Visual Comput (2008) 24: 1039–1051

DOI 10.1007/s00371-007-0204-x ORIGINAL ARTICLE

Tobias Salzbrunn Pathline predicates and unsteady flow

Christoph Garth
Gerik Scheuermann structures
Joerg Meyer

Published online: 1 August 2008 Abstract In most fluid dynamics decides if a pathline has a property of
© Springer-Verlag 2008 applications, unsteady flow is a nat- interest to the user. We will show that
ural phenomenon and steady models any suitable set of pathline predicates
are just simplifications of the real can be interpreted as an unsteady flow
T. Salzbrunn (u) · G. Scheuermann situation. Since computing power structure definition. The visualization
Institut für Informatik, increases, the number and complexity of the resulting unsteady flow struc-
Universität Leipzig, of unsteady flow simulations grows, ture provides a visual description of
Postfach 100920, too. Besides time-dependent features, overall flow behavior with respect to
04009 Leipzig, Germany scientists and engineers are essen- the user’s interest. Furthermore, this
{salzbrunn,scheuermann} tially looking for a description of flow structure serves as a basis for
@informatik.uni-leipzig.de the overall flow behavior, usually pathline placements tailored to the
C. Garth, with respect to the requirements of requirements of the application.
Institute for Data Analysis their application domain. We call
and Visualization, such a description a flow structure, Keywords Unsteady flow
University of California,
One Shields Avenue,
requiring a framework of definitions visualization · Feature detection ·
Davis, CA 95616, USA for an unsteady flow structure. In this Computational fluid dynamics (CFD)
cgarth@ucdavis.edu article, we present such a framework
based on pathline predicates. Using
J. Meyer,
University of California, Irvine (UCI), the common computer science
644E Engineering Tower, definition, a predicate is a Boolean
Irvine, CA 92697-2625, USA function, and a pathline predicate is
jmeyer@uci.edu a Boolean function on pathlines that

1 Introduction the kind of knowledge to be extracted from the simulation

data. In this article, we assume that the user has a list of
Unsteady three-dimensional flow is a better model of nat- flow features of interest. Usually, this list is based on expe-
ural flow phenomena than steady flow in most cases. Since rience, theory, and intention of the simulation. Therefore,
the computing power, especially of PC clusters, is high a feature-based flow visualization is part of the solution to
enough now to compute such complex simulations, un- the analysis task. It is only a part, because, in our view, the
steady flow data has become quite common and we need user will also be interested in the overall flow in relation
good postprocessing tools for these datasets. Since the to the features. As an example, an engineer constructing
data is massive, interactive access is still difficult, es- a part of a water turbine outflow is usually interested in
pecially for realistic unstructured grids with millions of vortices, among other features. Therefore, a method for
elements. A meaningful combination of automatic analy- detecting and visualizing vortices through regions with
sis and visualization is needed to provide insight to en- large vortical motion is needed. But this is only a part of
gineers and scientists. Because we want to base the visu- the story, because the engineer will ask also the question
alization on the analysis task, we need information about of how the vortex influences the overall flow in the turbine.
1040 T. Salzbrunn et al.

This article introduces the concept of unsteady flow struc- There are several dense, i.e. comprehensive, visualiza-
ture to address both local and global aspects. We divide tion techniques for unsteady flow. Forsell and Cohen [5]
the flow into two parts: the pathlines influenced by the vor- extended the original LIC algorithm from Cabral and
tex and the remaining pathlines. Looking at the particles Leedom [4]. Verma et al. [30] presented a pseudo-LIC ap-
moving through the vortex vicinity reveals how these par- proach on sparsely placed pathline ribbons. A hardware-
ticles behave in comparison to their counterparts. Through assisted texture advection technique was proposed by
this analysis we will also see some circulating behavior. Jobard et al. [9]. An image-based approach (IBFV) was
Using again our framework reveals the part of the flow introduced by van Wijk [35]. Jobard et al. [10] pre-
showing this behavior and complementing our basic un- sented a Lagrangian–Eulerian advection scheme (LEA).
derstanding of the flow. We call such a partitioning of the Weiskopf et al. [33] proposed a spacetime-coherent frame-
pathlines, as in the previous two examples, an unsteady work for texture-based visualization (UFAC). Shen and
flow structure. Kao [22] presented a new LIC algorithm (UFLIC), which
In general, an unsteady flow structure is a partitioning was recently accelerated and extended to 3D by Liu and
of all pathlines that reflects the user’s analysis criteria with Moorhead [15]. Many of the dense visualization tech-
respect to the particle movement. Of course, this means niques work only for 2D unsteady flow, i.e. flow on sur-
that totally different flow structures can be obtained for faces. Even if there exists an extension for 3D unsteady
the same flow, but this is a direct consequence of varying flow, the usability is reduced by the occlusion problem in-
intentions specified by different users analyzing the same herent to dense visualization techniques. Park et al. [16]
simulation. addressed this problem by using multi-dimensional trans-
This article defines a framework for defining and com- fer functions. Bauer et al. [2] used special regions of
puting such flow structures. The framework is based on interest to selectively visualize 3D time-dependent vector
the notion of pathline predicates. Such a predicate assigns fields. This article shares the same spirit, and our approach
a Boolean value to each pathline depending on whether is described in the next section.
the respective pathline has a certain property or behavior. For 2D time-dependent vector fields, there exists
In this way, each predicate splits the set of all pathlines a number of contributions to extend the topological con-
into two classes. Combining different predicates based on cepts from steady flow. Early works come from Tric-
Boolean algebra allows a finer partitioning of the flow that oche et al. [28, 29] and more recent work from Theisel
goes beyond a simple partitioning into two parts. This pro- et al. [27]. Recently, Shi et al. [23] presented an informa-
vides the basis for definitions of a wide variety of flow tion visualization approach for 3D unsteady vector fields.
structures of different complexity. But, even for very com- Nevertheless, a convincing concept of structure in un-
plex flow structures, the user is provided with an exact steady flow is still missing. This article is an attempt to fill
meaning for each part of this flow structure, because of this gap.
the exact definitions of the underlying predicates. How- Another way to obtain useful information out of time-
ever, despite this objectivity, it is up to the user to choose dependent data is feature-based visualization. One of the
meaningful predicates in order to obtain a meaningful flow most challenging tasks in this context is to track the de-
structure that is worth its name. sired features over time. There are some general works
Flow structures provide a structural view of the be- on feature tracking (van Walsum et al. [31], Samtaney
havior of interest. Even though this is valuable infor- et al. [21], Reinders et al. [19], and Silver and Wang [24])
mation in its own right, users still might want to track and several works on special features and their track-
the traces of individual particles. Therefore, we will also ing, e.g. singularities (Garth et al. [6]), closed stream-
show how to compute a sparse pathline placement that lines (Wischgoll et al. [36]), vortices (Bauer and Peik-
takes the underlying flow structure into account and pro- ert [1]), or general line type features, isosurfaces, or vol-
vides the user with a dynamical view of the flow. We umes (Theisel and Seidel [26], Weigle and Banks [32],
will see that both views inspire and complement each and Ji et al. [8]).
other. The authors have recently published a work on stream-
line predicates [20] that can be considered as the steady
counterpart to the work presented here. In the following
2 Related work section, it will become clear that there are substantial dif-
ferences. The main handicap to a direct application of the
Pathlines and streaklines [14] are often used to show same concepts is the fact that unsteady flow is usually
particle traces in time-dependent flow fields. In add- given only for a finite time span. This prevents many path-
ition, Becker et al. [3] extended flow volumes to unsteady lines from being complete in the sense of connecting an
flows as a generalization of the concept of streaklines. inlet to an outlet of the domain via a curved line. It is also
Time surfaces as an extension of time lines can be han- apparent that different predicates are needed and that the
dled by a level-set approach as proposed by Westermann visualization makes substantial use of animation in the un-
et al. [34]. steady case.
 Pathline predicates and unsteady flow structures 1041

3 Unsteady flow features For any position a ∈ D and any time τ ∈ I, we define
a pathline pa,τ by
Feature-based visualization aims at a higher level of ab-
straction by extracting ‘phenomena, structures, or objects pa,τ : Ia,τ → D,
in a dataset, that are of interest for a certain research t  → pa,τ (t),
or engineering problem’ [18]. Typical examples in un- pa,τ (τ) = a,
steady flows are shock waves, boundary layers, recircula-
∂ pa,τ
tion zones, attachment lines, separation lines, separation (t) = v( pa,τ (t), t),
bubbles, vortices, shear flow regions, and vortex bursts. ∂t
In each application, in each dataset, and for each re- where v denotes the velocity field. Here Ia,τ ⊂ I is the
searcher, a different feature definition may be appropriate. maximal lifespan of the particle in D during I. Since we
Therefore, the first important task is to find an appropri- are interested in different particles, we call two pathlines
ate definition of the feature that permits the development pa,τ , pb,σ equivalent if they describe the life of the same
of a corresponding detection algorithm. Additionally, for particle, i.e. we have the equivalence relation ∼:
time-dependent features, one has to ensure a consistent
tracking over time. In a last step, the extracted feature may pa,τ ∼ pb,σ ⇔ pa,τ (σ) = b
be simplified and described quantitatively in order to be
(⇔ pb,σ (τ) = a, due to uniqueness theorem).
easily visualized.
Features can be seen as a characterization of a set of Hence, we can express the set of all different particles as
points in the spatial domain. With respect to unsteady flow equivalence classes of pathlines with respect to the previ-
features, the temporal domain will also be taken into ac- ous equivalence relation ∼ as
count. Based on information from a local neighborhood,
a subregion, or the whole field, each point has an attribute Pv := { pa,τ | pa,τ is pathline of v }/ ∼ .
assigned stating whether the specific feature exists or not
at this point. Hence, a feature definition can be formulated Since equivalence classes are not intuitive, we divide the
as a point predicate, i.e. a function that maps all points of set Pv into four distinct classes. If we assume that no par-
the domain to a Boolean value: ticle is created or destroyed inside D, a particle
(a) can be present in the interior D̊ at t0 and leave D at
Π : D × I → {TRUE, FALSE} a time τ ∈ (t0 , tn ],
(b) can be present in the interior D̊ at t0 and be present in
with the characteristic set
the interior D̊ at tn ,
(c) enters D at a time τ ∈ [t0 , tn ] and leaves D at a time
CΠ = {(x, t) ∈ D × I | Π(x, t) = TRUE},
τ ∈ [t0 , tn ) with τ > τ, or
(d) enters D at a time τ ∈ [t0 , tn ] and can be present in the
where D ∈ R3 denotes the spatial domain and I ⊂ R de- interior D̊ at tn (see Fig. 1).
notes the temporal domain.
Formally, we have

Pv = Pvb ∪ Pvp ∪ Pvc ∪ Pve

4 Pathline predicates
(b = begin, p = permanent, c = complete, and e = enter)
We are concerned with a continuous unsteady velocity with
field
Pvb = { pa,τ | pa,τ (t0 ) ∈ D̊, Ia,τ = [t0 , τ ]
v : D × I → R 3
and pa,τ (τ ) ∈ ∂D},
Pvp = { pa,τ | pa,τ (t0 ) ∈ D̊, Ia,τ = [t0 , tn ]
on a bounded domain D ⊂ R3 over a time span I =
[t0 , tn ] and the property that there exists a positive constant and pa,τ (tn ) ∈ D̊},
K > 0, K ∈ R, such that Pvc= { pa,τ | a ∈ ∂D, τ ∈ [t0 , tn ], Ia,τ = [τ, τ ]
and (τ < tn or (τ = tn and pa,τ (τ ) ∈ ∂D))},
∀x, y ∈ D and ∀r, s ∈ I,
v(x, r) − v(y, s) ≤ K(x, r) − (y, t) Pv = { pa,τ | a ∈ ∂D, τ ∈ [t0 , tn ], Ia,τ = [τ, tn ]
e

and pa,τ (tn ) ∈ D̊},

(Lipschitz property). We note that this property is needed
for the existence and uniqueness of the following path- where D̊ denotes the interior of D and ∂D the boundary
lines. of D. We note that we know the full path through D of
1042 T. Salzbrunn et al.

Based on pathline predicates, we develop a formal

definition for a flow structure of an unsteady flow. Since
the flow structure depends on the user’s interest in the flow
behavior, different flow structures of the same flow are
possible and useful. The chosen flow structure determines
the outcome of the visualization and the features that are
visible in the rendered flow. We start with a finite set G of
pathline predicates with disjunct characteristic sets, i.e.

G = {Pλ | λ ∈ Γ }, C Pλ ∩ C Pµ = ∅ ∀λ, µ ∈ Γ, λ  = µ

for a finite set Γ . Furthermore, we demand

C Pλ = Pv .
λ∈Γ

Fig. 1. Different types of particles of Pv . The classification depends This partitioning represents the unsteady flow structure
on the location of a particle in the domain D during the time inter- and serves as a formal definition. We note that, in practice,
val I = [t0 , tn ] (note that for illustrative purposes the domain D is many pathline predicates are defined only on the com-
one-dimensional) plete pathlines Pvc . In this case, G has to include the three
p
extra predicates P p = ( p ∈ Pv ) (the pathlines staying per-
the particle only if it belongs to Pvc . Additionally, we note manently in the domain), P e = ( p ∈ Pve ) (the pathlines
that we do not know the path of a particle outside the do- entering but not leaving the domain), and P b = ( p ∈ Pvb )
main D. Hence, a reentering of the domain by a particle (the pathlines present in the interior at the beginning and
would be regarded as an entrance of a new particle. leaving the domain).
A pathline predicate is a partial map
P : Pv → {TRUE, FALSE}, 6 Computation of pathlines
p  → P( p).
As a preprocessing step for the construction of unsteady
We allow a partial map here, because many predicates in flow structures, we essentially transform the velocity field
practice assume a full path of the particle through the do- from an Eulerian description (given as the resulting dataset
main, i.e. they can only be defined for p ∈ Pvc . In the of a simulation) to a Lagrangian description. That means
special case, where there is no inflow and no outflow, Pv that we want to analyze changes which occur as one fol-
p
consists only of Pv and all pathlines are known within lows a fluid particle along its trajectory, instead of ex-
the time interval [t0 , tn ]. On this basis the particles can be amining changes as they occur at a fixed point in the
compared. The corresponding characteristic set C P is de- velocity field. This step needs to be done only once for
fined by a given dataset. The large computational effort of this pre-
processing step has to be related to the effort of the simu-
C P = {(x, t) ∈ D × I | P( px,t ) = TRUE}. lation itself. Even if this preprocessing step takes some
hours/days on commodity hardware, simulating 3D un-
This is the set of all points in spacetime visited by particles
steady flow takes days/weeks on supercomputers.
fulfilling the predicate. For a given time ti we define the
In theory, there exists an infinite number of pathlines.
restriction of the characteristic set C P to ti as
Depending on the predicate, close pathlines may have dif-
C P |ti = {(x, t) ∈ C P | t = ti }. ferent predicate values. Obviously, we must limit the com-
putational burden to obtain results in finite time. We do
not know a priori which predicates will be used in gen-
eral. This excludes every adaptive optimization approach
5 Unsteady flow structure based on a specific predicate. Hence, we decided to dis-
cretize the volume of particles such that two particles have
We are convinced that scientists and engineers try to ob- a maximal distance below a value of 2δ ∈ R at each time
tain a mental model of the overall flow behavior. A sub- step t0 , . . . , tn . Of course, other approaches are possible.
stantial part of this model is the movement of all particles, For this purpose, we voxelize the spatial domain D of the
which we call flow structure. Furthermore, the mental simulation data with voxels of edge length δ. (For mean-
model connects this movement to detected features of the ingful results, we can restrict the predicates to those that
unsteady flow. have characteristic sets with a diameter greater than δ in
 Pathline predicates and unsteady flow structures 1043

Table 1. Memory needed to store pathline, raster, and attribute data In both cases, one can use parallel machines to speed up
for different datasets the calculation.
Dataset Pathline data Rasterization Attributes

Cuboid 18 GB 1.2 GB 5.5 MB

Turbine 42 GB 27 GB 45 MB
8 Pathline predicate examples
There are two groups of pathline predicates that differ in
the amount of additional information needed to evaluate
most parts of the volume during simulation time.) With the predicate. Predicates from the first group need only
respect to the origin of the particles, there are two pos- the information about the pathline itself. This has the ad-
sibilities. They are either present inside D at time t0 or vantage that we do not have to compute some derived
they enter D through an inflow part of ∂D during the time quantities of the unsteady flow field. Examples for this
interval I . group are predicates that compute geometric characteris-
The flow volume present at t0 is represented by par- tics of the pathline like curvature or deviation from a given
ticles starting at the center of the voxels at time t0 . All the direction and demand a certain threshold in order to be
particles are propagated forwards in time over the com- evaluated as true. Another example uses the time infor-
plete time interval I using the vector data of the original mation of the pathline and computes residence time in the
grid. (Because of main memory limitations, it could be in dataset. As predicate, one can take a certain quantile of the
some cases faster to propagate all particles from one time resulting distribution of the values as a threshold.
step to the next.) In this process, it becomes clear that the The second group comprises predicates that rely on
flow does not cover the volume in a dense enough manner, data derived from the unsteady flow, e.g. derived fields
so we add particles to the centers of empty voxels at each like vorticity, helicity, and pressure, or features of the flow
time step. Once we have reached time step tn , we have the like shock-wave boundaries or vortex cores. These pre-
complete pathlines of all the particles present at t0 . But, dicates are usually computationally more expensive and
we are missing the history of the particles that we intro- need more storage, since the additional derived data has to
duced on the fly to obtain a dense covering. Therefore, we be stored. In the following sections we give an example for
iterate backwards through the time steps and compute the each of the two groups that we use in the remainder of the
missing parts of the pathlines of these introduced particles. article.
At the end of this process, we store each pathline pa,τ ,
its position at every time step (i.e. its rasterization), its
lifespan Ia,τ , and its type according to the classification of 8.1 Pathlines showing circulating behavior
Pv introduced in Sect. 4. By far the largest amount of stor-
age is needed for the pathlines and depends on the reso- In many engineering applications, such as fluid dynam-
lution of their discretization as polylines. But, the path- ics or circulatory systems in medicine, there is a great
lines do not have to be cached in the main memory. For the interest in recirculation zones. Often they lead to large
computation of the predicates, one could go through all the stagnant flow zones that hinder the overall flow, resulting,
pathlines one by one in a linear fashion. The storage de- for example, in inefficient turbines and pumps. In other
mands for our examples used in Sect. 10 can be found in applications recirculation zones are desired, e.g. to obtain
Table 1. a cleaner combustion in engines. Recirculation with re-
spect to a recirculation plane can be computed by adding
up the winding angle of the projection of the pathlines to
this plane. Values above 2π indicate at least one circle.
7 Computation of pathline predicates Computation load can be lowered with additional informa-
tion about the extent of the recirculation area. Based on
While the pathlines have to be computed only once per this information, control regions can be defined that in-
dataset, each pathline predicate needs to be evaluated on dicate circulating behavior if visited multiple times. The
these pathlines. As the examples in the next section show, resulting pathline predicate would state whether a pathline
many predicates require information about the complete shows recirculating behavior or not.
path of the particle through the volume. In this case, we
limit the calculation to the set Pvc . Each pathline is eval- 8.2 Pathlines accompanying a specific vortex
uated using a function for the predicate. Sometimes it is
faster to do this in parallel for all particles, especially if the Scientists and engineers are interested in the interplay be-
original vector field is needed. If this is not the case, e.g. tween vortices in the flow and flow outside the vortex re-
if the computation is based on the integration of curvature gion. Therefore, we examine pathlines with respect to the
along the pathline and on deciding if it belongs to the top behavior towards a vortex. We want to analyze the part of
n% (n = 10, for example), a sequential approach is faster. the flow that is influenced by a certain vortex. A common
1044 T. Salzbrunn et al.

imagination of ‘influenced’ is a particle which is close To obtain a view that exposes the dynamical behavior
enough and swirls around the vortex core. For steady flow of the flow, we need a pathline placement producing on the
a predicate that examines this behavior is given in [20]. one hand only a small number of particles for a given time
Extending this approach to unsteady flow requires an ex- step in order to avoid clutter and on the other hand enough
act tracking of vortex cores from one time step to the next. pathlines to be still representative of the underlying flow
This is a difficult task and very time consuming. There- structure. Looking upon a respective discretized charac-
fore, we concentrate on the distance between a particle teristic set as a four-dimensional region, a representation
and a vortex. We assume that a particle that accompa- that captures its essential topology can be obtained in the
nies a (moving) vortex core within a small neighborhood form of a skeleton representation (medial axis). Comput-
distance for some time is influenced by this vortex. Typ- ing this skeleton for such a hypervolume is still a very
ically, the distance lies within the vortex volume, that of challenging task and a subject of ongoing research (e.g.
course changes its diameter along the vortex core and over see the works of Jonker [11, 12]). But, we can take a sim-
time. To compute the vortex volume directly is a nontrivial pler approach that fits our needs. For every time ti we see
and error-prone task leading to complicated geometries. only the restricted characteristic set C P |ti . A representa-
Hence, the required inside-test would be computationally tive pathline placement should assure that for this time
expensive. To reduce the computational cost, we make span the respective particles should represent the structure
a simplifying assumption and use a volume with a con- of C P |ti . Therefore, we only compute for every time step
stant diameter. Of course, other loops could also be used. the skeleton of the respective restricted characteristic set
But, the circle allows a very fast inside-test (see Sect. 10). using a thinning approach from Kuba and Palagyi [13].
For a given pathline, we track the distance to the vor- Every voxel of the resulting skeleton represents all the
tex core and sum up the times the pathlines reside within voxels of the restricted discrete characteristic set that are
the required distance to the vortex during the lifetime of no closer to any of the other skeleton voxels. To com-
the particle. Doing this for all pathlines (or a representa- pute this set of voxels for every skeleton voxel (numbered
tive subset) results in a temporal distribution. Hence, we consecutively through all time steps) we use a flood-fill
can use a certain quantile as a minimum time threshold to algorithm starting for the first run with the skeleton vox-
define the predicate. If a pathline has a longer residence els themselves and continue by assigning the next-nearest
time than the given threshold, the predicate is evaluated neighbors (i.e. max. 26 voxels) to the respective skeleton
as true. voxels for the next runs. In the case of conflicting assign-
ments concerning two skeleton voxels we have to calcu-
late the actual distance and assign the voxel to the skeleton
voxel with the shortest distance. If the distance is equal
9 Pathline placement we take the first skeleton voxel. After several runs we ob-
tain a partition of the restricted discrete characteristic set
Flow structures show a partitioning of the flow. They pro- according to the assignment to a skeleton voxel.
vide those parts of the flow exhibiting a specified behavior. We define that a pathline ‘visits’ a skeleton voxel at
In this structural view, individual particles usually cannot time ti if the respective particle is situated in the skeleton
be seen. However, tracing the behavior of individual par- voxel itself or one of its assigned voxels at that time. A set
ticles within the context of a given flow structure could of pathlines visiting all skeleton voxels represents the re-
lead to new ideas for refining this flow structure. Hence, spective characteristic set. We use an heuristic approach
both the structural view and the dynamical view of indi- that minimizes the set of pathlines. Starting with the first
vidual particles cross-fertilize each other. skeleton voxel we look at all visiting pathlines and choose

Fig. 2. Snapshots of a particle placement for a given time step (from left to right): restricted characteristic set, double eroded restricted
characteristic set, resulting skeleton, particle placement from initial resolution, and particle placement from quarter resolution
 Pathline predicates and unsteady flow structures 1045

from this set the one visiting the most unvisited skeleton left transparent. At time step t0 there is no inflow. The
voxels of later time steps. We repeat this for the next un- particles from C P p |t0 , that will stay in the domain for
visited skeleton voxel until no unvisited skeleton voxel is the entire simulation, are especially located around the
left. The resulting set of pathlines is used for the place- cuboid. The complementary part of the flow comprises
ment. We applied two ways to further lower the number of particles from C P b |t0 . The next images for time steps t1 ,
pathlines. First, we eroded the discrete characteristic set to t2 , t3 , t4 , and t5 show the advancing inflow. Especially,
obtain the main dominating structure, using a simple mor- the flow around the cuboid mainly comprises particles
phological erosion operator. This results in fewer skeleton from C ST . Flow not hindered by the cuboid moves quickly
voxels and thus fewer pathlines. Second, we discretized to the outflow boundaries and hence comprises particles
the eroded voxelized characteristic set with a lower reso- from C ST . At time step t6 nearly the entire domain is
lution yielding again fewer skeleton voxels and pathlines filled with inflow and the structure of C ST |t6 can be seen:
(see Fig. 2). a meandering pattern resulting from the vortex street be-
hind the cuboid. In the next image the inflow of C P e |t7
can be seen at time step t7 . At time step t8 nearly all
particles from C ST and C ST left the domain. Particles
10 Results
of P pc are still mainly centered around the cuboid indi-
Our first dataset results from a direct numerical simula- cating circular behavior of the flow around the cuboid.
tion of fluid flow around a cuboid at a Reynolds number The series of images shows that even with this very sim-
of Re = 1000. The simulation was carried out with the ple pathline predicate and with the distinction of different
NaSt3DGP1 flow solver. A version of the NaSt3DGP code classes of Pvc , the user can obtain important insights into
as well as related information and documentation is avail- the flow. Based on these results further analysis can be
able for download at http://wissrech.iam.uni-bonn.de/ started.
research/projects/NaSt3DGP/index.htm. The underlying Next, we apply our framework to a dataset result-
regular grid contains 250 000 cells. The timespan is I = ing from a simulation of the flow in the upper part
[0.0, 100.0]. For the pathline starting points, we voxelize of a draft tube of a Francis turbine. The flow en-
the domain with a resolution of [99] × [99] × [24] × [120], ters the draft tube with a strong residual swirl. This
resulting in 2 588 600 pathlines. The first example shows helps to guide the flow around the 90 degree bend in
the importance and usefulness of discriminating different the draft tube. The bounding box of the draft tube is
particle classes of Pv . We analyze how long particles stay D = [−0.55, 0.55] × [−0.45, 0.25] × [−1.2, −0.07]. The
in the flow. Therefore, we compute the staying time for time span is I = [0.0, 9.6]. The underlying unstruc-
every pathline. From the resulting distribution of staying tured grid contains 5 million tetrahedra. For the path-
times we take a certain threshold tmin (for our example line starting points, we voxelize the domain with a high
we take the 97% quantile). We define the corresponding resolution of [82] × [62] × [75] × [1200], resulting in
pathline predicate as 5 838 785 pathlines.
For our first analysis of this dataset, we examine the
ST = ( p ∈ Pvc ) flow with respect to its vortices. Therefore, we extract the
dominating vortices and build up a flow structure based on
∧ ( p stays more than or equal
the pathline predicates from Sect. 8.
to tmin in the domain), First, we compute the vortex core lines for each time
ST = ( p ∈ Pvc ) step. The left-hand image in Fig. 3 shows the resulting
∧ ( p stays less than tmin in the domain). segments for one time step after applying the method
of Sujudi–Haimes [25] formulated with the parallel vec-
For our flow structure we use tor operator of Peikert and Roth [17]. One dominating
vortex is clearly visible but also much ‘noise’ (false
GStay = {ST, ST , P b , P e , P p }, positives or small vortices) and artifacts at the bound-
aries. Hence, we filter out all lines under a certain length
where P b , P e , and P p describe the pathlines that are (arc length = 0.7). The result of the filtering is shown
incomplete. Figure 4 shows snapshots of the restricted in Fig. 3 (center).
characteristic sets C ST |ti (colored in blue), C ST |ti (col- To compute the pathline predicate from Sect. 8.2, we
ored in turquoise), and C P p |ti (colored in magenta) at need an effective method for the necessary distance cal-
times t1 = 0.0, t2 = 3.3, t3 = 11.6, t4 = 19.7, t5 = 35.8, culations. For each time step, we compute a distance
t6 = 45.8, t7 = 77.5, and t8 = 95.0. C P b |ti and C P e |ti are field for the dominant vortex on the positions of the
dataset grid. This reduces minimum distance calculations
1 NaSt3DGP was developed by the research group in the Division of Sci-
to a simple interpolation in the distance field at an in-
entific Computing and Numerical Simulation at the University of Bonn. It terrogated position in space and time. Figure 3 (right)
is essentially based on the code described in a book by Griebel et al. [7]. shows the isosurface of the distance field concerning the
1046 T. Salzbrunn et al.

The predicates P b , P e , and P p describe the pathlines that

are incomplete. We use the finite set of pathline predicates

GVortex = {S, S , P b , P e , P p }

to build up the flow structure. Figure 5 shows snap-

shots of the restricted characteristic sets C S |ti (colored in
blue) and C S |ti (colored in turquoise) at times t1 = 0.024,
t2 = 0.104, t3 = 0.312, t4 = 0.456, t5 = 0.488, t6 = 0.608,
t7 = 6.92, and t8 = 9.512. C P e |ti and C P b |ti are left trans-
parent. At time steps t1 and t2 , an early stage of the simu-
Fig. 3. Vortex core lines of Sujudi–Haimes for a single time step lation can be seen. The main inflow is at the top of the
before (left) and after filtering (center). After the filtering one cen- draft tube and a minor inflow at the bottom. The flow
tral vortex remains. The right-hand picture shows an isosurface of
the corresponding distance field (for the distance value used for close to the walls streams much faster than the remain-
pathline predicate S) ing flow. The basic periodic movement of the vortex can
be seen at t3 , t4 , and t5 . The vortex alternates from the
left- to the right-hand side of the draft tube. The vortex
main vortex for an isovalue of 0.04 (which we use as loses particles and strength, especially when hitting the
maximal neighborhood distance for our computations) right-hand side. The particles outside the vortex travel to-
in a total range from 0.0 to 1.0. For each pathline, we wards the outlet. At t6 and t7 , one can see a small opposite
calculate the residence time within the neighboring dis- vortex fed with particles from the main vortex. The vor-
tance to the vortex. From the resulting distribution of tex moves particles upwards, contrary to the main vortex.
residence times, we take the value of the 50% quan- Once the particles hit the main vortex again they change
tile as minimum residence time tmin = 0.056 and de- their direction and flow again downwards. This indicates
fine the following pathline predicate S and its opposite the existence of a circulation flow embedded in the over-
S as all flow. Finally, at t8 , we see a larger set C P e |t8 that will
stay in the domain until the end. Additional to the remain-
S = ( p ∈ Pvc ) ing downwards flow of C S |t8 and C S |t8 there is a small
∧ ( p stays more than or equal part of C S |t8 residing at the top of the draft tube, indicat-
to tmin in the vortex neighborhood), ing a circle in the flow. Figure 6 shows individual particles
resulting from our pathline placement method. A subset
S = ( p ∈ Pvc ) of these particles is shown together with faded traces of
∧ ( p stays less than tmin in the vortex neighborhood). the last three time steps. The traces’ length gives a hint of

Fig. 4. Structural view of GStay : boundaries of restricted characteristic sets C S |ti (colored in blue), C ST |ti (colored in turquoise), and
C P p |ti (colored in magenta) at times t1 = 0.0, t2 = 3.3, t3 = 11.6, t4 = 19.7, t5 = 35.8, t6 = 45.8, t7 = 77.5, and t8 = 95.0. C P e |ti
and C P b |ti are left transparent
 Pathline predicates and unsteady flow structures 1047

Fig. 5. Structural view of GVortex : boundaries of restricted characteristic sets C S |ti (colored in blue) and C S |ti (colored in turquoise) at
times t1 = 0.024, t2 = 0.104, t3 = 0.312, t4 = 0.456, t5 = 0.488, t6 = 0.608, t7 = 6.92, and t8 = 9.512. C P e |ti and C P b |ti are left
transparent

Fig. 6. Dynamic view of GVortex : individual particles at three time steps t1 = 0.416, t2 = 0.544, and t3 = 0.624 are depicted in blue. Cor-
responding restricted characteristic sets C S |t1 , C S |t2 , and C S |t3 are colored in light gray. A subset of particles has an additional faded
trace of the last three time steps. This series of snapshots shows the loss of particles of the vortex colliding with the right-hand side of the
draft tube

the particle velocity. The particles of the subset result from whole motion of particles and the animation of the struc-
a pathline placement based on a coarser voxelization of the tural view we refer the reader to the videos accompanying
characteristic set (see Sect. 9). The series of particle snap- this article.
shots shows the vortex hitting the right-hand side of the Next, we investigate the part of the flow showing the
draft tube and at the same time losing particles. To see the circulating behavior indicated earlier. We compute the cir-
1048 T. Salzbrunn et al.

Fig. 7. Structural view of GRecirc : the upper series of pictures shows the restricted characteristic sets C R1 |ti (colored in dark green) and
C R |ti (colored in light green) at time steps t1 = 0.08, t2 = 0.28, t3 = 0.592, and t4 = 1.192. C P e |ti and C P b |ti are left transparent. The
1
lower series of pictures illustrates the flow at time step t5 = 6.04 using the same color scheme. To obtain the main region of C R1 |t5 , an
erosion operator is applied to C R1 |t5 two times (last but one picture). The last picture shows the eroded region of C R1 |t5 together with
C S |t5 (colored in light gray)

culating behavior as explained in Sect. 8 using an x y plane sides the recirculation area. To obtain the recirculation
at the center of the draft tube as a control plane. We define area, we therefore apply an erosion operation on C R1 |t5
the following predicate C1 and its opposite C1 : as shown in the second and third pictures of this series.
The fourth picture shows in light gray C S |6.04 together
R1 = ( p ∈ Pvc ) ∧ with the recirculation area, illustrating that this area is
( p makes at least one circle around the center), generated by the main vortex. Figure 8 shows a series of
R1 = ( p ∈ Pvc ) ∧ snapshots of individual particles showing this recircula-
tory behavior.
( p makes no circle around the center)

and the predicates P b , P e , and P p again describe the path- Table 2. Computation times for the turbine dataset
lines that are incomplete. For our flow structure we use
Task Computation time
GRecirc = {R1 , R1 , P b , P e , P p }.
Integration + rasterization + storage 89 h
The upper series of pictures in Fig. 7 shows the restricted Predicate vortex 2.5 h
characteristic sets C R1 |ti (colored in dark green) and Predicate circulation 1h
C R1 |ti (colored in light green) at time steps t1 = 0.08, Particle placement + animation 1.2 h
t2 = 0.28, t3 = 0.592, and t4 = 1.192. C P e |ti and C P b |ti
are left transparent. During this time period the recircu-
lation area is built up on the left-hand side of the draft Table 3. Computation times for the cuboid dataset
tube. The lower series of pictures illustrates the flow at
time t5 = 6.04 using the same color scheme. It seems Task Computation time
that the recirculation area now covers nearly the entire
Integration + rasterization + storage 35 h
draft tube. But, C R1 |t5 contains all particles leaving this
Predicate staying times 0.3 h
area downstream or getting into this area from above be-
 Pathline predicates and unsteady flow structures 1049

Fig. 8. Dynamic view of GRecirc : individual particles at three time steps t1 = 1.736, t2 = 1.808, and t3 = 2.016 are depicted in green. For
some particles an additional faded trace of the last three time steps is shown. The corresponding restricted characteristic sets C S |t1 , C S |t2 ,
and C S |t3 of the vortex pathline predicate (shown in light gray) illustrate the connection between vortex movement and recirculation of
the flow

The discussed examples proved pathline predicates us to formalize and compute a partitioning of the dynam-
and unsteady flow structures to be a useful tool for the ics, i.e. the selection of particles based on the interests
analysis of 3D unsteady flow. Working with large 3D un- of a user. Different structures can be defined on the same
steady vector fields by means of commodity PC hardware dataset, reflecting varying user interests. Based on the flow
is still one of the most challenging tasks because of stor- structure, we proposed a pathline placement allowing the
age demands and computational costs. Tables 2 and 3 user to track individual particles showing the behavior
show computation times on a PC (AMD Opteron 224, specified in the pathline predicates. Combining the struc-
8 GB RAM, single core) for the different tasks leading to tural view of a flow structure with the dynamical view
a flow structure. By far the most time consuming task is of tracking individual particles yields a proper illustration
the initial preprocessing step needed once for each dataset. of the flow the user is interested in. For our future work,
Additionally, the computation of the predicates and the fi- we intend to study more datasets and other predicates al-
nal animations of the structures take some time. Therefore, lowing different flow structures. Additionally, we want to
future work will exploit acceleration potentials especially improve computation times by exploiting all paralleliza-
by means of parallelization. Emerging multi-core CPUs, tion potentials.
the use of GPUs for integration purposes, and powerful PC
clusters should dramatically cut down computation times.

Acknowledgement The authors wish to thank Ronald Peikert from

11 Conclusion ETH Zürich for providing the datasets. Furthermore, thanks go to
all members of the FAnToM development team for their program-
This article has introduced the notion of pathline predi- ming efforts. This work was partly supported by DFG grant SCHE
cates and unsteady flow structure. These concepts allow 663/3-7.

References
1. Bauer, D., Peikert, R.: Vortex tracking in pp. 329–335. IEEE Computer Society unstructured 3D time-dependent datasets.
scale space. In: Data Visualization 2002, Press, Atlanta, GA (1995) In: IEEE Visualization 2004, pp. 329–336.
Proceedings of VisSym 2002, pp. 233–240. 4. Cabral, B., Leedom, L.C.: Imaging vector IEEE Computer Society Press, Austin, TX
Eurographics Association, Aire-la-Ville fields using line integral convolution. (2004)
(2002) Comput. Graph. 27(4), 263–270 (1993) 7. Griebel, M., Dornseifer, T.,
2. Bauer, D., Peikert, R., Sato, M., Sick, M.: 5. Forsell, L., Cohen, S.: Using line integral Neunhoeffer, T.: Numerical Simulation
A case study in selective visualization of convolution for flow visualization: in Fluid Dynamics, a Practical
unsteady 3D flow. In: IEEE Visualization curvilinear grids, variable speed animation, Introduction. SIAM, Philadelphia,
2002, pp. 525–528. IEEE Computer and unsteady flows. IEEE Trans. Vis. PA (1998)
Society Press, Boston, MA (2002) Comput. Graph. 1(2), 133–141 (1995) 8. Ji, G., Shen, H., Wenger, R.: Volume
3. Becker, B., Lane, D., Max, N.: Unsteady 6. Garth, C., Tricoche, X., Scheuermann, G.: tracking using higher dimensional
flow volumes. In: IEEE Visualization 1995, Tracking of vector field singularities in isosurfacing. In: IEEE Visualization 2003,
1050 T. Salzbrunn et al.

pp. 209–216. IEEE Computer Society 18. Post, F., Vrolijk, B., Hauser, H., 28. Tricoche, X., Scheuermann, G., Hagen, H.:
Press, Seattle, WA (2003) Laramee, R., Doleisch, H.: The state of the Topology-based visualization of
9. Jobard, B., Erlebacher, G., Hussaini, M.: art in flow visualization: feature extraction time-dependent 2D vector fields. In: Data
Hardware-assisted texture advection for and tracking. Comput. Graph. Forum 4(22), Visualization 2001, Proceedings of VisSym
unsteady flow visalization. In: IEEE 775–792 (2003) 2001, pp. 117–126. Eurographics
Visualization 2000, pp. 155–162. IEEE 19. Reinders, F., Post, F.H., Spoelder, H.: Association, Ascona (2001)
Computer Society Press, Salt Lake City, Attribute-based feature tracking. In: Data 29. Tricoche, X., Wischgoll, T.,
UT (2000) Visualization 1999, Proceedings of VisSym Scheuermann, G., Hagen, H.: Topological
10. Jobard, B., Erlebacher, G., Hussaini, M.: 1999, pp. 63–72. Eurographics Association, tracking for the visualization of time
Lagrangian–Eulerian advection of noise Vienna (1999) dependent two-dimensional flows. Comput.
and dye textures for unsteady flow 20. Salzbrunn, T., Scheuermann, G.: Streamline Graph. 26(2), 249–257 (2002)
visualization. IEEE Trans. Vis. Comput. predicates. IEEE Trans. Vis. Comput. 30. Verma, V., Kao, D., Pang, A.: PLIC:
Graph. 8(3), 211–222 (2002) Graph. 12(6), 1601–1612 (2006) Bridging the gap between streamlines and
11. Jonker, P.: Morphological operations on 3D 21. Samtaney, R., Silver D., Zabusky, N., LIC. In: IEEE Visualization 1999,
and 4D images: from shape primitive Cao, J.: Visualizing features and tracking pp. 341–348. IEEE Computer Society
detection to skeletonization. In: Discrete their evolution. IEEE Comput. 27(7), Press, San Francisco, CA (1999)
Geometry for Computer Imagery: 9th 20–27 (1994) 31. van Walsum, T., Post, F.H., Silver, D.,
International Conference, DGCI 2000, 22. Shen, H., Kao, D.: A new line integral Post, F.J.: Feature extraction and iconic
pp. 371–391. Springer, Berlin, Heidelberg, convolution algorithm for visualizing visualization. IEEE Trans. Vis. Comput.
Uppsala (2000) time-varying flow fields. IEEE Trans. Vis. Graph. 2(2), 111–119 (1996)
12. Jonker, P.: Skeletons in n dimensions using Comput. Graph. 4(2), 98–108 (1998) 32. Weigle, C., Banks, D.: Extracting
shape primitives. Pattern Recogn. Lett. 23. Shi, K., Theisel, H., Hauser, H., iso-valued features in 4-dimensional scalar
23(6), 677–686 (2002) Weinkauf, T., Matkovic, K., Hege, H., fields. In: Proceedings of the Symposium
13. Kuba, A., Palagyi, K.: A 3D 6-subiteration Seidel, H.: Path line attributes – an on Volume Visualization, pp. 103–110.
thinning algorithm for extracting medial information visualization approach to ACM, Research Triangle Park, NC (1998)
lines. Pattern Recogn. Lett. 19(7), 613–627 analyzing the dynamic behavior of 3D 33. Weiskopf, D., Erlebacher, G., Ertl, T.:
(1998) time-dependent flow fields. To appear in A texture-based framework for
14. Lane, D.: Scientific visualization of TopoInVis 2007 spacetime-coherent visualization of
large-scale unsteady fluid flows. In: 24. Silver, D., Wang, X.: Tracking and time-dependent vector fields. In: IEEE
Scientific Visualization, pp. 125–145. IEEE visualizing turbulent 3d features. IEEE Visualization 2003, pp. 107–114. IEEE
Computer Society, Los Alamitos, CA Trans. Vis. Comput. Graph. 3(2), 129–141 Computer Society Press, Seattle, WA
(1997) (1997) (2003)
15. Liu, Z., Moorhead II, R.: Accelerated 25. Sujudi, D., Haimes, R.: Identification of 34. Westermann, R., Johnson, C., Ertl, T.:
unsteady flow line integral convolution. swirling flow in 3D vector fields. Tech. Topology-preserving smoothing of vector
IEEE Trans. Vis. Comput. Graph. 11(2), Rep. AIAA Paper 95-1715, American fields. IEEE Trans. Vis. Comput. Graph.
113–125 (2005) Institute of Aeronautics and Astronautics 7(3), 222–229 (2001)
16. Park, S., Budge, B., Linsen, L., (1995) 35. van Wijk, J.: Image based flow
Hamann, B., Joy, K.: Dense geometric flow 26. Theisel, H., Seidel, H.: Feature flow fields. visualization. In: ACM SIGGRAPH 2002,
visualization. In: Brodlie, K., Duke, D., In: Data Visualization 2003, Proceedings of pp. 745–754. ACM, San Antonio, TX
Joy, K. (eds.) Proceedings of VisSym 2003, pp. 141–148. Eurographics (2002)
Eurographics/IEEE-VGTC Symposium on Association, Grenoble (2003) 36. Wischgoll, T., Scheuermann, G.,
Visualization 2005 (EuroVis 2005), 27. Theisel, H., Weinkauf, T., Hege, H., Hagen, H.: Tracking closed streamlines in
pp. 21–28. Eurographics Association, Seidel, H.: Topological methods for 2D time-dependent planar flows. In: Ertl, T.,
Aire-la-Ville (2005) time-dependent vector fields based on Girod, B., Niemann, H., Seidel, H.P. (eds.)
17. Peikert, R., Roth, M.: The parallel vectors streamlines and pathlines. IEEE Trans. Vis. Vision, Modeling, and Visualization 2001,
operator – a vector field visualization Comput. Graph. 11(4), 383–394 (2005) pp. 447–454. Aka, Berlin (2001)
primitive. In: IEEE Visualization 1999,
pp. 263–270. IEEE Computer Society
Press, San Francisco, CA (1999)
 Pathline predicates and unsteady flow structures 1051

T OBIAS S ALZBRUNN received an M.S. degree G ERIK S CHEUERMANN received his B.S. and trical Engineering & Computer Science and the
in computer science in 2003 from the Univer- M.S. degrees in mathematics in 1995 from the Department of Biomedical Engineering in the
sity of Kaiserslautern, Germany. Afterwards he University of Kaiserslautern, Germany. In 1999, Henry Samueli School of Engineering at the
worked as a research assistant at the same uni- he received a Ph.D. degree in computer science, University of California, Irvine. He joined UC
versity. Currently, he is working as a research as- also from the University of Kaiserslautern. From Irvine in 2002. Dr. Meyer is also affiliated with
sistant at the Department of Computer Science at 1995 to 1997, he conducted research at Ari- the Center of GRAVITY (Graphics, Visualiza-
the University of Leipzig, Germany. His research zona State University for about one year. He tion, and Imaging Technology) in the California
interests include dynamical systems and scientific worked as a postdoctoral researcher at the Center Institute for Telecommunications and Informa-
visualization, especially flow visualization. for Image Processing and Integrated Computing tion Technology (Calit2). He received his Ph.D.
(CIPIC) at the University of California, Davis, degree from the University of Kaiserslautern,
C HRISTOPH G ARTH received his M.S. degree in 1999 and 2000. Between 2001 and 2004, he Germany, in 1999. He held an appointment as
in mathematics and computer science in 2003 was an assistant professor of computer science a post-doctoral researcher and lecturer in the
and his Ph.D. in 2007 from the University of at the University of Kaiserslautern. Currently, Computer Science Department at the Univer-
Kaiserslautern. He is currently a postdoctoral he is a full professor in the Computer Science sity of California, Davis, from 1999 to 2000,
researcher with the Institute for Data Analysis Department of the University of Leipzig, Ger- and maintains an adjunct assistant professorship
and Visualization at the University of California, many. His research interests include algebraic at the Computer Science and Engineering De-
Davis and a collegiate member of the DFG IRTG geometry, topology, Clifford algebra, image pro- partment at Mississippi State University, where
1181 “Visualization of Large and Unstructured cessing, graphics, and scientific visualization. He he was also affiliated with an NSF-sponsored
Datasets”. His research interests include the vi- is a member of the ACM, IEEE Computer Soci- Engineering Research Center (2000–2002). His
sualization of large and unstructured datasets, ety, and GI. research interests include scientific visualization,
the application of topological methods in visu- computer graphics, biomedical imaging, and vir-
alization, and specifically the visualization and J OERG M EYER is an assistant professor with tual reality.
analysis of numerically simulated flows. a shared appointment in the Department of Elec-