion to estimate the likelihood of each grid
vertex to lie close to a separation resp. attachment line. In essence, this stage is
comparable to the method of Okada and Kao [9]. However the density computed
here is not directly visualized but serves as input for the next step of the algorithm.
Moreover the local criterion used can be chosen arbitrarily by the user. In particular,
the pattern matching idea underlying Kenwright’s method *7+ can be reformulated to
yield a distance value. The second step leverages these local estimates to monitor
the global convergence of streamlines toward separation resp. attachment lines.
This step is justified by the asymptotic streamline convergence that takes place
along separation and attachment lines in the shear stress vector field. Observe that
streamline integration is global in nature, which makes the method both more robust
and more flexible than Kenwright’s approach. Practically, the ridge and valley lines
of the density function computed previously are extracted. The resulting skeleton
is then used to restrict the candidate seed points for streamline integration to a
set of isolated lines that are discretized at a predefined resolution. Streamline
convergence is measured on a cell-wise basis by incrementing a local counter every
time a streamline crosses a cell. The final step consists in extracting the ridge and
valley lines of this convergence map. This yields an approximation of the
locationdd

Figure 3. Successive steps of the extraction method for separation and

attachment lines [19]. Top left: ridge and valley lines of density function. Top right:
streamlines seeded along ridge and valley lines. Bottom left: cell-wise accumulation mon-
itoring. Bottom right: separation and attachment lines.

of separation and attachment lines, the accuracy of which is determined by the grid
resolution. The actual geometry is eventually provided by streamline integration.
An overview of these successive steps is shown in Fig. 3.

3 Topology-based Visualization
Vector field topology is a powerful approach for the visualization of planar flows.
Topology-based methods leverage basic results of the qualitative theory of dynami-
cal systems to generate effective depictions characterized by a high level of abstrac-
tion and an accurate segmentation of the domain in regions where the flow exhibit
a uniform behavior. Formally, this classification is defined with respect to the limit
sets of the streamlines. Additionally, parametric topology and the notion of bifur-
cation can be used to extend this technique to time-dependent flows and account
for the structural transformations that their topology undergoes over time.
Unfortunately, the application of this methodology to three-dimensional prob-
lems has not so far demonstrated the same usefulness in visualization applications.
One explanation is the intricacy of the resulting pictures: the topology of volume
flows involve stream surfaces that are plagued by self-occlusion and visual clutter.

Figure 4. Linear planar critical points

Another problem concerns the lack of intuitive connection between topological struc-
tures and major features of interest in fluid dynamics problems, as described in the
previous section. Neither vortices nor separation lines are, in general, topological
in nature. Thus topology-based methods fail to extract them properly.
In the following we provide a short introduction to essential notions of planar
and three-dimensional vector field topology. Our presentation is driven by the needs
of visualization algorithms discussed in the next section. For a more complete survey
of existing methods in topology-based flow visualization, we refer the reader to [13].

3.1 Topological Skeleton of Steady Flows

The topology of a vector field is the decomposition of its phase portrait into regions
where all streamlines have the same limit sets. The phase portrait considers all
points located along the same streamline as a single equivalence class. In other
words, it interprets the domain in terms of its dense coverage by streamlines. A
limit set is defined as a set of points that constitute the asymptotic limit of a
streamline, either forward or backward. We restrict our considerations to the two
most common types of limit sets: critical points and cycles.

Limit Sets
The critical points of a steady vector field are the locations where the field magni-
tude vanishes. Because of the uniqueness of the solution of a dynamical system with
respect to its initial conditions, critical points are the only locations where stream-
lines can meet asymptotically. In the non-degenerate, linear case, the nature of a
critical point is determined by the eigenvalues of the Jacobian matrix. The different
types are illustrated in Fig. 4 and Fig. 5 for the planar and three-dimensional case,
respectively. In both cases, the eigenvalues are shown along with the associated
configuration. When all the eigenvalues have positive (resp. negative) real parts,
the critical point is a source (resp. sink). If both positive and negative real parts
are present, the critical point exhibits both source and sink behavior and is called
a saddle point.
Cycles are closed streamlines that correspond to periodic solutions of the
dynamical system. The non-degenerate case corresponds to cycles that act as sink
or sources with respect to the surrounding streamlines.

Figure 5. Linear critical points in three-dimensions

Separatrices
The limit sets present in a vector field induce a segmentation of the domain into
regions where all streamlines share the same limit sets for forward and backward
integration. The boundaries between such regions are called separatrices.
Specifically, in the planar case two major types of separatrices exist: cycles and
streamlines that attach to a saddle point along its eigenvectors. Refer to Figure 4.
As in the planar setting, separatrices of the three-dimensional topology are either
periodic manifolds or they start at saddle points along their eigenvectors. However
these separatrices are either 1D (streamlines) or 2D (stream surfaces). The latter
are spanned by both eigenvectors associated with the eigenvalues whose real parts
have same sign. Again, refer to Fig. 5.

3.2 Topology Extraction

In practice critical points are extracted on a cell-wise basis. In the piecewise linear
case, the equation to solve in each cell to determine the position of a zero vector is
linear. In the planar, bilinear case, the same problem leads to a quadratic equation.
Trilinear interpolation requires numerical schemes like Newton-Raphson to locate
critical points. If such a point is found in the interior of a given cell, its type
is determined by solving the associated Eigensystem. In the case of saddle points,
this analysis also provides the eigenvectors along which the integration of separating
streamlines or stream surfaces is carried out.
The extraction of cycles is more involved. Wischgoll and Scheuermann pro-
posed the first method for that purpose [20]. The basic idea of their algorithm is
to first identify a cell-wise cycle in which a streamlines appears to be captured.
The next step consists in verifying this property by integrating streamlines from
the boundary of the cell cycle to ensure that the vector field does indeed prevent
any streamline from leaving the corresponding region. This result is then combined
with the Poincare-Bendixson theorem which states that in the absence of any crit-
ical point in the cell cycle, the limit set present in this region must be a cycle. A
method derived from the same principle was later proposed by the same authors
for the 3D case [21]. Observe that tori that are stable under the flow are another
 Figure 6. Fold bifurcation

Figure 7. Hopf bifurcation

possible generalization of cycles in a three-dimensional setting.

3.3 Parametric Topology and Bifurcations

When a vector field depends on a parameter (e.g. time), changes in the parameter
value induce changes to its topology. These transformations are called bifurcations
and exist in an infinite variety. Their common property nonetheless is to replace a
stable structural configuration by another stable configuration through an instan-
taneous, unstable pattern. Here, stability is defined with respect to the ability of a
given structure to remain qualitatively unchanged after a small but arbitrary mod-
ification of the vector field. Bifurcations are either local or global depending on the
extent of the region they impact.
For the need of this presentation, we consider only local bifurcations. The most
common ones in the planar case fall in two categories: fold and Hopf bifurcation.
The former corresponds to the pairwise annihilation (resp. creation) of a saddle
point and a sink or source. An example is shown in Fig. 6. The latter is characterized
by the transformation of a sink into a source (and vice versa) and the simultaneous
creation (resp. annihilation) of a cycle surrounding the critical point. See Fig. 7.
Similar transformations occur in the 3D case. A simple example can be derived
from a 2D fold bifurcation by adding a one-dimensional source behavior to a saddle
point and a source affected by the transformation. This creates two 3D saddle
points that merge and vanish in the very same way. This is illustrated in Fig. 8.
 Figure 8. Three-dimensional fold bifurcation

3.4 Topology Tracking

A simple algorithmic solution to track the continuous evolution of the topology
and detect the associated bifurcations was proposed by Tricoche et al. in [16].
The method was designed for two-dimensional unsteady flows. Its basic idea is
to embeds the discrete space-time domain of the data in a three-dimensional grid.
More precisely, assuming that the grid is a fixed triangulation, each triangle is
expanded along the time axis which creates a prism that connects two consecutive
time steps. In each prism, a linear interpolant along the time axis extends to 3D
the planar, piecewise linear interpolation defined at each time step.
In this prism grid, the singularities are tracked in a cell-wise manner. Given a
critical point present in a cell at a time step the algorithm uses the equation of the
3D interpolating function to determine the path of this critical point throughout
the prism. Observe that the interpolation is chosen such as to ensure that at most
one critical point is present in a cell at any given time. Therefore only two cases are
possible for this path: it can either cross the prism from one time step to the next
(i.e. move within the same triangle cell in a 2D perspective) or cross the side faces
of the prism before reaching the next time step (that is, leave the triangle where it
was located initially between two time steps). When a critical point leaves a prism,
its path must be followed further in the corresponding neighbor.
As far as bifurcations are concerned, they can take two forms in that frame-
work. The only local bifurcation that can occur in the interior of a cell (thus
involving a single critical point) is a Hopf bifurcation. It is detected by checking
the persistence of the sink/source type of the critical point along its path. The
cycle originating at the bifurcation point can be tracked by extracting it at each
subsequent time step using the scheme of Wischgoll and Scheuermann [20]. Fold
bifurcations are constrained to take place on the side faces of a prism. In that case,
the path of a saddle point contained in a prism connects with the path of a sink
(resp. source) located in a neighboring prism. As explained previously, this config-
uration can either cause the creation or the annihilation of both critical points. The
separatrices emanating from a saddle point can be integrated at each point along the
path and reconnected along the time axis to span surfaces in the space-time domain.
The extension of this scheme to three-dimensional vector fields was described by
Garth et al. in [4]. The basic idea is the same as in the 2D case. However the space-
time domain is decomposed into 4D simplices in their implementation. This choice
greatly simplifies the equations to be solved to track the critical points and
determine their type everywhere along the time axis. The application of this method
to the temporal analysis of a vortex breakdown bubble is discussed in section 4.3.
Remark that an alternative technique for topology tracking in two and three dimen-
sions called Feature Flow Field was proposed by Theisel and Seidel [15]. However,
their method requires all time steps to be available in memory at once which makes
it non suitable to handle the large data sets considered in the next section.

3.5 Cutting Plane Topology

To finish this overview of topology-based flow visualization methods, we briefly
present a technique called Cutting Plane Topology introduced recently by Tricoche
et al. [18]. It is based on the topology tracking algorithm mentioned previously and
permits to extract and explore complex three-dimensional flow structures. Specifi-
cally, a steady three-dimensional flow is investigated through the parametric topol-
ogy of its 2D projection onto a plane that is swept along a prescribed curve across
a volume of interest. In other words, the curve controls how the cutting plane is
moved to span a 3D volume and is interpreted as the parameter space for topology
tracking. The choice of this curve is therefore application specific. We describe
in section 4.2 how the inherent symmetry of the considered flow structures can
be exploited to create effective visualizations that unravel intricate flow structures.
Another important aspect to consider here is the orientation of the plane as a con-
tinuous function of its position along the curve. Once again this choice must be
dictated by the flow to yield meaningful results. The solutions proposed in [18]
range from a fixed orientation imposed by a symmetry axis to a direction chosen
in order to maximize the amount of flow crossing the plane. Once the curve and
the plane orientation have been decided, topology tracking can be carried out in a
computational space where the successive positions of the plane and the associated
sampled vector values are aligned to satisfy the original configuration of [16]. The
extracted paths and bifurcations often require low-pass filtering in post-processing
to discard the short-term artifacts introduced by the choice of plane orientation.
Additional details can be found in [18].

4 Vortex Breakdown Analysis

4.1 Background
In both the civil and military fields, the demand for shorter flight times and faster
aircrafts has been a driving force behind research in recent years. Although it is
not a recent development, together with supersonic speeds becoming more attrac-
tive, the delta wing design has found its way back into aircraft construction, as
is demonstrated by a number of military aircrafts and the transatlantic passenger
jet Concorde. An increased perception of security and the ever-shortening take-off
and landing intervals on modern airports mandate a thorough examination of delta
wing configurations with the aim of controllable flight even in exceptional flight
situations, e.g. at a high angle of attack at subsonic speed. Furthermore, in mili-
tary airflight, exceptional maneuverability is of prime importance. Due to this, the
understanding of flow phenomena related to delta wing setups has become a major
point of research activity. Among the most interesting of these is v ortex break-
down due to its severe impact on flight stability. Delta wing designs differ from
more conventional wing designs in that a substantial part of the lift on the wing
is not only created by Bernoulli’s principle (i.e. the velocity and hence pressure
difference above and below the wing), but by large vortex systems above the wing.
Breakdown of these vortices, i.e. the sudden loss of coherent vortical flow struc-
tures above the wing, can have catastrophic consequences, due to the sudden loss
in lift and structural failure of the wing construction. While the phenomenon has
been known for quite some time, there is still a lack of good theories of its genesis.
Numerical research can be extremely helpful in allowing to study the occurrence of
vortex breakdown in computer simulations. In the following, we describe a dataset
from a simulation of a delta wing configuration that exhibits breakdown and apply
a number of visualization techniques to study it.
The dataset results from an unsteady computation of the Navier-Stokes equa-
tions in a volume surrounding a delta wing. As the simulation progresses, the angle
of attack increases. The data is given on an unstructured adaptive-resolution grid
with about 12 million cells and includes 1000 time steps. This dataset poses a
significant challenge for visualization algorithms, owing to both size and numerical
resolution issues.

4.2 Formation of Vortex Systems

In order to verify the correctness of the simulation and get an insight into the
vortex systems above the wing, the extraction of vortices above the wing surface
is an important task. The vortices above the wing actually form two symmetric
vortex systems consisting of primary, secondary and tertiary vortices each.
Figure 9 (left) provides an overview showing the basic configuration. A stream
surface started just below the wing apex illustrates the flow of air going over the
edges and rolling up into the two primary vortices. The right image provides a view
of the same surface from an opposing angle, revealing the three vortices on each side
of the wing. It is interesting to note that while the primary and secondary vortices
are almost circular in shape, the tertiary vortex is extremely elliptic. In the left
image,another pair of stream surfaces wraps around the primary vortex core lines
and exhibits spiraling behavior, as illustrated by proper color coding. About two
thirds along, the coherent motion of these surfaces is interrupted and replaced by a
bubble shaped structure. Despite the general symmetry of the dataset, the vortex
breakdown it exhibits is asymmetric. On the (in direction of flight) right side, a so-
called breakdown bubble forms, while the left-side structure looks chaotic.
The stream surfaces make for a good illustration of the major phenomena
in the dataset. Since the vortex axes move over time, the seeding curves for the
streamlines need to be determined manually for every time step, which is tedious.
 Figure 9. Left: Overview of the delta wing dataset with vortex creation at apex
and the two primary vortices, breaking down differently. Right: Formation of primary,
secondary and tertiary vortices at wing apex. Note how the shape of the tertiary vortex is
strongly elliptic. For both images, the surface color coding depends on the starting curve
location of individual streamlines (s-parameter).

Figure 10. Vortices as extracted by the vortex surface technique. False posi-
tives (indicated by arrows) are reliably discarded. The closeup (right image) shows clean
separation between the three vortices that form the vortex system. Note the strongly elliptic
shape of the tertiary vortex (red).

Automatic vortex core line extraction can help to some extent. Figure 10 (left)
shows the results of an application of the Sujudi-Haimes algorithm (dark lines).
The results are of mixed quality in that they include the vortex systems as well as
some false positives. Using stream surfaces and the vortex surface technique detailed
above, it is possible to discard the false positives, extract smooth vortex core lines
(magenta lines) and obtain good vortex regions for each of the three vortices (red,
green and blue surfaces). The right image shows a closeup. The separation between
the individual vortices is excellent, and the elliptic shape of the tertiary vortex is
extracted well.
To further understand the relation between the different vortices in each sys-
 Figure 11. Application of cutting-plane topology to a delta wing simulation.
Left: Plane travels along the wing symmetry axis, revealing the primary vortices and ex-
tracting their core lines as spiral-type critical point paths. Middle: Plane travels along the
primary vortex core line. The full vortex system is visible. Right: A single slice (from the
middle image) allows a detailed observation of the structural interaction between different
vortices (marked by green arrows) and the corresponding separation and attachment (see
also Fig. 12) on the wing surface (blue and red arrows).

tem, we have applied cutting-plane topology. The resulting structures are shown
in Figure 11. The left image gives an overview of the structures that are revealed
when the cutting plane travels along the symmetry axis of the wing. Although it is
not orthogonal to the vortical structures, a good overall picture of the flow situation
results. The primary vortex cores can be identified as the paths of critical points
as the plane travels. The middle and right images show a closeup of the vortex
system structure. More detail is extracted as the plane travels orthogonal to the
primary vortex core. The interaction of the different vortices can be inferred from
the topology. Effectively, the visualization of the vortex system is here reduced
to two-dimensions, where it is much easier to comprehend. The separation in the
wing shear flow appears as a natural part of the vortex system (individual vortices
indicated by green arrows) and is extracted as the paths of saddle critical points
close to the wing surface (red arrow in the right image). This primary separation
is essentially the boundary of the influence regions of the primary and secondary
vortices.
The separation and attachment structures can also be extracted directly, al-
though with some difficulty. Figure 12 (left) shows results obtained using the ap-
proach described in Section 2. The primary separation is clearly visible. Together
with the topological visualizations presented before, a complete picture of the com-
plex vortex dynamics above the delta wing can be obtained and compared to a
theoretical model (cf. [2]). Although it is not of direct use in the analysis of vortex
breakdown, it is of great value with respect to a validation of simulation results. In
the next section, we focus on a direct visualization of the breakdown structure.

4.3 Structure and Evolution of the Breakdown Bubble

The complicated structure of the breakdown bubble presents a serious challenge
for established visualization techniques. Since it is three-dimensional in nature,
naive geometric visualizations such as using streamlines suffer from issues of spatial
 Figure 12. Left: Separation and attachment lines in the wing shear flow as ex-
tracted with the method from [17]. Right: Illustrative sketch of the three-dimensional flow
structure above a delta wing and the resulting shear flow structure on the wing (from [2]).

Figure 13. A stream surface illustrating the flow structure of the breakdown
bubble in the delta wing dataset. While the opaque rendering (left) fails to provide insight,
application of a clipping plane (middle) or transparent rendering (right) details the internal
flow structure, essentially consisting of an asymmetric recirculation zone rotating around
the original vortex core line.

perception issues. Stream surfaces can deliver better images here by providing a
surface primitive that helps in understanding the three-dimensional structure of the
flow. Fig. 13 provides an example. While the stream surface completely wraps the
breakdown bubble, inner structure is easily revealed by applying a clipping plane.
The observed motion is of a recirculation-type, overlaid by a simultaneous rotation
around the original vortex axis.
To better identify the recirculation zone, cutting-plane topology is an ideal
tool as it allows to discard the superposed rotation. In this case, the cutting-
plane rotates on the vortex axis. The recirculation can easily be identified by a
closed, strongly curved vortex core (cf. Fig. 4.3) that appears as a spiral-type
critical point path in the cutting-plane topology. It is interesting to note that
these recirculation vortices are very hard to extract using conventional schemes
due to their strongly curved nature. The right image shows three recirculation
zones, hinting at several occurrences of vortex breakdown of the primary vor-
tex. While the cutting-plane topology is immensely useful in the analysis of the
breakdown bubble flow structure, it cannot provide an understanding of the dy-
 Figure 14. Application of cutting-plane topology to vortex breakdown analysis.
Visualization of topology is enhanced by volume-rendered isosurfaces of velocity magnitude.
Left: Right-side breakdown bubble. Structure is revealed through cutting-plane topology
(plane revolving on vortex axis). The recirculation core is extracted as a spiral-type path
(yellow). Right: Left-side staggered breakdown. Although a breakdown bubble is not dis-
cernible, several recirculation zones are extracted (yellow), hinting at multiple breakdown
bubbles.

namic of the flow in this case since it is essentially limited to a single time slice.
Topological methods can still be useful
860 in this context. It has been known that
the occurrence of a breakdown bubble
is accompanied by stagnation points in
780
the flow, i.e. critical points of the flow
vector field, that lie on the vortex axis
and essentially “delimit” the bubble.
timestep

700

While these critical points are of saddle-

type, topological visualization of the
620 separating surfaces is essentially equiv-
alent to a direct application of stream
stagnation points left
stagnation points right
surfaces (Fig. 13). It is interesting,
540
1 however, to apply critical-point track-
singularity position
1.5

ing in this context. Although the re-

Figure 15. Structural graphs of sulting visualization is essentially four-
left (blue) and right (red) side breakdown bub- dimensional (the critical points move
ble evolution. While both sides start out al- in three-dimensional space over time),
most identically, the left side deteriorates into it can be reduced to two dimensions
chaotic breakdown and shows several rapidly by observing that the movement of the
oscillating bubbles in later time steps. stagnation points is limited to the axis
of the primary vortex. The benefit of
this procedure is twofold: First, the resulting structural graph allows to infer the
structural evolution over all time steps. Second, since the dataset under consid-
eration is asymmetric, it provides a simple and effective means to compare the
evolution on both sides (Fig. 15). Comparing these results with those from Fig. 4.3
confirms that the left-side breakdown is chaotic (several small rapidly oscillating
bubbles).
5 Evaluation of Flow in Engine Components
With the general progress of state-of-the-art CFD simulations, the discipline of
engine design is made accessible to both numerical simulation and visualization of
the resulting datasets, allowing for rapid testing of engine designs. In the following
we give an example of the application of topological and feature-based visualization
methods for the analysis of prototype simulations.
Among the many design goals of combustion engines, the mixing process of
fuel and oxygen occupies an important place. If a good mixture can be achieved, the
resulting combustion is both clean and efficient, with all the fuel burned and minimal
exhaust remaining. In turn, the mixing process strongly depends on the inflow of the
fuel and air components into the combustion chamber or cylinder. If the inlet flow
generates sufficient kinetic energy during this valve cycle, the resulting turbulence
distributes fuel and air optimally in the combustion chamber. For common types of
engines, near-optimal flow patterns are actually known and include, among others,
so-called swirl and tumble motions. We show two examples of simulation datasets
showing each of these two types of flow patterns (henceforth termed “swirl-motion”
and “tumble motion”). The basic geometries of the datasets and the respective
desired motion patterns are shown in Figure 16.

Diesel Engine
This simulation is the result of a the simulation of steady charge flow in a
diesel engine, based on a stationary geometry, resulting in a simple and stable
flow. The main axis of motion is aligned with the cylinder axis and is constant
in time. The spatial resolution of the single time step is high with a total of
776,000 unstructured cells on an adaptive resolution grid.

Gas Engine
This dataset results from an unsteady simulation of the charge phase of a gas
engine. As the piston moves down, the cylinder volume increases by an order
of magnitude and the fuel-air mixture entering the cylinder is drawn into a
gradually developing tumble pattern. The overall motion is highly transient
and unstable. Both spatial and temporal resolution are relatively low, with
the data given on 32 time steps and the grid consisting of roughly 61,000
unstructured elements at the maximum crank angle.

For both datasets, the simulation results are given in the form of a vector field
defined in the interior of the respective cylinder geometries. As is quite common
in CFD simulations, the flow is required to vanish on the domain boundary (the so-
called no-slip boundary condition) in order to correctly model fluid-boundary
friction. Nevertheless, values on the boundary of the domain are easily inferred by
e.g. extrapolation of volume values next to the boundary. We remark that in classi-
cal automotive engineering analysis, visualization is rarely performed for volume or
boundary data but instead on two-dimensional slices. The main visualization goal
in these cases is is the extraction and visual analysis of the swirl- and tumble-motion
patterns.
 Figure 17. Application of cutting-plane topology and boundary topology
to the gas engine dataset. Left: Cutting-plane topology provides a good overview
of the overall motion pattern. Spiral-critical point paths (green) indicate rotation
centers. Middle: In combination with boundary topology, interactions of boundary
and volume flows are visible. Spiral critical points occur on the boundary where
vortices intersect it. Right: Rotational centers enable an enhanced interpretation of
conventional particle visualizations. Particles are color coded according to velocity
magnitude. It is visible how the rotation centers capture particles in small-scale
rotations.

Unlike the previous dataset, in these

datasets methods for the extraction of
separation and attachment lines could
not be applied since the resolution of
the boundary flow field is too low.
However, in our experiments we found
that boundary topology is successful in
extracting separation and attachment
structures in the boundary fields due to
the existence of critical points. In com-
bination of with cutting plane topol-
ogy, it is an effective approach to the
extraction of swirl or tumble patterns
and gives a good impression of the mu- Figure 16. Left: Stable, circu-
tual influences of boundary and volume lating flow pattern in a diesel engine des-
flows. Figure 17 illustrates this in a sin- ignated as swirl motion, with the cylinder
gle time-slices of the gas engine dataset. axis as the axis of rotation. The flow en-
From these images, one can recognize ters tangentially through the intake ports.
that the overall desired tumble motion Right: Transient tumble motion in a gas
actually consists of three independent engine. The axis of motion moves as
rotational centers that together form the cylinder expands downwards and stays
the tumble motion. Since one of these halfway between the top cylinder wall and
vortices is rotating opposite to the other the piston head at the bottom.
two, the tumble motion is much weaker
 Figure 18. Visualization results for the diesel engine. Interaction of vol-
ume and boundary flows can be observed where separatrices on the boundary indicate
a separation between two vortices in the volume close to the boundary. Left: A com-
bination of boundary topology with a volume rendering of λ2. The transfer function
is chosen to indicate rotation strength. Middle: Here, the transfer function is
chosen to represent the direction of the rotation (red vs. blue). Counter-rotating
vortices close to the cylinder top take away kinetic energy from the formation of
the main swirl pattern. Right: Boundary topology laid over a LIC image of the
boundary flow.

than expected, and the design must be improved. Here, the axis for the cutting-
plane approach is quite naturally parallel to the desired tumble axis, discarding
patterns of motion that are not considered important for this application.
For the diesel engine (where a swirl-type pattern is desired), topological meth-
ods can provide results of similar quality, especially in combination with other tech-
niques. Since the topological visualizations are mostly sparse in the sense that they
provide a concise line-type depiction of the flow structure, it makes sense to combine
them with feature extraction techniques that create a dense visualization (such as
the λ2-criterion) or other visualization techniques. Figure 5 gives several examples
of hybrid visualizations of this type. The combination of boundary and volume
visualizations gives a good understanding of the general nature of the flow. Again,
we find that the overall swirl pattern is a combination of several smaller vortices.
There is only one large but weak vortex extending all the way to the bottom of the
engine cylinder. The achieved pattern is therefor suboptimal.
For both the gas and diesel engines, the presented visualizations can be con-
structed without user assistance. This guarantees that visualizations are compara-
ble between different simulation datasets of the same type, an important property
when using visualization as a design analysis tool