Next-Generation

Data Visualization Tools

by

William Ribarsky and James Foley

GIT -GVU-94-27
May 1994

Graphics, Visualization & Usability

Center
Georgia Institute of Technology
30332-0280
Atlanta GA

Next-generation data
visualization tools
7

Bill Ribarsky and Jim Foley

.

7.1. INTRODUCTION
Scientific data visualization has finally come of age as an important and
accepted discipline. While scientists have been using computer graphics to
visualize experimental data and computational results for at least 30 years,
recent improvements in cost/performance of graphics workstations, more
readily available software, and the new-found name "scientific data visual
ization" have solidified the discipline. Many thousands of scientists regu
larly use visualizations as part of their work. Our thesis is that scientists are
forced to work too hard to create these visualizations, but that the evolving
set of visualization tools can greatly reduce the requisite effort. Further,
these tools, and those of the next generation, will open significant new ways
to understand increasingly complex and large datasets. Principal causes for
this will be the increased interactivity that the tools will naturally offer and
the need for designs based on careful consideration of scientific and
engineering analysis requirements.
One of the keys to success of visualization tools is ease of use. The tools'
users - scientists whose interest is finding meaning in data - have neither
time nor patience for graphics subroutine packages, graphics jargon, and
unnecessary detail. Ease of use can be provided in several ways. First, tools
which allow visualizations to be created as easily as with contemporary 2D
presentation graphics packages are important, and indeed, several exist.
However, the predefined presentations in these packages are fixed and are
not readily extensible. The scientist should be able to create new visual
izations by interactively binding data values to visualizations, such as three
data values to position, one data value to isosurfaces, another to isocurves
drawn on the isosurfaces and three more data values to vectors anchored on
the isosurface. Such capabilities exist for 2D data presentation: they need to
be generalized to 3D.
Going one step further, automatic generation of such visualizations
based on characteristics of the data, information needs of the user, effec
tiveness of various techniques (color, isosurfaces, vectors), effectiveness of
various combinations of techniques, and characteristics of the display
devices is an appropriate goal for the next generation of visualization
DATA VISUALIZATION
ISBN

Copyright 7 994 Academic Press Ltd

All rights of reproduction in any form reserved

104

Interface Technology and Perception

systems. Again, some of this has been done for 2D data presentations: the
work needs to be extended. Expert system technology can be helpful in
making the complex decisions involved here, but human factors and per
ceptual psychology knowledge are needed if good decisions are to be made.

7.2. WHAT IS VISUALIZATION?

A useful definition of visualization might be the binding (or mapping) of
data to a representation that can be perceived. The types of binding could be
visual, auditory, tactile, etc., or a combination of these. As Senay and
Ignatius (1990) have pointed out, the visual bindings could be further
subdivided into sets of marks (e.g. points, lines, areas, volumes) that
express position or pattern and retinal properties (e.g. color, shape, size,
orientation, texture) that enhance the marks and may also carry additional
information. The auditory bindings could be used for iconization or to map
data onto elements such as pitch or spatial localization. The tactile bind
ings could be through force feedback devices, controls with multiple degrees

of freedom, and so on.

We can use our expanded definition of visualization to guide us in
building very general and extensible libraries of visualization primitives. If
in addition we design these primitives with attachable elements and then
allow the user to bind data variables interactively to these elements, we
have a truly powerful system for visualization and analysis. Such a system
will make possible the easy building of mixed visual representations of data
(e.g. containing both point glyphs and volume visualizations) that users can
customize to best represent their data rather than being arbitrarily re
stricted to predetermined representations. In addition such systems will
prcve particularly useful in representing several variables at once and in
such a manner that their interdependencies can be made clear. Finally, as
we shall show below, the interactive binding coupled with related forms of
interactivity will significantly enhance one's ability to analyse complex
data.
Next-generation visualization tools need:
a general data model;
effective and efficient graphical representations;
the means to bind from data to graphics.

However, our next-generation data require that we also integrate into our
visualization tools:
the means to convert (or filter) data to appropriate mappings;
a general analysis model.

The next-generation data will scale to breathtaking size and complexity.

Already molecular dynamics (MD) simulations on the CM-5 model as
many as 250 million interacting particles, each changing position over
time. On the horizon are simulations of a billion particles or more. These
systems are approaching macroscopic dimensions and will provoke

Next-generation data visualization tools

detailed comparisons with continuum models and with experiments. In

underwater seismic oil exploration, the data easily number 1 Tbyte. The
spatial structures and other variables embedded in these data must be
studied in detail and compared with simulations. Finally, in the next 5-10
years, data flowing from the Earth Orbiting System of satellites will total
more than 1 Tbyte per day on atmospheric chemistry, air and surface
temperatures, winds, cloud patterns, ocean currents and other variables. It
will be fruitless to pre-analyse data of this size fully; that will occur on theffly
and most likely as a result of what the user has seen in a previousx

visualization.
Furthermore, comparison of data from different sources in the same
visualizations will become commonplace, as Treinish (1991) and others
have pointed out. In these circumstances tools for filtering or converting
data will be necessary and must be tightly integrated with the visualization
process. We call this integrated approach visual analysis. In order to make
our visual analysis effective, it is necessary to build an analysis model and
integrate it with our visualization model. The analysis model should be
founded on a set of principles that are applicable in general to scientific and
engineering data. We will discuss the building of such a foundation below.

7.3. A VISUALIZATION ENVIRONMENT

What we have described implies the visualization environment of Figure
7.1. If we have a set of data objects and separately have a set of graphical
objects, we can delay decisions about how they are interactively joined
together (in the binder) until those decisions need to be made - as the data
are being viewed and explored. This is the most flexible way to create data
visualizations. (We shall see below that it can become too flexible, with the
result that the user may need expert assistance in choosing the most
effective representations.) Furthermore, if we attach a 3D graphical editor
to define and group graphical objects (such as glyphs) for subsequent
binding, we can put all these modules in a feedback loop that processes
through the user's brain to refine the visualization iteratively or to focus on
certain aspects of the data.
2D, 3D Graphics

Editor

1
Data

Data
Management

Data Model
Figure 7. 1 . A general visualization environment.

Binding

Visualization

105

106

Interface Technology and Perception

It is important that the environment be based on a well-thought-out data

model that allows both "expressive" and "effective" data visualizations (in
the sense of Mackinlay, 1986). In addition, if the data model is appropri
ately constructed, the highly flexible binding interface can allow one to
distinguish data visually at the level of the individual element (if desired)
and can also allow cross-linking of variables to bring out correlations at any
level of detail. To permit easy input and use of data from any source, the
model should be simple but general and "self-describing", as in Treinish
and Gough (1987). Next-generation data will come more and more from
several sources and the metadata contained in the data model will be
necessary to construct meaningful comparisons. Given the ever-expanding
size of next-generation datasets, the data model must be integrated into a
carefully developed and very efficient data management structure. We will
describe our initial attempts at some of these aspects of a data model in our
description of the Giyphmaker.

7.4. GLYPHS AND THEIR USES

Glyphs, which are graphical objects whose elements (e.g. position, size,
shape, color, orientation, etc.) are bound to data, have proven useful in
depicting spatially complex, multivariate data such as are found in current
calculations and experiments. These objects can clearly be effective in
depicting discrete data such as macromolecular structure and interactions,
but they have also proved their use in representing variables such as -wind
speed and direction in atmospheric dynamics simulations and obser
vations ( Treinish, 1993). Glyphs owe their effectiveness to the ability of the
eye-brain system to discern finely resolved spatial relationships and differ
ences in shape. They thus allow the user to display and correlate several
variables at once; a researcher can, for example, choose to distinguish two
variables by binding them to the visually orthogonal representations of
shape distortion and pseudocoloring. We have recently shown (Ribarskyet
al. 1993a) that complex, three-dimensional structure from molecularjdyG
namics simulations and all components of the stress tensor acting on
individual atoms can be distinguished and correlated by using shape dis
tortion and spot coloring of spherical glyphs bound to the atomic positions
(see Figure 7.2.). It is important to reveal this interdependence because the
atomic stresses force the plastic deformations that result in significant
structural changes.
In quite a different context, Ellson and Cox (1988) have shown that
glyphs are quite useful in depicting the flow patterns that result from
finite-element simulations of hot plastic injected into a mold. In this work
100-200 glyphs were placed according to a finite-element grid; each glyph
was a 3D object with the velocity in a finite element represented as a shape
distortion of appropriate magnitude and direction, and temperature and
pressure represented by different color scales on the glyph and its base,
respectively. Other important work employing glyphs concerns the use of
color or gray scale "icons" that merge color and texture perception to

Next-generation data visualization tools

'-

'

*V;v.. * .

-AV*- .*

*.

Figure 7.2. Atomic Force microscope with Ni tip being raised from Au surface.
On-axis color spots and shape distortion of the spheres are used to represent
stretching/compressive forces and shear forces, respectively.
represent multiple parameters (Lefkowitz, 1991; Pickett and Grinstein,
1 988). Here the icon is a generalization of the single pixel to objects having

perceivable features and attributes. Typical icons are stick figures or other
simple collections of lines with such characteristics as line orientation or
length representing different parameters. Color coding is also used to bind
additional parameters to the icons. The method has been used, for
example, on satellite imaging or MRI data (Smith et al., 1991) where
thousands of small icons are used to build up collective perceptions of
texture and contour.
Some final examples in the current use of glyphs come from work by
Haber ( 1990) on the use of tensor glyphs in visualizing data from engineer
ing mechanics. All these and many other examples show- the range and
power of glyph representations. They are often superior to traditional
surface and volume techniques in, for example, displaying and elucidating
3D vector or tensor fields. It is apparent from this selection of applications,
that glyphs can be effectively used to depict a variety of data, including
continuum data, in addition to the obvious data (e.g. molecular modeling
applications).
Taken together, all this research explored a wide range of glyph represen
tations starting from a few hundred highly detailed glyphs (Ellson and Cox,
1988) and extending to several thousand glyphs where it was most import
ant to perceive "gestalt" or collective effects rather than individual glyph
characteristics (Lefkowitz, 1991). Our previous work explored the middle
ground where we showed that one could effectively perceive and use (for
analysis! individual glyph elements even when there were thousands of
glyphs (Ribarsky et al., 1993a). Such an example, where several variables
(tensor components) are depicted at once, is shown in Figure 2. With
Glyphmaker we hope to make it possible for the user to represent, visualize
and analyse data effectively in all these categories.

107

108

Interface Technology and Perception

7.5. AN EXAMPLE: GLYPHMAKER

We have recently developed Glyphmaker, a tool for data visualization/anal
ysis (Ribarsky et al., 1993b,c) that allows users easily to build customized
representations of multivariate data and provides interactive tools to ex
plore the patterns in and relations between the data. Glyphmaker is in
tegrated into the SGI Data Explorer to take advantage of the features and
flexibility of dataflow visualization systems. The main parts of Glyph
maker are the Data Reader, the 3D Glyph Editor and the Glyph Binder. It
provides non-expert users with a quick and easy way to set up visualizations
of complex data, to refine the -visualization, add to the set of variables
depicted and focus on regions of interest. We provide here a brief overview
of Glyphmaker; see Ribarsky et al., 1993b,c) for a more complete descrip
tion with a fuller discussion of design principles and applications.

7.5.1

Foundations of Giyphmaker

Our objective is to provide a highly interactive and efficient means for users
to create visual representations, change them, bind data to the represen
tations, change the bindings, bring in new data or glyphs with associated
bindings, and quickly see the results of all these operations and updates in a
visual display. The historical motivation for our approach comes from work
of Foley and others. In a system developed for dynamic process visual
ization, Foley and McMath (1986) showed how non-programmers could
construct and customize dynamically updated scenes displaying process
information. This information came from real-world environments such
as factories, power plants and refineries and was displayed (and the display
modified) to provide precise monitoring of these complex processes. The
idea was to make available a library of icons ( 2D glyphs) or allow the user to
create them; the user then employed binding tables to connect variables to
icon elements. The user was thus able to create a customized, graphical
representation describing a set of time-varying processes variables and to do
so without intensive programming. Later Foley (1990) proposed extending
this idea to more general visualization schemes, such as with scientific
data. Here the data might be 2D or 3D, and the user might need to analyse
several variables at once.
The recently developed dataflow approaches provide a general, compat
ible framework for Glyphmaker. Although dataflow and similar schemes
based on a visual programming style had been investigated for a period of
years, the first widely-used dataflow system was AVS (Upson et al., 1989).
Now there are several additional systems built around the same basic
dataflow concept including the SGI Iris Explorer , IBM Explorer, Khoros,
and apE . They all allow users to build their own applications by construct
ing graphical networks of modules - such as for data input, filtering,
transformation to geometry, pseudocoloring and rendering.

Next-generation data visualization tools

7.5.2. Glyphmaker architecture and interfaces

The Glyphmaker system has 3 main interactive parts:
the Read Module, which converts input data from a user-specified format
the Explorer internal format so that it can be read-in by any existing
Explorer module;
the Glyph Editor, which creates glyphs and arranges them, if desired, into
compound glyphs;
the Glyph Binder, which allows interactive bindings or binding changes
between data variables and glyph elements. It also allows one quickly
to alter the range of data variables or to alter the range of change in the
glyph elements as the data varies from minimum to maximum. Finally,
the Binder permits the saving of new bindings or reloading of previous
to

ones.
We will describe these parts below. In addition, some Explorer modules
are important enough that we integrate them into our system; two main
modules here are the Generate Colormap module and the Renderer mod
ule. Figure 7.33 shows how the parts of Glyphmaker fit into the Explorer
dataflow environment. Note that the structure in Figure 7.3 fits into the
general structure shown in Figure 7.1. Data management is handled
through the Data Transcriber and the Header/Label Transcriber, and the
data model is organized around the Explorer data structures and the
Header/Label file.

The Read Module

To deal with the need to retrieve ever-growing amounts of data from a
variety of formats and place them in a common structure for easy compara
tive viewing or for analysis with a common set of tools, we have developed a
simple "self-describing" data structure that is consistent with other work
on this problem (Treinish and Goettsche, 1991). Each self-describing file
has a header, which describes the data and their format, followed by the
data values themselves. In the header each variable is labeled with a name,
primitive type, number of instances, minimum value and maximum value.
The label, minimum and maximum values are important information that
is passed along to the Glyph Binder. The number of records following in the
file with identical format is indicated at the top of the header; this allows
concatenation of an animation sequence or different views of the data. Each
header can also contain comments about what the file contains. Thedata in
the file then follow in the order given in the header. Headers (and the data
following) can be placed anywhere in the file. They can also be placed
together at the top with all the data following the last header; this allows
one to merge files and still keep header information accessible.
We chose this very simple format because it was easy to understand, easy
to create, and it fulfilled the basic needs of the Glyphmaker. In the future we
will modify the Read Module to handle more complicated data formats

109

110

Interface Technology and Perception

Raw Data

Raw Data

T
Data Transcriber

Header File

Explorer Data Structure

(lattice, pyramid)

Header File Transcriber

Conditional Box

Variable Data Structure

Binder

Geometry

Glyph Data Structure

Editor

Generator)

Explorer Geometry

Renderer

Figure 7.3. Glyphmaker visualization environment. The figure shows the flow
of data between the parts of Glyphmaker as described in the text. Raw data flow into
the Data Transcriber (in the Read module), are converted to Explorer data struc
tures, and then used to generate geometry. The Header file (which can be part of the
data file) is used by the Binder to make data variable headings that are joined with
glyphs from the Editor.

than described here, but we will try to do so while retaining the basic
simplicity and intuitive feel of our present format.

The Glyph Editor

The Glyph Editor is a simple 3D editor used to draw the glyphs and,
if desired, to arrange and orient them in compound glyphs. There are
four selection menus, seen in the upper right hand comer of Figure 7.4:
Glyphs, Views, Mode, and Action. The Glyph menu items are: lines,
spheres, cuboids, cylinders and arrows. The Views menu helps the
user manipulate the glyphs in 3D space by providing menu items for
selecting either perspective, top, side or front views singly or all four views

Next-generation data visualization tools

"i* "iY
I-

*' !S" is*'

sr

Figure 7.4. 3D Glyph Editor interface. Selection menus are in the upper right
comer; a compound glyph composed of a sphere and a vector glyph is displayed.

simultaneously. The Action menu has bind, render, ungroup and clear
items for (respectively) bringing up the Binder, rendering a set of glyphs
(with full lighting and shading - useful for seeing how they will look in the
finished visualization), ungrouping grouped glyphs, and clearing all glyphs
from the Editor. We will soon include options to save and load useful
glyphs, since creating good glyphs takes care and testing.
Each glyph has several elements that can be bound to data. For example,
one might bind the radius, position (x, y,z), color, overall size and transpar
ency of a sphere to individual data variables. Shape, such as elongation or
shortening along each axis, might be bound to additional variables. Similar
sets of elements are available for all glyphs in the Glyphmaker library. One
must choose which elements to make active for the next phase; only these
elements appear in the Glyph Binder menu.

The Glyph Binder

For binding, the user chooses the bind item from the Action menu in the
Editor. A window pops up that shows all the active elements and the
variables as toggle buttons (see Figure 7.5). To bind a variable to an
element, the user simply selects an active element followed by a variable or
a variable followed by an active element. As shown in Figure 7.5, the user
can also get a pop-up window with binding information by clicking an
already-lit toggle button. Bindings can also be deleted and reset with the
mouse. Attached to each variable and active element are text widgets

111

112

Interface Technology and Perception

specifying the minimum and the maximum value for that variable or active
site. Default values appear initially in these widgets the default for the
variable is taken from the self-describing file header. The use of defaults
means that bindings can be established in a few clicks, and the results
displayed immediately. Of course, the user has the ability to change any of
these values, and the result will be transmitted to the Tenderer for update of
the display.
We have tried to create the Binder with the interactivity and ease of
control necessary for effective data exploration. The variables and elements
are in separate scrollable lists (Figure 7.5) so that one can quickly find the
variables/elements of interest, even if the lists are longer than the window
size. The Binder window is also resizable and movable using the mouse so
that one can easily rearrange the display real estate - this is important since
the Binder will often be used concurrently with the Tenderer for interactive
refinement of data visualizations. The Render button at the bottom of the
window pipes any changes made in the Binder to the Tenderer for display.
The OK button dismisses the Binder, but with data/glyph bindings remain
ing in place for the interactive visualization. At the top left of the window,
the File button has a pull-down menu with the usual save/reload options.
Saving and loading of bindings is important since it may take some time to
find an ideal set of bindings and associated minimum/maximum values for
a particular set of data In addition to the local presentation of binding
information obtained by clicking on a particular toggle button, the user
needs a global view - especially when several glyphs are involved. To handle
the global view, we are near completion and will soon implement a binding
visualization panel in the Binder window that will illustrate the overall
binding structure.

>-

FUt Edit Properties Variable*.

Active Seraents

Meetsum

Minimum

Minimum

Variables

30j

Maximum

Sphered:col or

0.000000

j 99.000

NePosjfl

0.000000

100X001

Sphered xadlus

|0.000000

i 1.0000

NePos_yO

>0.000000

1 100X0

Sphered :cen x

I -1.000000

1 1.0000

NePos.zO

I 0.000000

I 100.00 |

SphsreO:ce.t y

i-i.OOOOCO

; 1.0000

NeRadiusO

1 0X100000

|100.00

NeKElocO

10.000000

I 1020X

1.0000 ;

Sp| connection
MePos.zO

J Spheral:color
jSphere! radius

jSpherelrcenx

jO.QOOOOC

1 99.000

jNeKEavgO

|XOOOOO

! 100.00

1 0.000000

110000

jNeTempO

1 0.000000

1300.00

uxxxxx

jNaPos.xl

j 0.000000

i ioo.ooa

-1.000000

jphereOtctn z bound to NePosjti

OK

Apply

j Cancel| Rcndarj

Figure 7.5. The Glyph Binder Interface. On the left are the glyph active elements
and on the right are the variable names (in this example including position coordi
nates and properties for each atom type).

Next-generation data visualization tools

Conditional Box
We have implemented a specific and particularly useful instance of a
concept from statistical graphics called focusing and linking (Buja et al.,
1991). (As discussed further below. ) This is a 3D box that can be moved and

sized to enclose a subregion of the data after it has been selected for display
and bound to particular glyphs. The box allows one to distinguish what is
inside from what is outside in effect to reclassify the data according to this
condition. Thus the data, which may have been originally classified accord
ing to atom type can be additionally classified according to this spatial
condition. One can then choose new glyph bindings - for example, new
colors or glyph shapes for what is inside the box - to focus on or even isolate
an interesting spatial region.
Our current implementation of the Conditional Box is shown schemat
ically m Figure 7.3. The box module receives information from the Explorer
data structure and from the Geometry Generator to generate a box with the
appropriate scale and position with respect to the data. Then data and box
geometry are piped to the Tenderer for simultaneous display. Interactive
controls can update the box geometry and position it with respect to the
data. In the future we will pipe the new classifications determined by the
Conditional Box (e.g. in or out) directly to the Binder, allowing the user to
employ the full range of glyph elements. At the moment we only allow a
limited set of glyph rebindings based on the Conditional Box.

A Giyphmaker example
We present an example using Giyphmaker on data from a molecular
dynamics (MD) simulation investigating the dynamic behavior of atomis
tic materials as they move and interact with one another. External and
internal forces may cause the build-up of large, localized and transient
stresses and heating resulting in such dynamic events as plastic flow,
localized melting and resolidification, and even rupture or splattering of
material. It is necessary to analyse the complex interrelation of stresses,
heating and dynamic behavior in order to understand the processes that
occur in these simulations. The simulation presented here was run on a
Cray Y-MP at the DOE National Energy Research Center by Dr Hai-Peng
and Professor Uzi Landman of the School of Physics, Georgia Instutute of
Technology (Hai-Peng and Landman, 1993).
The simulation depicts a "nanocluster" of 64 atoms of NaCl hitting an
Ne surface and becoming embedded in the surface. Although the system
was initially quite cold (50 K) and the impact velocity moderate (3km/s),
upon impact the cluster and surface in its vicinity heats up greatly with
much disorganization and eventually some vaporization. We have chosen
simple sphere glyphs for binding to the separate components of the system
(Na and Cl separately in the cluster, dynamic Ne atoms in several layers
stacked on several layers of crystalline substrate). Properties of this system
include localized atomic kinetic energy and temperature and also the
atomic stress tensor - in the visualizations presented here we have bound
the temperature to the color of the glyph for each atomic type. We have
applied the Conditional Box to the system at a time step just before the

113

114

Interface Technology and Perception

cluster penetrates the surface (Figure 7.6). This is to avoid the common
problem with large scale simulations; often important features are embed
ded in the system and hidden from view, as shown in Figure 7.7 where the
cluster has fully penetrated the Ne (with binding of color to local tempera
ture turned on). Using the interactive rotation and zooming capabilities in
the Tenderer and the Conditional Box controls (shown at the lower right in
Figure 7.6), we adjust the semi-transparent box so that it encloses the
cluster and the region of the Ne surface that the cluster is about to strike.
Now we can choose to view only this selected region, the excluded region, or
both (perhaps using the transparency control to make the excluded region
semi-transparent). We can also retain the Conditional Box classification
shown in Figure 7.6 for later time steps; Figure 7.8 shows the selected
region after the cluster embeds itself (with local temperature binding
turned off and on, respectively'!. In this case, we can see (by comparing with
previous time steps) that Ne atoms have diffused out from the center of the
impact but the edges of the selected region have not yet moved out much
Finally, we show in Figure 7.9 the fuil system with the cluster removed
(achieved by just clicking off the NaCl bindings and updating the display).
One can now explore the shape and development of the hole using the
interactive rotation and zoom controls in the Tenderer. Also, though it was
not apparent when the cluster was visible, one can see clearly that the
greatest heating of the Ne surface remains in a concentrated region below
the cluster impact. The color bar in Figure 7.9 shows the temperature
binding used throughout. A temperature range of 0 to 300 K is mapped onto
the absolute scale of 0 to 100 on the color bar. We can thus see that even
after total embedding of the cluster, the outer edges of the surrounding Ne
remain quite cool. However, the cluster itself and Ne atoms in the center
are between 200 and 300 K- much greater than the vaporization tempera
ture of Ne.
In Figures 7.10 and 7.11, we show how more properties can be built into
the visualization by depicting the effect of localized, on-diagonal atomic
stresses using a vector glyph. The same time step as in Figure 7.7 is shown,
but here we have removed the binding to the Ne layers and enabled the
binding to the crystalline substrate atoms below so that we can see clearly
how the latter behave. Since the stress tensor is symmetric, it has only six
components that can be divided intuitively into on- and off-diagonal sets.
The on- diagonal set represents stretching or compression along the com
ponent axes depending on whether the sign of the component is positive or
negative, respectively; the off-diagonal set represents shear forces. We have
depicted the on-diagonal forces both with a grouped glyph consisting of a
sphere and vector (not shown here, the spheres locate atoms precisely) and
with vectors alone. One can switch between the two representations with a
few clicks in the Binder. The direction and magnitude of the vector in
dicates the amount and direction of on-diagonal force. (The magnitude of
the vector is also proportional to the pressure at that point) . Let us label the
horizontal and vertical directions in Figure 7.11 as x and y with z perpendic
ular to the image. Figure 7.10 shows that columns of atoms (in the z
direction) directly below the cluster (and below rows along x and y that
intersect these atoms) have vectors mostly pointing downward indicating
compressive forces along z. However, Figure 7.11 shows that columns that

Next-generation data visualization tools

nu

It t ?

Figure 7.6. The full MD system with the semi-transparent Conditional Box
surrounding the NaCl cluster and Ne in the impact region.

are more along the diagonals from the center of the image have larger
along the crystalline surface plane. This crystal-structure
dependent distribution of stresses is brought out by the Glyphmaker bind
ings. One could also represent the off-diagonal stresses by adding another
vector, or the stress tensor could be represented in a variety of other ways.
We work out the application of Glyphmaker to tensor quantities in Ribarsky et al. (1993c); we also present a completely different application there,
visualizing a computational fluid dynamics simulation of airflow around
helicopter blades.

components

.napif
Figure 7.7. The full MD system at a time step after the cluster has embedded
itself in the Ne; here the color glyph elements are bound to local temperature.

115

116

Incerface Technology and Perception

Figure 7.8. The part within the Conditional Box at the same time step as Figure

7.7.

7.6. EXTENDING GLYPHMAKER

It is evident that the Conditional Box offers a general means to control the
data visualization. If the Conditional Box is used to provide a new set of
class labels (by sending the new label information to the Header/Label
Transcriber in Figure 7.3), then bindings of glyphs to data can be inter
actively reset, based on these new classes. This provides the power to bind
at any level of detail, even down to the individual datum, that we referred to

Figure 7.9. The full MD system (same time step as Figure 7.7) with the NaCl
cluster bindings turned off and with color bound to local temperature. The color bar
for the local temperature binding is in the lower left comer; the 0 to 1 00 scale maps
linearly onto 0 to 300 K.

Next-generation data visualization tools

vs

Figure 7.10. The NaCl cluster and crystalline substrate with Ne bindings
turned off (same time step as Figure 7.7) with visualization of the on-axis (stretch
ing/compressive) components of the atomic stress tensor added.

previously (e.g. the spatial Conditional Box can be set to any size). Of
course, the Conditional Box need not only be applied spatially; it could be
applied to any property variable. Thus one could set classes based on ranges
cf energy, temperature, force components or other variables. If a mechan
ism is provided to save and recall each of these classifications, they could be
collected for comparative viewing. In addition, there is no reason to restrict
the user to just one Conditional Box; he or she should be able to generate
multiple boxes and choose several ranges, if need be. Finally, of course,
one should be able to apply conditions more complicated than the selection

-3* *

\
*

v '

Figure 7,11. A rotated view of the system in Figure 7.10. The viewpoint is
directly over the NaCl cluster and perpendicular to the crystalline substrate surface.

1 IT

1 IS

Interface Technology and Perception

of variable ranges. These might involve algebraic expressions for data

variables or correlations between two or more Conditional Boxes.
A common thread in all these extensions is that the Conditional Box
provides the means for focusing and linking to bring out correlations
between data. These are processes that have proved quite successful in
statistical graphics where multiple views, each containing partial infor
mation about a complex dataset, are either isolated or linked (e.g._ percep
tually) so that the user can integrate them into a coherent picture of the data
relations (Buja et al., 1991 ). In fact, if we pipeline the reclassifications of the
Conditional Box to a predetermined set of bindings and then move the box
interactively, we have a version of the brushing techniques as well (Buja et
al., 1991 ). We are working on extending Glyphmaker to make many of
these Conditional Box operations possible. The advantage of the Glyphmaker approach is that one will be able to apply a condition to the data, see
the visualization updated based on the condition and adjust accordingly.

This capability for iterative refinement will produce optimized visual

analysis.

7.7. OD, 1 D, 2D, 3D, OR BEYOND?

Glyphmaker and the other glyph -based examples mentioned above display
zero degree of freedom (OD) data, since icons describe points and their
associated properties. How can we handle higher dimensional represen
tations (e.g. ribbons, isosurfaces, volumes)? Our examples and those of
other workers have shown that OD glyphs can be effectively applied to a
broad range of data types, including continuum data. Further, it is apparent
that point glyphs have certain advantages over other representations. They
can, for example, be applied to original (even noisy) data and do not require
the amount of interpolation and smoothing that, say, surface represen
tations impose. Point glyphs could also be more effective in bringing out
strong variations that may exist in all three directions in spatially dense
data since higher dimensional representations could obscure interior data
or not show all variations at the same time. Finally, point glyphs are
especially effective, as we have shown, in carrying and displaying several
variables simultaneously.
Nevertheless, despite the flexibility of point glyphs, there are cases where
other representations are more effective. For example, ribbons are best for
showing twisting spatial stmctures or flowlines that vary in time. Surfaces
or volumes may be best for bringing out the general features and patterns in
certain continuum data, especially data that are fairly smooth. Even though
some of these higher dimensional representations already exist in the
dataflow environment, it would be advantageous to put them in Glyph
maker to make use of the interactive binding and conditional focusing and
linking that we have come to value. Also, our above example shows that it
will be possible (especially as we enrich the set of glyphs) to define unique
representations based on these fonns that are customized for the data at
hand.
How then can we put the higher dimensional representations in Glyphmaker? We can certainly take a cue from spreadsheets, which allow the

Next-generation data visualization tools

interactive binding of datasets to lines, columns, piecharts, etc.. The con

cept we have followed so far is "one-to-one" binding of data objects (points
plus associated properties! to glyphs. (Actually we also allow "one-tomany" since the Glyphmaker binder allows any number of glyph elements
to be bound to one variable.) We need to extend this to include the concept
of "many-to-one" (i.e. many data objects to one glyph). For example, the
glyph might be a ribbon with associated elements of color, width, or a set of
transverse spatial dimension that are orthogonal to the ribbon trajectory.
The latter, along with color, have been used by Delmarcelle and Hesselink
( 1 993) to depict continuous tensor fields. The appropriate data in this case
would be a collection of data objects along a trajectory. Thus whether the
glyph is a ribbon, surface or volume it implies a particular data structure for
the data objects and also a data model that allows the handling of the glyph
in the same way in the editor and the binder, no matter what its dimension
ality. Beyond that a particular data preparation is also implied. It does not
make sense to bind data to a ribbon if they are not processed as a trajectory
(e.g. computed as a streamline or hyperstreamline in the case of CFD). The
best way to ensure that trajectories are available when needed is to provide
the facility to compute them from more primitive data within the visual

analysis system. This approach fits the analysis model, expanded upon in
the next section, that we argued must be integrated with the visualization
tools. Similar considerations hold for surface and volume data. In principle
it appears that a data model and an appropriate set of higher dimensional
glyphs with elements can be developed and inserted into Glyphmaker. This
will provide the glyph-based approach with the widest possible set of visual
representations for depicting all types of data.

7.8. ELEMENTS OF VISUAL ANALYSIS

We have argued for an integration of visualization and analysis because
next-generation data will be so numerous and heterogeneous that they
must be analysed on the run, and because of the great advantages offered by
interactive visualization techniques in this process. In order to establish a
general approach, we present here certain essential elements. There is not
space to expand on these ideas, but we are working towards a fuller expo
sition to be presented elsewhere.
For present purposes let us focus on data likely to come from scientific
and engineering applications. Clearly a general analysis approach should be
based on a study of the general characteristics of the physical quantities
associated with the data; that is, characteristics reflected in all (or most)
physical data. For example, much of the data come in the form of either
scalar, vector, or tensor fields that depend on space and time. These fields
have groupings in 3, 31, and 32 components, respectively. In general, 3n
dimensional fields are possible, and all these quantities imply the primacy
of the spatial coordinates in the physical system since they are usually
derived from expansions of source fields in the spatial coordinates. These
groupings imply necessary groupings for visual analysis - if possible, we
would want to look at all three vector components or all nine tensor

119

120

Interface Technology and Perception

components in the same visualization. The glyph-based work and other

research mentioned above have already provided several methods to do this

effectively. However, to be certain that we have a rich enough set of visual
representations, we should make a general study of likely physical quanti
ties. A good start has already been made on this for certain classes of
physical systems (Delmarcelle and Hesselink, 1993; Haber, 1989), and
there are general studies that can be applied (Borisenko and Tarpov,_1979).
Beyond this question, there may also be other "necessary multidimen
sional groupings" that are not 3n. It is necessary to see whether the
representations (or modifications of them) we have built for the spatial
fields will be effective for these cases and what additions to our visual
vocabulary may be needed.
For quantitative analysis, we need an algebra for combining variables
that reside in our data objects. This certainly should include a standard set
of functions (log, exp, root, power, etc.), generally useful transformations
(linear algebra operations such as diagonalization, etc.) or statistical tech
niques (regression analysis, smoothing, etc.) . Perhaps a command line
interface would be the most flexible for these operations. In any case, our
data model must be flexible enough to handle these transformations effipientlv and produce new data objects of the appropriate type for further
visual analysis.
Finally, we should address the deeper question of how researchers tend to
analyse and then develop understanding of complex data. For example,
they often reduce the dimensions of complex results by averaging over one
or more variables (including time). They then study these simpler quanti
ties and develop conceptual models that may lead to analytic models. If the
data come from observations, this may eventually lead to more complete
analyses, based on simulations or multivariable phenomenological
models, where the whole physical system is studied. This sequence of
investigation suggests that we should insert ways to choose "moments of
the data" in our interactive visual analysis system. Obvious moments
could be selected by, for example, putting boxes around a spatial region of
the data and then averaging over a chosen spatial dimension. The re
searcher could then follow the length, width or height of this region in time.
Of course, extensions of this idea could be made to other variables. Certain
data may also exhibit periodic behavior or other patterning in space, time or
other variables. Fourier and related transforms should be available and
applicable interactively to reveal this behavior. This is often quite a power
ful way to provide a unique signature for the data (analogous to the struc
ture factors from crystallography) that may be revealed in only a few
dominant components. These signatures may suggest causes for the pat
terns in the data or their dynamic behavior.

7.9. TOO MUCH VISUALIZATION EXPERTISE REQUIRED

The preceding direction for visualization software concerns rapid end-user
creation of the data visualizations. Unfortunately, even with the best
currently-known tools (in our opinion, a combination of dataflow

Next-generation data visualization tools

diagramming and glyphmakers), the scientist seeking to visualize his

information still has to know too much. Going beyond these tools, we want
to help the scientist choose an appropriate visualization without the need
to tryout a large number of possibilities. Consider the case of a dataset with
ten 2D scalar fields, and with eight ways to depict a scalar field (hue,
saturation, value, isolines, projection into 3D, size of a glyph, orientation of
a glyph, elongation of an elliptical glyph). Without considering the combi
nations of these display methods which do not make sense, there are 108
ways to bind the scalar fields to the visual representations. How is the
scientist to decide which to choose? There is really no practical way to
decide, other than to make the same choices that others have made in the
past, or to experiment painstakingly with a small number of the possi
bilities. This is an unsatisfactory situation.
What we really need are tools which allow the scientist to work at a higher
level of abstraction than currently. The scientist should be able to say to the
computer "I would like to see a comparison between the stress tensor and
temperature fields from time tj to time t2". The computer should then
propose to the scientist one or several possible visualizations which meet
that request.
What we are talking about here is declarative rather than procedural
specification of what we want from the computer. Rather than telling the
computer how to get what we want, we simply tell the computer what we
want, and expect the computer to deduce how to get here from there. This is
exactly what is done in the database world, where a typical query might be
"retrieve all data files concerning ozone levels over the north pole between
1970 and 1990 for which the average ozone level is greater than 50 metric
tons per square mile below the seasonal norm".
The current solution to this problem is to have the scientist talk not with
the computer but with a "visualization designer" who understands graphic
design, has perception, and has some knowledge of scientific concepts and
needs. NCSA calls this the "Renaissance Team" approach (Smarr, 1988).
This can work in the short term, but in the long run, it is neither practical
nor economical to provide each scientist with easy access to a graphic
designer. Nor is it practical to train scientists to become expert visual
ization creators.
The only viable longer term solution is to imbue visualization tools with
some graphic design expertise. This could be done by having the computer
serve as a critic of what has been designed by the scientist, or by having the
computer actually generate one or several possible visualizations. The first
approach, acting as a critic using the stated visualization objective, still
leaves the creation process to the user. The automatic generation alterna
tive is the more attractive, although also the more difficult: it is always
easier to criticize an artifact which someone else has created than to create
the artifact in the first place.
There are at least two general approaches to automatic generation, both
developed in the artificial intelligence community: the classical generateand-test and the more recent case-based reasoning.

Generate-and-test. The generate-and-test strategy requires a set of graphic

primitives, rules for generating visualizations from the primitives, and
criteria against which to test each generated visualization. For instance, in

121

122

Interface Technology and Perception

.'-t

of*4'

the natural language domain, words form the primitives of the English
language, grammar mles tell us how to generate sentences from words, and
our knowledge of the real world is the criterion by which we test a sentence
to determine if it makes sense.
In the graphical domain, Mackinlay (1986) was the first to use the
generate-and-test paradigm. He worked with 2D discrete data presented in
business-graphics formats such as bar, scatter and line charts. His graphics
primitives include areas (such as circles and bars), lines, marks jsuch as +
an5? and visual attributes such as color, size and orientation. The primi
tives are organized into primitive graphical languages. Each of these primi
tive languages uses one or several graphics primitives. The primitive
graphical languages include: horizontal axis, vertical axis, line chart, bar
chart, scatter plot, color, shape, size, saturation, texture, orientation, tree
and network.
The grammar for generating alternative representations is defined by
three composition mles, which combine the primitive graphical languages
into complete graphs. The rules are double-axis composition, single-axis
composition and mark composition. Double-axis composition superim
poses two sets of data encoded with identical vertical and horizontal axes on
a single set of axes, such as showing the sales data by year for two different
divisions on the same set of axes. Single-axis composition juxtaposes two
sets of data which share one common axis (but have a different second axis)
such that the common axis occurs just once. An example here would be two
bar charts side by side, one encoding gross national product (GNP) by
country and the other encoding population by country. Country is the
common axis, while GNP and population are the different second axes.
Mark composition uses two different attributes of the same area or mark to
encode two different variables, such as using the size of a circle to encode a
state's population and the color of the circle to encode the average age of the
state's population.
Rules for testing the generated alternatives are derived from a combi
nation of experimental studies (Cleveland and McGill, 1984; Cleveland,
1985; also Ware and Beatty, 1988, the published version of earlier technical
report used by Mackinlay) and experience-based guidelines (Bertin, 1981,
1983;Tufte, 1983).Therulesconcemwhattypesofinformation(nominal,
ordinal, quantitative) can be expressed by different primitives, (example)
and the relative effectiveness of alternative primitives which could be used
for the same type of information. For instance, size and saturation can both
be used to express quantitative data, but size is more effective than is
saturation.
Senay and Ignatius (1990, 1991, 1992), working with Foley, Hodges and
Sibert, extended Mackinlay's conceptual framework to 3D visualization.
Additional composition rules are defined: in addition to double-axis com
position, singie-axis composition and mark composition, Senay and Igna
tius define superimposition (placing one set of data on top of another),
composition by transparency (making one set of data partially transparent
on top of another), composition by intersection (using the 3D intersection
of two sets of data to depict both sets), and composition by interleaving
(spatially interleaving marks encoding one dataset with marks encoding
another). Figures 7.12 and 7.13 illustrate some of this work.

Next-generation data visualization tools

IN
Figure 7.12. A VISTA-created visualization created in response to a request
(created via dialog box fill-in) to show weather data variables from Peru, in decreas
ing oider of importance: rainfall, shown as a surface (because this is the most

effective available technique for showing continuous 2D data,- humidity, shown as

contours on the surface (a less-effective technique than the surface, chosen because
the user stated that humidity was less important than rainfall), temperature, shown
as color (even less effective than the contours); wind velocity, shown as blue vectors,and observation data location, shown as the small white triangles . The composition
operator of superimposition has been used to combine the contour plot and station
locations onto the surface depicting rainfall. (Courtesy H. Senay.)

Roth and Mattis (1990), in their SAGE system, extend Mackinlay's

work in another way, adding more semantic description to the data and
their relationships. For example, they represent algebraic relationships
such as total-costs materials-costs + labor-costs, allowing them to
reflect this in the graphs they produce, such as a bar chart showing totalcosts but with separate bar segments for the bar showing material-costs
and labor-costs.
Casner's (1991) BOZ takes a different approach, focusing first on the
user's objective, or task, in requesting the visualization. Different visual
izations are generated. For each visualization, the sequence of visual search
and comparison operations the user will have to perform to meet the stated
objective is also generated. The test criterion, used to determine which of
the generated alternatives to accept, is essentially the time which will be
required for the user to perform the sequence of operations. Casner exper"Ipiented with visualizations generated by BOZ and found that users' per
formance was indeed better than with less well-designed visualizations,
and that the predicted task time correlated well with the measured task
times. Using the generate-and-test design strategy, the sequence of visual
scanning and comparison operations needed to achieve the objective with
different potential visualizations is used as the criteria for testing each of
the alternatives.
Generate-and-test has been well-explored, and shows considerable
promise. Its application depends on formalizing the generation process and

123

124

Interface Technology and Perception

Figure 7.13. A second visualization of the same weather data, in which the
supermposition operator has been replaced by what Senay calls union and Mackinlay calls double axis composition. The result is that the humidity contours, wind
velocity vectors, and station locations are on the ground plane rather than on the
rainfall surface, creating a less effective presentation because some of the infor
mation is obscured. (Courtesy H. Senay.)

quantifying the test metric; for more complex visualizations, or new

domains, this is a non-trivial and not always possible task.

Case-based reasoning. In contrast to the generate-and-test strategy is casebased reasoning (Kolodner, 1993). The fundamental strategy is to organize
a large collection of existing visualizations as cases and to design new
visualizations by adapting and combining the past cases. New problems in
the case-based approach are solved by adapting the solutions to similar
problems encountered in the past. The technique has been used in a wide
variety of applications. It has not yet been applied to scientific data visual
ization, but we believe it holds promise as an alternative to generate-andtest because it requires less formalization.
The important issues in building a case-based visualization advisor are
developing a large library of visualizations, developing an indexing scheme
to access relevant cases, and determining a closeness metric to find the best
matches from the case library. We are exploring this area, but have not yet
initiated serious work. The idea would be to index the cases by the types of
data to be presented (nominal, ordinal, interval, scalar), the relative import
ance of each of the datasets, the dimensionality of the data space (typically
2D, 3D or 4D, with time possibly being one of the dimensions) and the
tasks to be performed with the visualization (such as compare, find relative
change, correlate, determine value). For purely exploratory data analysis, of
course, stating of a goal could be difficult.
Whether generate-and-test, case-based or other approaches are used, it is
also important that the visualizations generated or retrieved by the system
be editable by the user. Creating a system environment in which the system

Nexc-eeneration data visualization tools

presents the user with the "best" or "right" design is to make the arrogant
presumption that the system is always going to do at least as well as the user
could do - an unlikely proposition. We know how important this flexibility
is from our work with automatic generation of user interface dialog boxes
and menus |de Baar et al., 1992; Kim and Foley, 1993). In the context
of scientific data visualization, an attractive way to do this is to generate
the dataflow diagram and parameter settings needed to create the visual
ization: then the user can fine-tune the visualization with a well-known set

of tools.

ACKNOWLEDGEMENTS
This work is supported in part by grant numbers CDA-9200635 and
NCR-9000460 from the National Science Foundation.

REFERENCES
J. ( 1 98 1 1. Graphics and Graphics Information Processing, de Gruyter, New
York. (Translated by W. Berg and P. Scott from La Graphique et le Traitement
Graphique de 1' Information, Flammarion, Paris, 1977.)
Benin, J. (1983). Seminology of Graphics. University of Wisconsin Press, Madison,
WI. (Translated by W. Berg from Smiologie Graphique, Editions Gauthier-Villars, Paris,- Editions Mouton &. Cie, Paris-La Haye,- and Ecole Pratique des Hautes
Etudes, Paris, 1967.)
Borisenko, A. I. and Tarpov, I. E. (1979). Vector and Tensor Analysis with Appli
cations. Dover Publications, New York.
Buja, A., McDonald, J. A., Michalak, I. and Stuetzle, W. (1991). Interactive data
visualization using focusing and linking. In Proceedings of Visualization /91, San
Diego, pp. 156-163.
Casner, S. (1991). A task-analytic approach to the automated design of graphic
presentations. ACM Transactions on Graphics 10 (2), 1 11-151.
Cleveland, W. S. and McGill, R. ( 1 984). Graphical perception: Theory, Experimen
tation and Application to the Development of Graphic Methods. Journal of the
American Statistical Association 79 (387), 531-554.
Cleveland, W. S. (1985). The Elements of Graphing Data, Wadsworth Advanced
Books and Software, Monterey, CA.
deBaar, D., Mullet, K. and Foley, J. (1992). Coupling Application Design and User
Interface Design. In Proceedings CHI92 SIGCHI 1992 Computer Human
Interaction Conference, ACM, New York. pp. 259-266.
Delmarcelle, T. and Hesselink, L. (1993). Visualizing Second-Order Tensor Fields
with Hyperstreamlines. IEEE Computer Graphics and Apphcations 13 (4),
25-33.
Ellson, R. and Cox, D. (1988). Visualization of Injection Molding, Simulation 51
(5), 184-188.
Foley, I. and McMath, C. (1986). Dynamic Process Visualization, IEEE Computer
Graphics and Apphcations 6 (3), 16-25.
Foley, J. D. (1990). Scientific Data Visualization Software: Trends and Directions.
International Journal of Supercomputer Applications 4, 154.
Grinstein, G, Pickett, R. and Williams, M. (1989). EXVIS: An Exploratory Visual
ization Environment, In Proceedings Graphics Interface fS9, CIPS, Toronto,
pp. 254-261.
Benin,

125

126

Incert'ace Technology and Perception

Haber, R. (1989). "Scientific Visualization and the RIVERS Project at the National
Center for Supercomputing Applications," IEEE Computer 22 (8), 84-89.
Haber, R. and McNabb, D. A. ( 1 990). "Visualization Idioms: A Conceptual Model
for Scientific Visualization Systems." In B. Shriver and G. Nielson (eds), Scien
tific Visualization. IEEE Computer Society Press, Los Alamitos, CA.
Hai-Peng, C. and Landman, U. (1993). Controlled Deposition, Soft Landing, and
Glass Formation in Nanocluster-Surface Collisions. Science 260 (May),

1304-1307.

Kim, W. and Foley, J. (1993). Providing High-level Control and Expert Assistance in

f)

User Interface Presentation Design. In Proceedings INTERCHI 1993 bp. 430b

437. ACM, New York.
Kolodner, J. (1993). Case-Based Reasoning. Ablex.
Lefkowitz, H. (1991). Color Icons: Merging Color and Texture Perception for
Integrated Visualization of Multiple Perception. In Proceedings, Visualization
/91, San Diego, p. 164.
Mackiniay, J. (1986). Automating the Design of Graphical Presentations of Re
lational Information, ACM Transactions on Graphics 5 (2), 110-141 .
Pickett, R. M. and Grinstein, G. G. (1988). Iconographic Displays for Visualizing
Multidimensional Data. In Proceedings of the 1988 IEEE Conference on
Systems, Man, and Cybernetics, Beijing and Shenyang, vol. 1, pp. 514519.
Ribarsky, M. W., Hodges, L. F., Minsk, R. and Bruchez, M. (1993a). Visual Rep
resentations of Complex, Discrete Multivariate Data. The Visual Computer,
accepted for publication.
Ribarsky, M. W., Ayers, E., Eble, J. and Mukherjea, S. (1993b). Using Glyphmaker
to Create Customized Visualizations of Complex Data. Report GITGVU-93-25, 1993. (Submitted IEEE Computer.)
Ribarsky, W., Mukherjea, S. and Ayers, E. (1993c). Glyphmaker: An Interactive,
Programmerless Approach for Customizing Visual Data Representations. Report

#
"-r

GIT-GVU-93-26.
Roth, S. and Mattis, I. (1990). Data Characterization for Intelligent Graphics
Presentations. In Proceedings CHI '90, April, pp. 193-200. ACM, New York.
Senay, H. and Ignatius, E. (1990). Rules and Principles of Scientific Data Visual
ization. Technical Report GWU-IIST-90-13, The George Washington Uni
versity, Washington, DC, May. (Also in State of the Art in Scientific
Visualization, SIGGRAPH '90 tutorial notes, August.)
Senay, H. and Ignatius, E. (1991 ). Compositional Analysis and Synthesis of Scien
tific Visualization Techniques, InPatrikalakis, N. (ed.) Scientific Visualization of
Physical Phenomenon, Springer-Verlag, June, pp. 269-281.
Senay, H. and Ignatius, E. ( 1992). VISTA: A Knowledge Based System for Scientific
Data Visualization, George Washington University Technical Report GWUIIST-92-10, March. IEEE Computer Graphics and Apphcations, to appear.
Smarr, L. (1988). Supercomputing Offers New Challenges. Supercomputing
Review, p. 39.
Smith, S., Grinstein, G. and Bergeron, R. D. (1991). Interactive Data Exploration
with a Supercomputer. In Proceedings Visualization /91 , San Diego, p. 248.
Treinish, L. A. and Gough, M. L. (1987). A Software Package for the Data-Independent Storage of Multidimensional Data. EOS Transactions of the American
Geophysical Union 68 (28), 633-635.
Treinish, L. A. and Goettsche, C. (1991). Correlative Visualization Techniques for
Multidimensional Data. IBM foumal of Research and Development, 35 (1/2),
184-204.
Treinish, L. (1993). Inside Multidimensional Data. Byte 18 (4), 132-133.
Tufte, E. (1983). The Visual Display of Quantitative Information, Graphics Press,
Cheshire, CT.
Upson, C., Faulhaber, T., Kamins, D., Laidlaw, D., Schlegel, D.( Vroom, J., Gurwitz, R. and van Dam, A. (1989). The Application Visualization System: A
Computational Environment for Scientific Visualization. IEEE Computer
Graphics and Applications 9 (4), 30-42.
Ware, C. and Beatty, J. (1988). Using Color Dimensions to Display Data Dimen
sions, Human Factors, 20 (2), 127-42.

J**#/

Next-generation data visualization tools

12

About the author

James Foley is Professor of Computer Science and Director of the Graphics, Visual
ization, and Usability Center at Georgia Institute Technology. He was previously Pro
fessor and Chairman of Electrical Engineering and Computer Science at The George
Washington University. He earned his Ph.D. at the University of Michigan. His research
interests include user interfaces and interactive computer graphics; his work, which has
been sponsored by DEC, NSF, NASA, Sun, the Software Productivity Consortium and
Siemens, focuses on building UIDE, the User Interface Design Environment. He is co
author, with A. van Dam, of Fundamentals of Interactive Computer Graphics, and with
van Dam, S. Feiner and J. Hughes, of the recently-published Computer Graphics:
Principles and Practice. Foley is a Fellow of the IEEE, and a member of ACM, the Human
Factors Society and Sigma Xi. He is editor-in-chief of ACM Transactions on Graphics,
serves on several editorial boards and consults regularly for governmental and industrial

organizations.

William Ribarsky is Senior Research Scientist, Manager of the Scientific Visualization

Lab, and Associate Director for Service of the Graphics, Visualization, and Usability
Center, all at Georgia Institute of Technology. Dr Ribarsky's background is in computa
tional physics. His recent research has been in programmerless methods for constructing
visual representations of multivariate data, the semiotics of visual representations andthe
development of visual interfaces. Most recently he has been studying user-defined virtual
spaces for scientific visualization including investigating interfaces, graphical represen
tations, modes of interaction and methods of analysis.

\V

lc