A data model and algorithm design for scientific data visualization

A well-proven model for data visualization is to extract individual data features based on numerical fields or to construct parameterized geometric objects at discrete points in space. However, most visualization software in use today offer turn-key systems making it difficult for the user to go beyond a simple application of individual visualization techniques. Moreover, most implementations have a strong bias towards generating graphics-ready primitives; these graphics objects are not suitable as inputs to other visualization techniques because they only contain rendering information (RGB values and normals) and no numerical data, or because their data type is not compatible with the strict typing rules enforced by the newer visualization systems.; Some object-oriented systems offer no class hierarchy and are not open to class (and type) derivation, making the production of custom data types difficult at best. Other systems offer a monolithic data model providing too much generality.; In this thesis, we endeavored to augment this visualization paradigm. A new object oriented design with a data model and algorithm architecture provides two open hierarchies of data types and operations at the cell level and at the grid level. Structured and unstructured data grids are fully supported with hierarchies of data types which are naturally based on the zero-, one-, two- and three-dimensional worlds. All common visualization techniques for gridded data are implemented to efficiently and safely exchange re-usable Field Visualization Objects (FVO). Each FVO is a genuine instance of our class hierarchy, inheriting fixed behaviors and interfaces, with direct access to the numerical fields under study. The serial composition of visualization techniques on all FVOs is emphasized by providing a data model which allows composite visualization feature extractions, while offering many different intermediate visual representations. This Functional Composition proves very useful for studying interaction and correlation between input data fields. A uni-variate dataset can also benefit from this environment by associating a field with derived quantities such as grad(), curl(), div(), etc. Each class in the class hierarchies is provided with numerical and display methods which are not tied to particular hardware rendering platforms.; To support the serial composition of techniques on FVOs, stronger requirements are placed on some common visualization algorithms. A new algorithm for the construction of iso-value surfaces in heterogeneous grids is presented to provide well-behaved, fully connected and consistently oriented surface meshes. It is designed as a surface-tracking algorithm.; We use many surface-based rendering methods, but also developed a unique grid traversal technique for the photo-realistic rendering of tetrahedral grids. It is a ray-tracing technique which uses the data mesh itself as an unstructured bounding shape. It provides a tight bounding box and an adaptive subdivision of space and yields performances comparable to classic ray-tracing implementations.