Interactive scientific visualizations are widely used for the visual exploration and examination of physical data resulting from measurements or simulations. Driven by technical advancements of data acquisition and simulation technologies, especially in the geo-scientific domain, large amounts of highly detailed subsurface data are generated. The oil and gas industry is particularly pushing such developments as hydrocarbon reservoirs are increasingly difficult to discover and exploit. Suitable visualization techniques are vital for the discovery of the reservoirs as well as their development and production. However, the ever-growing scale and complexity of geo-scientific data sets result in an expanding disparity between the size of the data and the capabilities of current computer systems with regard to limited memory and computing resources. In this thesis we present a unified out-of-core data-virtualization system supporting geo-scientific data sets consisting of multiple large seismic volumes and height-field surfaces, wherein each data set may exceed the size of the graphics memory or possibly even the main memory. Current data sets fall within the range of hundreds of gigabytes up to terabytes in size. Through the mutual utilization of memory and bandwidth resources by multiple data sets, our data-management system is able to share and balance limited system resources among different data sets. We employ multi-resolution methods based on hierarchical octree and quadtree data structures to generate level-of-detail working sets of the data stored in main memory and graphics memory for rendering. The working set generation in our system is based on a common feedback mechanism with inherent support for translucent geometric and volumetric data sets. This feedback mechanism collects information about required levels of detail during the rendering process and is capable of directly resolving data visibility without the application of any costly occlusion culling approaches. A central goal of the proposed out-of-core data management system is an effective virtualization of large data sets. Through an abstraction of the level-of-detail working sets, our system allows developers to work with extremely large data sets independent of their complex internal data representations and physical memory layouts. Based on this out-of-core data virtualization infrastructure, we present distinct rendering approaches for specific visualization problems of large geo-scientific data sets. We demonstrate the application of our data virtualization system and show how multi-resolution data can be treated exactly the same way as regular data sets during the rendering process. An efficient volume ray casting system is presented for the rendering of multiple arbitrarily overlapping multi-resolution volume data sets. Binary space-partitioning volume decomposition of the bounding boxes of the cube-shaped volumes is used to identify the overlapping and non-overlapping volume regions in order to optimize the rendering process. We further propose a ray casting-based rendering system for the visualization of geological subsurface models consisting of multiple very detailed height fields. The rendering of an entire stack of height-field surfaces is accomplished in a single rendering pass using a two-level acceleration structure, which combines a minimum-maximum quadtree for empty-space skipping and sorted lists of depth intervals to restrict ray intersection searches to relevant height fields and depth ranges. Ultimately, we present a unified rendering system for the visualization of entire geological models consisting of highly detailed stacked horizon surfaces and massive volume data. We demonstrate a single-pass ray casting approach facilitating correct visual interaction between distinct translucent model components, while increasing the rendering efficiency by reducing processing overhead of potentially invisible parts of the model. The combination of image-order rendering approaches and the level-of-detail feedback mechanism used by our out-of-core data-management system inherently accounts for occlusions of different data types without the application of costly culling techniques. The unified out-of-core data-management and virtualization infrastructure considerably facilitates the implementation of complex visualization systems. We demonstrate its applicability for the visualization of large geo-scientific data sets using output-sensitive rendering techniques. As a result, the magnitude and multitude of data sets that can be interactively visualized is significantly increased compared to existing approaches.