Uncertainty-Aware Deep Neural Representations for Visual Analysis of Vector Field Data.

Area of Science:

• Scientific Visualization
• Machine Learning
• Computational Science

Background:

• Deep Neural Networks (DNNs) are increasingly used in scientific visualization.
• DNNs lack inherent mechanisms for quantifying prediction uncertainty.
• Understanding DNN uncertainty is vital for informed decision-making in scientific applications.

Purpose of the Study:

• To develop uncertainty-aware implicit neural representations for modeling steady-state vector fields.
• To evaluate Deep Ensemble and Monte Carlo Dropout for uncertainty estimation in vector field visualization.
• To enhance the interpretability and resilience of DNNs for analyzing complex vector field data.

Main Methods:

• Developed uncertainty-aware implicit neural representations.
• Applied Deep Ensemble and Monte Carlo Dropout techniques for uncertainty estimation.
• Evaluated models on several steady vector field datasets.

Main Results:

• Uncertainty-aware models provide informative visualizations of vector field features.
• Incorporating prediction uncertainty improves DNN model resilience.
• The developed models enhance the interpretability of DNNs for vector field analysis.

Conclusions:

• Uncertainty-aware DNNs are effective for modeling and visualizing steady-state vector fields.
• The proposed methods improve the reliability and applicability of DNNs in scientific visualization.
• This work enables more robust and interpretable analysis of non-trivial vector field datasets.