Browsing by Author "Grottel, Sebastian"

Now showing 1 - 1 of 1

Open Access
Point-based visualization of molecular dynamics data sets
(2012) Grottel, Sebastian; Ertl, Thomas (Prof. Dr.)
The analysis and especially the exploration of large data sets from simulations often benefit from visualization. If it is not possible to calculate some well-known characteristic values, or if it is known that spatial features like distributions may play an important role, directly looking at a visual representation of the data becomes a first and important step in the analysis process. This is especially true for scientific work where simulations with novel methods or novel problem scenarios produce data with potentially unknown features and effects. Particle-based simulation methods, like molecular dynamics, smooth particle hydrodynamics, or the discrete element method are prominent examples as these simulation techniques themselves, as well as the scenarios they are applied to, are in the focus of current research of the corresponding application fields, e.g. physics, thermodynamics, biochemistry, material science and engineering. The sizes of the simulation scenarios, and as a consequence the sizes of the resulting data sets, steadily increased during the last years to close the gap between length scales accessible through experiments and length scales reachable by simulations. This is not only because the increased compute power of individual machines or the availability of relatively cheap compute clusters to be used on-site, but also because of significant improvements of the simulation algorithms themselves. Most of the available visualization tools are insufficiently optimized to cope with the data set sizes they have to face nowadays, as they often require the whole data set to be stored in main memory or exclusively use mesh-based rendering methods. For rendering particle-based data sets point-based ray casting has become the state-of-the-art approach and large data sets are typically addressed by different data-streaming and out-of-core techniques. This thesis investigates the performance bottlenecks of secondary storage and the data transfer between main memory and graphics hardware, as well as the impact of different compression techniques using spatial data structures and quantisation of coordinates, to optimize the foundation for such streaming techniques. Point-based ray casting is presented for different kinds of graphical primitives, like spheres, cylinders, compound glyphs, and polyhedral crystallites, extending the applicability of this rendering approach. To optimize the rendering performance the required calculations are reduced to the necessary minimum employing advanced culling techniques, allowing for interactive visualization of data sets with hundreds of millions of particles. A second and probably more important aspect when visualising large particle data sets remains: creating meaningful and useful visualizations. Even a system capable of rendering millions of particles will usually generate images that are prone to aliasing, visual clutter, and other effects hindering good perception and thus the understanding of the presented data. This thesis presents two approaches of advanced shading and lighting to remedy this issue. An image-space method to estimate normal vectors for the structure implicitly formed by the particles addresses the aliasing problem, while the perception of the global structure and depth of the data is enhanced by a specially adapted ambient occlusion technique approximating global illumination. For more efficient visualizations, relevant structures need to be derived from the original particle data. Such consolidated visualizations provide a better overview of the structure of the data sets. However, as they reduce the visual information, such representations must be created with care to be sure not to omit important data or introduce misleading artefacts from the applied methods. This thesis presents several examples, highlighting two aspects: spatial structures and representatives for the dynamics of the data. The examples of the first group range over molecular clusters, i.e. droplets in the context of thermodynamics nucleation simulations, dislocations and stacking faults from material science, and generic surface representations, similar to molecule surface descriptions, known from biochemistry. The dynamics of data are given by examples of the interaction of molecule clusters, the clustering of path lines of water-protein interactions and the tracking of expelled material in laser ablation simulations. To reliably avoid problems, like introducing artefacts or misleading presentations, with derived representations it is important to always involve experts from the corresponding application domain, because only those are able to judge usefulness and correctness of a visualization compared to the original data. Such close collaborations are most fruitful if the application domain expert can also be the actual user of the visualization tools, e.g. allowing to experiment with parameter settings.