Deep Learning Approaches For Exploring Collections of Visual Features of Scalar Fields

Abstract

Scalar fields are a common type of scientific data from numerical simulation, scanning, or sensor networks. We consider two types of visual features computed from scalar fields for analysis, the images of the scalar fields and the images of topological features of the scalar fields. Especially, we address the problems of studying those visual features in collections. In most settings of scalar field analysis, it is not sufficient to study one instance of a static image. We present two projects, one for each type of visual feature, to demonstrate our approach to utilizing deep learning to explore collections of visual features. In the first project, we study a technique called direct volume rendering, which studies volumetric scalar fields by mapping each volume element to a color and opacity and projecting them onto an image screen. This process has a substantial parameter space for camera setup and color and opacity mapping. It is challenging for researchers to choose the best rendering parameters before exploring the parameter space. Our first project considers the volume-rendered images of different parameters as a collection of visual features. We utilize a deep learning approach to learn a representation of mapping from the space of parameters to their associated visual features. We then provide a visualization tool to help users explore the space of the parameters associated with their resulting volume-rendered images to aid users in choosing the desired parameters. Our second project looks at a topological data analysis approach that investigates the Morse complex of a scalar field. Given a collection of scalar fields, commonly found in time-series, multivariate studies, or ensemble studies, understanding what types of Morse complexes are in the collection and examining the relationship between Morse complexes are non-trivial tasks. In this project, we consider the collection of Morse complexes as a collection of visual features representing the domain's topological decomposition. We present a proof-of-concept deep learning approach to extract the geometry of the decomposition and apply visualization techniques to study the space of features we extracted. Our approach can study collections of Morse complexes from different sources, including synthetic functions, time-varying flow vorticities, and microscopy images.