Trustworthy Deep Learning-Assisted Visualization and Analysis for Distribution-Based Ensemble Scientific Data...

Area of Science:

• Computational Science
• Data Science
• Scientific Visualization

Background:

• Ensemble datasets from scientific simulations are crucial for studying complex phenomena.
• Large datasets pose challenges for traditional data analysis due to I/O and storage limitations.
• Distribution-based representations offer a compact alternative but sacrifice spatial information.

Purpose of the Study:

• To develop a method for reconstructing high-fidelity data volumes from distribution-based representations.
• To address the loss of spatial information in traditional distribution-based methods.
• To enhance downstream data analysis and uncertainty quantification.

Main Methods:

• Introduced a deep learning model utilizing the Sinkhorn operator and Gumbel trick.
• The model learns to map samples from a distribution to spatial locations within a data block.
• Reconstruction is performed directly from the distribution representation, not a direct prediction.

Main Results:

• The deep learning model successfully reconstructs data volumes from distribution representations.
• The reconstructed data preserves spatial information, improving analysis precision.
• The method supports high-quality downstream analysis, visualization, and point-wise uncertainty quantification.

Conclusions:

• The proposed deep learning approach effectively reconstructs data volumes while preserving spatial fidelity.
• This method alleviates I/O and storage constraints for large scientific datasets.
• Enables advanced data analysis, visualization, and uncertainty quantification for complex simulations.