Neural Representations of Flow Data for Visual Analysis
Flow visualization serves as a crucial tool for scientists and researchers to gain valuable insights into complex flow data. However, with the growing complexity and scale of the numerical simulations, the resulting datasets have expanded considerably, presenting challenges in terms of storage, transmission, and subsequent analysis. Motivated by these challenges, this dissertation explores the integration of neural representations within flow visualization frameworks. By leveraging the latest developments in deep learning, novel approaches are developed to enhance computational efficiency, improve data reduction and enable interactive visual analysis. First, we investigate the fusion of neural representation with visualization tasks. We propose a technique to integrate streamlines within the superresolution framework enabling the generation of high-resolution flow fields that enhance the fidelity of flow visualization. Next, we propose a technique to reconstruct arbitrary flow map samples using an implicit neural representation (INR), providing a compact and accurate representation for comprehensive analysis. To further bridge neural representations and visualization, we propose a technique for incorporating scale information within INRs, offering a compact model that permits filtering, sampling, and progressive representation of data. Finally, a comprehensive framework is proposed for learning neural representation of flow maps that strikes a balance between accuracy, computational efficiency, and scalability. Overall, we demonstrate the potential of neural representations as a tool that scientists and researchers can utilize to extract essential flow insights, manage data effectively and enable accurate visual analysis.