by Diana Phillips Mahoney, Computer Graphics World, July 1995.
With their vibrant peaks and valleys, mesmerizing geometric gardens, and dizzying fractal horizons, the landscapes of scientific visualization often exemplify computer graphics at its best. Yet, scientific visualization is not about beautiful computer graphics. It's about inquiry and discovery - the quest to make sense of vast collections of numerical and physical data representing an endless array of natural phenomena.
In a scientific visualization, every pixel has meaning; each morsel of data potentially holds a key to unlocking one of nature's many secret doors. Variations in color and texture highlight distinctions in the data. Animation demonstrates the evolution of the data over time. The ability to display multiple dimensions within a single plot showcases relationships among the data that might otherwise be hidden. That the resultant imagery is awesome is merely a consequence. "Scientists are not trying to make good-looking stuff", says Tom DeFanti, co-director of the Electronic Visualization Laboratory at the University of Illinois in Chicago. "They're trying to make things that show the parameters of what they're seeking to discover and to get the intuition they need to do that. Visualization is part of the experiment. It's certainly not the end goal."
Computer-generated scientific visualizations are not new. For years now, scientists have been looking to visual representations of their research data to better understand such things as hurricane patterns, planetary evolution, and the propagation of cancerous cells through human tissue. Yet, over the past few years, the visualization market has begun a subtle, though not insignificant, transformation. Early on, scientists and visualization experts typically worked independently of each other -the former collecting the data, the latter generating the visual representation of it. Today, many of the visualization tools are in the hands of the scientists themselves, whith visualization specialists stepping in to help customize the software environment to best serve the researchers' needs or to generate a "finished product," such as a videotape for professional and peer presentation.
"Early visualization work was extraordinarily sophisticated in the sense that it took a number of people with specialized knowledge to get things done," says industry analyst Carl Machover of Machover and Associates (White Plains, NY). "You needed people familiar "with computers and programming, you needed people familiar with databases, and you needed people familiar with science. Rarely would one person wear all three hats." In the past few years, however, "there's been a lot of work going on to get this stuff directly into the hands of the scientist. There are still some very exotic visualizations that require a lot of involvement, but for the most part, more people are able to make use of the technology," he says.
Shane Mayor, visiting scientist with the National Center for Atmospheric Research (NCAR) in Boulder, Colorado, agrees. At NCAR, Mayor works with the Optical Remote Sensing Group, which develops laser instruments called lidars (for light detection and ranging) to probe the atmosphere. He uses PV-Wave visual data analysis software from Visual Numerics (Houston) to analyze and visualize windshear, turbulence, and ozone data gathered via different types of lidars. Previously a programmer for NASA Langley, he has experienced the benefits of today's visualization software first-hand. "When I started programming at NASA, I was using FORTRAN and NCAR Graphics. If somebody came to me and said, 'I'd like to plot a time series of ozone,' it was a bear. Now it's relatively easy and can be done interactively. From a command line, you can say, 'Plot X and Y, where X is the time array and Y is the ozone array.' [The software] makes some assumptions, then automatically puts it up and scales it for you."
While the software's automation is great, Mayor says, it can be easily overridden via customized routines. "I can define a normal coordinate system and just go in and write PV-Wave code to draw line segments and points one at a time, make my own tick marks, put in my own annotation. I have complete control with it. "In fact, Mayor says he hasn't found anything he can't do with the program. I've read variable-length, binary data files. I read and write ASCII stuff. It generates PostScript. It has a complete math and stat library built in, so I have all sorts of curve-fitting functions. Just name it."
In addition to software improvements, hardware capabilities, wich are expanding at an explosive rate, are contributing to the visualization market's expansion (according to Machover, the annual growth rate for this market is 16% to 18%, compared to a 10% to 15% growth rate for the computer graphics market in general). Powerful multiprocessor and parallel-processin workstations as well as supercomputing servers are often handling tasks once reserved for supercomputers. And supercomputers themselves are more powerful. As a result, researchers are able to manipulate their data in ways never before possible.
The ability to "steer" computations is an example of these new capabilities. "In most classic computations, people have essentially run a machine for, 100 hours, created a big dataset, then tried to figure out what's in the dataset by doing data mining of some sort," says DeFanti. While that pattern has not disappeared, new methods are proving far more efficient. "People have begun saying, 'Let me try and steer around this data and see if, by interactively working on it, I can get some intuition that I can't get the other way.'"
According to David Uhlir, product manager for Research Systems (Boulder, CO), developers of IDL visualization software, "The technology's finally at a point where we're able to see fullscreen animations that look real, and we're able to render things in a quick enough period of time to satisfy impatient people."
These developments are a function not only of hardware improvements, but also of more creative software development, says Uhlir. "It's not so much that we're learning new tricks as it is we're learning ways not to do things. Also, we're figuring out things like how to be conservative of disk space and input/output and how to do the calculations that go into a realtime animation off-line so its performance will be in real time." Additionally, researchers are learning better ways to sift through their data before visualizing it, he says, "instead of taking the approach of 'let me see everything, then I'll figure out what I don't need.'"
Technical considerations aside, much of the success of any visualization is a function of subjective, conceptual considerations. Researchers must determine the type of representation (an isosurface, volume rendering, or ball-and-stick model, for example) that would best suit the data being considered as well as how to use colors, textures, and animation to enhance understanding. Because the goals of visualizations differ, each visualization needs to be considered individually, says Dan Clark, marketing director at Visual Numerics. "A problem in computational fluid dynamics is very different from one in chemical-reaction optimization, for example. The distinctions in one process may be subtle, while those of the other may be significant. So you use different types of plots, different functions, different filters -different tools in general- to manipulate and display the data."
The number of dimensions a researcher wants to study will also impact the mode of representation. A single plot can display up to, and perhaps more than, seven variables, says Clark. "Suppose you choose a picture of a mountain. Your X and Y axes are your latitude and longitude, and Z is the height. If you drape color over that to represent snow depth or biomass, you've got a 4D plot. If you animate that, showing how that fifth variable changes as a function of X, Y, and Z, plus over time, you've got a five-dimensional plot. If you make vectors, you could do a six-dimensional plot and show the direction that those things are moving in at any point in time, if that's relevant. Then you could do a color-graded vector, or a vector that has a scalar magnitude of change in length or width. "Ultimately, he says, "you'd end up with a seven-dimensional plot from which you could extract a tremendous amount of information that you'd never be able to see any other way."
While there's no one-size-fits-all process for generating an effective visualization, there are a number of steps, in addition to choosing the appropriate mode of representation, that are critical to this objective. One is the use of color. "Once we get the data into a visualization environment, we look at what type of color maps would best show what we're trying to see," says Theresa Marie Rhyne, lead scientific visualization researcher for Lockheed Martin at the US EPA Scientific Visualization Center (Research Triangle Park, NC). Rhyne notes that personal preference can also be a factor in color definition. "Some of our scientists only want discrete color maps. They don't like maps where the colors gradually transcend from one to the next. They want it very defined so they can put a little label out there that says, where it's yellow it's 20 parts per billion, and when it goes to green it's 30 parts per billion."
Sometimes color choice is based on accepted standards. "In physics, people talk about the red shift and the blue shift. Red shifts mean that you're shifting the frequency positively, or motion away, and blue shifts are shifts in the negative frequency direction, or toward you," says NCAR's Mayor, who has built two color tables for his group's lidar research based on these color conventions. He notes that building color bars can be tricky business. "I spend so much time listening to complaints about the color bars I implement. Someone says, 'There's not enough contrast between the 11th and 12th bins,' or 'I can't distinguish those two colors well enough.' After I go in and change it, some guy down the hall can't make the distinctions. It's almost an aesthetic thing."
Texture mapping provides another way to highlight distinctions in complex data. "Scientists don't normally think of using textures, but when they do, it makes the images much livelier," says EVL's DeFanti. "People play around with different textures to see if it gives them information about motion or about shape. Then they can make the textures semitransparent and pile them on top of each other."
Some visualization applications have specific requirements that may not be applicable across the board. For example, to analyze data representing a large geographic domain, the information must be geographically registered, says Rhyne, whose work at the EPA Visualization Center involves providing visualization support for studying such things as air-pollutant concentrations and water quality. "If the visualization is over a geographic domain, the dataset has to be examined to see how it's going to map onto the geographic region." Examples of the group's "bloopers," she says, include putting Florida where Maine is.
Another consideration is "cleaning" the data, says Visual Numerics' Clark. "Some people just take their data and plot it, then go back and get rid of the noise. Other times, people decide to filter the noise before plotting, so it looks smoother." What's important to realize, he says, is that people don't generally go from data to finished plots. "There's something else involved in terms of signal or mathematical processing or statistical analysis."
As an example, Clark describes the visualization process often implemented for geophysics research. "These researchers often have very sparse data that's scattered over a wide region. Because their data is regularly sampled, they need to do what's called gridding to get a uniform sampling of data points. Because most gridding algorithms do a smooth interpolation, any sign of a fault in the Earth's crust would be obliterated." Consequently, researchers must manipulate the data by applying a specialized gridding algorithm that can handle faults and discontinuities while allowing the remainder of the dataset to be interpolated.
Such requirements, however, raise the sensitive question of data accuracy. If the geophysics researcher isn't aware that a fault might be present, how does she know to look for it in her data? And if she doesn't look for it and apply specialized algorithms, wouldn't the resultant visualization be invalid? "That's the whole reason for exploring the data and applying different techniques," says Clark. "The researchers might try two or three different gridding algorithms, and if they see something that's incongruous between them, they know more investigation is in order. If you're in the San Francisco Bay area and your seismic analysis shows no faults, you ought to be worried about your data."
Because the generation of visualizations involves some subjective judgment, the risk of misrepresentation is always present, but that's not really a computer graphics issue, according to many people in the industry. "It's always been possible to lie with statistics; you don't need a computer to do that," says Rhyne. "It's just that now we're able to produce very seductive pictures, so people are more focused on the accuracy issue."
In this sense, the appeal of the imagery is a Catch-22, says Research Systems' Uhlir. "When USA Today picks up some of the visualizations that have been done, particularly those in the global-change realm, they just show the pretty pictures but they don't highlight the 'gotchas'." Uhlir considers one of the most offensive "gotchas" to be the establishment of questionable cause-and-effect relationships. "When you're dealing with an extremely complex system, a one-to-one correlation between, say, the heat of the Pacific Ocean and the weather in North America does not by itself mean cause and effect because there are all these other variables that are not being tracked that could be more responsible for the observed effect. I think that's the biggest abuse or misleading use of visualization."
Chances are, the seductive qualities of scientific visualizations are only going to grow, as an even sexier technology -virtual reality- becomes part of the mix. Currently, various research facilities are evaluating immersive capabilities for analyzing complex data.
In the Virtual Environment Laboratory at NASA's Goddard Space Flight Center, researchers have modified existing visualization tools to accommodate various VR devices, including a counterbalanced stereo display, called a BOOM, from Fakespace (Menlo Park, CA), a gesture glove, and a virtual wand. The goal is to increase information throughput between the data itself and the scientist, says project manager Stephen Maher. "Our intent is to take advantage of the natural interaction metaphors that VR touts: Just turn your head, and you move within the world; put on a glove and directly manipulate objects. We have an immense amount of data [generally oceanographic and atmospheric data] that the scientists need to identify patterns in, so the more we can bombard their senses, the more likely they'll be able to comprehend the data."
In the Electronic Visualization Lab at the University of Chicago, researchers are navigating through their 3D datasets via the CAVE, a projection-based system in which stereo images are displayed on screens surrounding the viewer. "The single major benefit of the technology right now," says EVL's DeFanti, "is the feel we get for the three dimensionality through the tracking, especially in spaces that people fundamentally don't understand."
Besides the often-mentioned technological and financial obstacles plaguing virtual reality across all applications, a roadblock particularly debilitating to scientific applications has been the inability to record virtual experiences. "If you can't record your science, it doesn't count as science, fundamentally," says DeFanti. Without this capability, he adds, "scientists are required to go back to the workstation after they do the VR part and somehow reduce what they're doing to a number someplace, or to create a single picture." A collaborative project between the EVL and the National Center for Supercomputing Applications, called the Virtual Director, is addressing this issue. "It allows you to basically have a camera in the space and move around the space and create a path. Then you go back and edit it, change timings, and fit splines through points," says DeFanti.
In addition to research applications, scientific visualizations are often used as bases for communication. For example, the visualization researchers with the EPA's Scientific Visualization Center generate high-end animations using Wavefront software to demonstrate the agency's major research activities and educate policymakers. For example, says Rhyne, when the Clean Air Act was being developed, the group created an animation of air-Pollution buildup and illustrated various potential scenarios if one or another, control strategy was implemented. "We were trying to show people who were going to set policy what was going on with photochemical modeling. It turns out photochemical modeling was written into the Clean Air Act legislation. "
It is applications such as these that drive visualization technology, says Research Systems' Uhlir. "Ten and 15 years ago, people were looking at some of the Grand Challenges kinds of problems and thinking, 'I can't do this in my life-time, given the computer, power, I have.' Today they're saying, 'I can do this.'" And, he adds, while many of the scientific problems are still bewilderingly difficult to comprehend, "the good news is that all these complex systems that people are studying aren't getting any more complex, so the hardware and software are catching up."