Local Latent Representation Based on Geometric Convolution for Particle Data Feature Exploration.

Area of Science:

• Scientific visualization
• Computational science
• Data analysis

Background:

• Particle data analysis is crucial for scientific simulations like fluid dynamics and cosmology.
• Conventional methods rely on hand-crafted features, which are often application-specific and labor-intensive.
• Transforming particle data into a latent space simplifies feature identification and extraction, but existing neural networks struggle with unstructured particle data.

Purpose of the Study:

• To develop a novel method for creating effective latent representations of scientific particle data.
• To overcome the limitations of grid-based neural networks for processing unstructured particle data.
• To enable robust feature extraction and tracking in particle-based simulations.

Main Methods:

• Utilized Geometric Convolution, a neural network component designed for 3D point clouds.
• Generated latent representations that encode particle positions and physical attributes within their local neighborhoods.
• Applied clustering and mean-shift tracking algorithms in the latent space for feature identification and tracking.

Main Results:

• Successfully created latent representations for scientific particle data from diverse applications.
• Extracted and tracked features from particle data using the proposed latent space approach.
• Demonstrated that the extracted features and tracking results are comparable to established hand-crafted feature methods.

Conclusions:

• Geometric Convolution provides an effective way to generate latent representations for scientific particle data.
• The proposed method offers a generalizable approach for feature extraction and tracking across different simulation types.
• This technique enhances the analysis of complex particle datasets in scientific research.