Visualizing Aircraft Aerodynamic Design
Steve Bryson
NASA Ames Research Center
Moffett Field, California
The virtual wind tunnel (VWT) is the result of the application of virtual reality (VR) interface techniques to the visualization of the results of computational fluid dynamics (CFD) simulations. CFD is used to simulate the airflow around aircraft, thus allowing engineers controlled exploration of various aerodynamic concepts without their having to resort to expensive time in real wind tunnels. CFD simulations often result in complex, time-varying flow patterns, the details of which can be critical in determining the aerodynamic properties of the aircraft being simulated. Understanding these complex patterns is one of the challenges of aircraft design and development. The VWT helps address this challenge by providing the user an environment for the rapid investigation of CFD simulations in a fully three-dimensional control and display context, using a variety of visualization tools. However, several serious problems, determined by the engineering requirements of the simulations, have had to be addressed during development of the VWT.
VR interfaces use a variety of interface technologies in order to allow the user to interact with the computer-generated environment in a fully three-dimensional context. There are two basic components to the VR interface used in the VWT: display and interaction. The VR display is based on the use of a head-tracked stereo display to provide an image from the user's point of view, with an update rate of at least 10 frames per second. When combined with conventional three-dimensional graphics, this technology provides to the user various three-dimensional structure and depth cues, most particularly head motion and stereo parallax. By moving around in the scene, the user gets a very strong sense of the three-dimensional structure of objects in the virtual environment. The resulting, greatly enhanced three-dimensional per-
ception allows CFD researchers to see the complex three-dimensional structures in their simulations much more accurately and easily than in conventional visualization systems.
VR interaction is based on the use of three-dimensional tracking systems, which provide the user's hand position and orientation, along with simple command information. This hand position and orientation data allow the user to move visualization objects about in the environment, exploring the flow simulation in near-real time. To support such exploration, visualizations must be recomputed continually based on the user's current hand position. In order to provide good responsiveness, these recomputations should occur with delays of less than one tenth of a second.
Thus, the VR interface implies strict performance requirements of both a 10-frames-per-second (or faster) rerendering of the entire scene as well as delays of less than one-tenth of a second between user control of a visualization and when that visualization is displayed. Meeting these performance demands has been one of the primary challenges in the design of the VWT. These performance demands are encountered in the computation of visualizations, the rendering of the visualizations, any network traffic involved in the operation of the system (not addressed in detail in this abstract), and access to the data being visualized.
Computations arising in typical visualization problems include the integration of vector fields to produce streamlines and moving particles, the computation of isosurfaces and cutting planes, and the extraction of geometry reflecting various derived features of interest to the CFD researcher. These computations are performed by a choice of algorithms, which typically involve an inverse relationship between speed and accuracy. Although modern workstations may be able to compute and render a very few simple visualizations within the performance requirements described above, the complexity of modern CFD simulations makes complex visualization environments, with many active visualization techniques in the scene, highly desirable. The resulting computational burden typically saturates the computational power of even the most powerful modern workstation, particularly for time-varying simulations where all visualizations must be recomputed with every frame.
In order to meet the computational performance requirements, the time required to complete the computation must be controlled so that it is not too long. This time may be controlled in a variety of ways, depending on the details of the computation. Typical control options include the choice of more or less accurate algorithms and the number of times the algorithm is iterated for extended visualization objects. When asked, many CFD researchers have indicated that they would rather have a fast and less-complete or less-accurate answer in order to rapidly explore the simulation. When they then identify an interesting feature, that feature can be verified with a slower, more-accurate algorithm. The challenge for the VWT then becomes deciding how to control
the time required for a computation while delivering the most complete and accurate answer within the specified time constraint. This is done using time-critical computation techniques, where each visualization is given a time budget. That time budget and the time taken in the previous computational frame is used to determine the parameters of the computation so that the computation does the best job possible in the time available.
The geometry resulting from time-critical computations typically can be rendered much more quickly than it was computed, so a time-critical graphics approach has not proven to be necessary for the VWT.
By far the most difficult challenge in development of the VWT has been the very large sizes of data sets encountered in time-varying CFD simulations. These simulations can range in size up to 300 gigabytes, as of 1996, with each time step containing 60 megabytes of data. In order to maintain a time-varying animation rate of 10 frames per second, each time step can take no longer than one-tenth of a second to load, implying a bandwidth from wherever that data is stored of 600 megabytes per second. The only storage medium with the required bandwidth is the physical memory inside the computer, yet the largest memory available in graphics workstations is only 16 gigabytes. This remains an unsolved problem, and it will only worsen as simulation sizes increase. Existing compromises include restricting the data set to allow more time steps to fit into physical memory and using widely striped disks for increased bandwidth (NASA Ames has a disk system that has attained bandwidths in excess of 300 megabytes per second). Approaches based on loading only the data required for a visualization run into the problem of disk latency, which proves to be unacceptably long for VR applications. Data compression also has failed to address this problem, since compression schemes that cause losses such as subsampling or Fourier-based methods intolerably distort the data, while compression schemes without loss such as run-length encoding often increase the size of the data set. At this time multiresolution compression schemes such as those based on wavelets provide some hope, but they have yet to be demonstrated for CFD-type data.
Another set of challenges for the VWT involves the design of the user interface. Most visualization systems are based on an ''external command model," where the user specifies a visualization using either text or a graphical user interface, which is then computed and rendered. The nature of the VR display requires that the user interface be imbedded in the visualization environment, and the three-dimensional interaction capability strongly suggests a direct manipulation interface. A direct manipulation interface implies mapping the user's hand position to the specification of a visualization. These considerations have led to the concept of a "local visualization," which uses the value of a data field at a point in space to generate the visualization. Using local visualizations, a CFD researcher can explore the phenomena in a simulation by "waving the visualization about in space," rapidly sampling
features in a volume of data. The visualizations are controlled by interactive tools, which "emit" collections of visualizations. It is these tools the user manipulates directly.
There are several features of the VWT architecture critical to its success: the computation and graphics are implemented in separate, asynchronous, concurrent processes so that a slowdown in the computation does not impede the head-tracked aspects of the display; the visualizations and tools are implemented in an object-oriented class hierarchy designed so that new visualizations or tools may be added without having to modify existing visualizations or tools; and a number of display and interaction technologies are supported, ranging from conventional workstation and mouse through head-mounted displays and gloves, with a unified interaction paradigm.
The VWT currently is under evaluation release at a number of NASA sites, with a general release expected in early 1997. For more information on the VWT, contact the Web at www.nas.nasa.gov/NAS/VWT. For papers on the VWT, contact the Web at www.nas.nasa.gov/~bryson/home.html.