GV: Small: Collaborative Research: An Information-Theoretic Framework for Large-Scale Data Analysis and Visualization

超级计算机不断增长的能力为科学家更详细地模拟更复杂的问题的能力提供了重大进步，导致了高影响力的科学和工程突破。为了充分理解海量数据，科学家需要可扩展的解决方案，可以在不同的细节级别执行复杂的数据分析。多年来，可视化已经成为分析各种计算密集型应用程序生成的数据的重要方法。然而，可视化参数的选择和重要特征的识别主要是以特别的方式完成的。为了使用户能够系统和有效地探索数据，在这项由俄亥俄州立大学和密歇根理工大学参与的合作研究中，PI探索了一个信息理论框架来评估可视化的质量并指导算法参数的选择。研究小组计划开发一个基于信息理论的四层分析框架。该框架的最底层由信息度量的组成部分组成，其中数据被建模为概率分布。在信息度量组件的基础上，在框架的第二层中，评估和优化了最常用的可视化算法，包括等值面提取和流水线生成，以有效地揭示数据中的最大信息量。PI还研究与图像空间中的信息测量相关的问题，并优化直接体绘制结果。该框架的第三级侧重于对时变和多变量数据集的分析。将发展在时变数据集中识别重要的时空区域和测量多变量数据集中的信息流以识别不同变量之间的因果关系的方法。在该框架的第四层中，信息论用于评估多分辨率体积和图像中不同细节级别的质量，并选择细节级别以优化可视化质量，同时满足潜在的性能约束。该项目的关键完成将是开发一个严格的基于信息论的解决方案，以帮助科学家理解大规模模拟和有效可视化产生的海量数据。为了将这项研究瞄准现实世界的应用，PI正在与桑迪亚国家实验室的燃烧科学家合作，后者处于各自领域的前沿，使用极端规模的计算来解决最具挑战性的问题。四层信息论框架将使用将向一般用户发布的可视化工具包(VTK)来实施。项目中开发的新算法和新技术将通过项目网站(http://www.cse.ohio-state.edu/~hwshen/Research/NSF_GV2010)，传播，并在私人投资机构积极参加的年度可视化和具体应用会议上作介绍。传播计划还将包括通过新闻、故事和演示文稿接触普通受众，以增强他们对可视化价值的理解和欣赏。该项目在计算科学、大规模数据分析和可视化领域为研究生、本科生和代表性不足的学生提供培训。

The growing power of supercomputers provides significant advancements to the scientists' capability to simulate more complex problems at greater detail, leading to high-impact scientific and engineering breakthroughs. To fully understand the vast amounts of data, scientists need scalable solutions that can perform complex data analysis at different levels of detail. Over the years, visualization has become an important method to analyze data generated by a variety of computationally intensive applications. The selection of visualization parameters and identification of important features, however, are mostly done in an ad-hoc manner. To enable the user to explore the data systematically and effectively, in this collaborative research effort involving the Ohio State University and the Michigan Technological University, the PIs explore an information-theoretical framework to evaluate the quality of visualization and guide the selection of algorithm parameters.The research team plans to develop a four-tier analysis framework based on information theory. The bottom tier of the framework consists of the components of information measures where data are modeled as probability distributions. Based on the information measurement components, in the tier two of the framework the most common visualization algorithms including isosurface extraction and flowline generation are evaluated and optimized to effectively reveal the most amount of information in the data. The PIs also investigate issues related to information measurement in image space and optimize the direct volume rendering results. The tier three of the framework is focused on the analysis of time-varying and multivariate data sets. Methods for identifying important spatio-temporal regions in time-varying data sets and to measure the information flow in multivariate data sets to identify the causal relationship among different variables will be developed. In the fourth tier of the framework, the information theory is used to assess the quality of different levels of detail in multi-resolution volumes and images, and to select the level of detail to optimize the visualization quality while satisfying the underlying performance constraints.The key accomplishment of this project will be the development of a rigorous information theory based solution to assist scientists in comprehending the vast amounts of data generated by large-scale simulations and effective visualizations. To target the research at real world applications, the PIs are collaborating with the combustion scientists at Sandia National Laboratories who are at the forefront of their field to employ extreme-scale computing to solve the most challenging problems. The four-tier information-theoretic framework will be implemented using the Visualization Toolkit (VTK), which is to be released to general users. New algorithms and techniques developed in the project will be disseminated through the project web site (http://www.cse.ohio-state.edu/~hwshen/Research/NSF_GV2010), presentations at the annual visualization and application-specific conferences that the PIs have been actively participating in. Dissemination plan will also includes reaching general audiences through news, stories, and presentations to enhance their understanding and appreciation of the value of visualization. This project provides training to graduate, undergraduate, and underrepresented students in the area of computational science and large-scale data analysis and visualization.