Deep Learning for Scientific Visualization: Data Generation and Visualization Synthesis

Scientific simulations generate massive, time-varying volumetric data that pose significant challenges in I/O, storage, visualization, and analysis. This dissertation introduces deep learning–based frameworks to improve efficiency and flexibility in scientific visualization.

First, the Generalized Multivariate Translation (GMT) model enables scalable one-to-many and many-to-many variable translations, reducing training overhead while improving accuracy. Then, ViSNeRF leverages neural radiance fields for efficient, high-quality visualization synthesis from sparse images, supporting parameter-space exploration with multi-dimensional representation. Building on this, ReVolVE reconstructs volumetric scenes from rendered images, enabling visualization enhancement without access to raw volume data. A comprehensive study of neural surface reconstruction methods benchmarks their effectiveness on scientific datasets. Finally, VolSegGS introduces a Gaussian-splatting approach for real-time segmentation and tracking in dynamic volumetric scenes.

Together, these contributions advance deep learning–driven solutions for scientific visualization, offering faster, more flexible, and higher-fidelity analysis tools.

科学模拟会生成海量的时变体数据，这类数据在输入/输出、存储、可视化与分析环节均面临显著挑战。本论文提出了基于深度学习的框架，以提升科学可视化的效率与灵活性。

首先，广义多变量翻译（Generalized Multivariate Translation, GMT）模型支持可扩展的一对多与多对多变量转换，在降低训练开销的同时提升了模型精度。随后，ViSNeRF借助神经辐射场（Neural Radiance Fields）实现了基于稀疏图像的高效高质量可视化合成，并支持通过多维度表征开展参数空间探索。在此基础上，ReVolVE可通过渲染图像重建体素场景，使得无需原始体数据即可开展可视化增强成为可能。本研究还针对神经表面重建方法开展了系统性调研，并在科学数据集上对各类方法的有效性进行了基准测试。最后，VolSegGS提出了一种基于高斯溅射（Gaussian Splatting）的方法，可在动态体数据场景中实现实时分割与跟踪。

上述各项研究成果共同推动了面向科学可视化的深度学习驱动解决方案的发展，提供了更快速、更灵活且保真度更高的分析工具。