Advanced Visualization Techniques and Data Representations for Large Scale Scientific Data

Scientific simulations provide a critical means for understanding and predicting important natural phenomena, often having significant impact on policy-making and the environment's well-being on the regional and global scales. The output of a typical leading-edge simulation is so voluminous and complex that advanced visualization techniques are urgently needed to explore and interpret the computed results. The new challenges of visualizing large simulation data are mainly imposed by the fact that data are too massive for transferring, storing, and processing. The gap between data generation and scientific discovery is getting wider. A viable solution to bridge the disparity is based on the concept of in-situ processing that can greatly reduce data movement and storage requirements by coupling visualization with simulation. It thus requires designing and deploying new parallel visualization techniques on cutting-edge high performance systems characterized by heterogeneous processors, a high level of concurrency, and deep memory hierarchies.;This dissertation makes contributions to the design of new visualization and data representation techniques to facilitate large-scale visualization on highly parallel distributed systems. We carefully study novel data representations of large and complex simulation data, and explore corresponding data partitioning and distribution schemes to ensure the stability of a visualization system in a large heterogeneous computing environment. Another task of this research is to exploit intra-node and inter-node parallelism at a high level of concurrency to improve parallel efficiency of visualization algorithms. We also study the communication patterns and data access patterns of parallel visualization process, and evaluate and enhance our new data representations to minimize inter-node data exchange. Lastly, we pair these techniques with multi-resolution advantage of data abstraction guided by an uncertainty-driven approach to make it possible to realize scalable visualization solutions for large simulations.;We carry out the experimental study based on selected, representative simulations and corresponding applications, such as high-performance and high-quality visualization of climate models and efficient data representations for the analysis of large-scale flow simulations. We demonstrate that well-designed visualization techniques and data representations for simulation data can facilitate more responsive and intuitive studies of visualization at large scale, and hence enhance scientists' potential to discover complex patterns and understand numerical simulations.