Deep Learning for Scientific Data Representation and Generation

Scientific visualization (SciVis) is one of the core components in supporting fundamental scientific discoveries and engineering designs. For example, scientists perform numerical simulations and produce 3D scalar and vector data to visualize, analyze, and understand various kinds of natural phenomena, such as climate change and star formation. However, the cost of these simulations is expensive when time, ensemble, and multivariate are involved and the scientific data are presented in diverse forms including streamline, pathline, stream surface, volume, and isosurface. A core problem in SciVis is how to efficiently and effectively produce and analyze these diversified data. In this dissertation, I develop novel deep learning methods to enable more effective and efficient frameworks for scientific data representation and generation.

In scientific data representation, I propose a unified framework that processes both streamlines and stream surfaces through auto-encoder decoder structure. Moreover, I build an interface that allows users to explore the relationships between the learned features and visual representations. I also utilize geometric deep learning (e.g., graph neural network) to extract node-level and surface-level features in an unsupervised fashion for node clustering and surface selection tasks. In scientific data generation, I introduce a comprehensive pipeline for variable selection and translation through feature learning, translation graph construction, and variable translation. This framework can serve as a data extrapolation and compression solution to reduce simulation costs. Besides, I develop an end-to-end generative framework that can synthesize spatiotemporal super-resolution volumes with high fidelity. Further, to improve network generalization, I propose an unsupervised pre-training stage using cycle loss. This spatiotemporal super-resolution approach can upscale data up to 512 times in spatial dimension and 11 times in temporal dimension, which offers scientists an option to reduce data storage.