<![CDATA[Various scientific applications, including cosmological simulation, fluid simulation, and molecular dynamics, depend heavily on the analysis of particle data. Although there are techniques for feature extraction and tracking regarding volumetric data, it is more difficult to do such tasks for particle data due to the lack of explicit connectivity information. Even though one could transform the particle data to volume beforehand, doing so runs the risk of incurring error and growing the data size. In order to facilitate feature extraction and tracking for scientific particle data, we adopt a deep learning (DL) method in this research. In order to capture the relation between physical features and spatial locations in a neighborhood, we use a DL model that generates latent vectors. Through clustering the latent vectors, characteristics could be retrieved from the vectors. The Cam-shift tracking algorithm, which just needs inference of the latent vector for chosen regions of interest, is implemented in the feature space to accomplish quick feature tracking. With the use of two datasets, we test our approach and contrast it with other approaches already in use.]]>
