Off-campus UNL users: To download campus access dissertations, please use the following link to log into our proxy server with your NU ID and password. When you are done browsing please remember to return to this page and log out.
Non-UNL users: Please talk to your librarian about requesting this dissertation through interlibrary loan.
3D Tracking and Analysis on Multivariate Time-Varying Scientific Data
Abstract
Modern supercomputers enable domain scientists to conduct simulations by solving dynamic systems of unprecedented complexity and generate vast amounts of data involving hundreds or even thousands of time steps. The collected data, in their raw format, is often redundant and hard to be processed. Therefore, it is beneficial to acquire an efficient and meaningful representation of the data. There are two viewpoints when it comes to the modeling of time-varying scientific data. The first is the Eulerian viewpoint. We can consider the time varying dataset as a 4 dimensional dataset, which is composed of 3 spatial dimensions and 1 temporal dimension. We train deep models to learn the concise representation of the 4D dataset, which captures the spatial features and temporal dynamics surrounding each voxel. The second is the Lagrangian viewpoint, which tracks a single "particle" through space and time and establishes the dynamic equation of its locations. Given the key frames, we effectively reconstruct any intermediate frames by tracing all the particles. We first train a model to reconstruct the datasets using linear interpolation, assuming each particle moves linearly when the time span is small enough. Then we generalize our model and trace the particles in a probabilistic manner, that is, each trajectory can be interpreted as a stochastic process. Thus in this way, we can describe the uncertainty for each trajectory systematically. Furthermore, analysis of these probabilistic trajectories can be conducted by clustering them and comparing the patterns of various clusters. Extensive experiments and visualizations on various dataset are also conducted, showing our method can help scientists gain useful insight on the underlying simulation results. To assess the deep model as a good probability estimator, we further conduct extensive experiments on synthetic and real world dataset to pinpoint the factors which influence the probability prediction error.
Subject Area
Computer science|Information science
Recommended Citation
Pan, Yu, "3D Tracking and Analysis on Multivariate Time-Varying Scientific Data" (2021). ETD collection for University of Nebraska-Lincoln. AAI28715963.
https://digitalcommons.unl.edu/dissertations/AAI28715963