Dissertation

Improved Visualization and Interactivity for Flow Field Exploration and Rendering

Han, Mengjiao

2024

Overview

Flow field exploration and visualization are crucial for understanding fluid dynamics. The progress in computational power, especially in high-performance computing, allows for higher-resolution simulations of flow fields. However, these advancements introduce challenges that traditional post hoc exploration and visualization techniques need help to meet, such as limited interactivity and poor visual perception. Additionally, complex visualization algorithms can be inaccessible to scientists needing more expertise in computational visualization, highlighting the need for innovative, user-friendly approaches. This dissertation addresses these challenges by proposing new methods to enhance the interactivity and visual perception of post hoc exploration and visualization. It also aims to democratize access to advanced techniques through open-source tools. Initially, it introduces a deep-learning-based neural network for Lagrangian-based particle tracing. As the first to employ deep learning in this context, it lays the groundwork for the proposed method through extensive experimentation, including evaluating flow map extraction strategies and the effects of training samples and integration durations. Various sampling techniques and optimal hyperparameter configurations are also explored. Building on this foundation, This dissertation comprehensively evaluates the Lagrangian-based particle tracing neural network. It assesses the model’s performance across different settings, such as two-dimensional (2D) and three-dimensional (3D) time-varying flow fields, flow fields from multiple applications, varying complexities, and structured and unstructured input data. An empirical study guides best practices in model architecture, activation functions, and training data structures. Comparative analysis with existing techniques using flow maps for visualization is also included. Moreover, this dissertation explores integrating the particle tracing model with various visualization interfaces to enhance interactivity. It introduces an interactive web-based interface and integrates high-fidelity visualization capabilities with an OSPRay-based viewer, leveraging the neural network’s efficiency. To further address the need for high-performance and high-fidelity flow field rendering, this dissertation proposes a technique for ray tracing generalized tube primitives. This method, suitable for visualizing line-type data with variable radii, bifurcations, and accurate transparency, is implemented within the OSPRay open-source framework. It provides interactive, high-quality rendering with minimal memory overhead, marking a significant advancement in flow visualization.

• Publisher: ProQuest Dissertations & Theses
• Subject: Computer Engineering / Computer science / Fluid mechanics
• ISBN: 9798302152428