High performance structure extraction and analysis for visualization

Samer S Barakat, Purdue University

Abstract

The amount of data produced by simulations and experiments in science, engineering and medicine is growing at an exponential rate. Visualization research aims to address the resulting data analysis challenge through graphical representations that facilitate the identification of interesting attributes of these datasets. In particular, a large effort in the visualization community targets the creation of methods that can automatically identify so-called features of interest and therefore focus the attention of the user on the aspects of the data that are most likely to benefit his or her analysis. Unfortunately from a methodological standpoint, these features are typically domain-specific and lack a clear and objective definition. In this dissertation, we consider the general problem of characterizing the geometric structure of numerical datasets, which we define as a set of distinguished manifolds exhibited by the data in its spatial embedding. Indeed, this research is premised on the idea that the precise and efficient visualization of these remarkable manifolds can greatly improve the interpretation and assessment of scientific datasets across a wide range of application scenarios. In particular, our work focuses on the analysis of the fluid flow phenomena that are ubiquitous in scientific and industrial research. This general approach raises a number of research questions that we address in the work presented here. The first problem concerns the definition of the distinguished manifolds at the core of our proposed model. We apply to their characterization concepts from dynamical systems theory and differential geometry that are motivated by physical, perceptual, and topological considerations. Specifically, generalizations of the notions of Lagrangian coherent structures (LCS) and ridges are presented that enable the structural analysis of multivariate and multifield datasets. A second major difficulty stems from the computational cost associated with the geometric extraction of these manifolds. Extracting nonlinear manifolds spanning multiple scales from very large time-dependent datasets is a challenging task that is further compounded by the need to afford the user interactive control over the visualization. To address this problem, a method is introduced for the simultaneous extraction and visualization of LCS in large transient fluid flows. In addition, two high-performance techniques are described that allow for the efficient visual representation and geometric processing of high-quality ridge manifolds in large-scale datasets. Finally, a new scalable method for the approximation and adaptive refinement of the flow map is proposed that opens the door to an application of Lagrangian fluid flow analysis at scale in numerical simulations. Applications of these techniques are presented that span computational fluid dynamics, aeronautics, combustion research, high energy physics, climate study, and biomedicine.

Degree

Ph.D.

Advisors

Tricoche, Purdue University.

Subject Area

Computer science

COinS