VDL-Surrogate: A View-Dependent Latent-based Model for Parameter Space Exploration of Ensemble Simulations

Area of Science:

• Computational Science
• Data Visualization
• Scientific Computing

Background:

• Surrogate models accelerate parameter space exploration of ensemble simulations by enabling previews without costly computations.
• Existing surrogate models often lack sufficient resolution for detailed visualization and analysis due to computational limitations.

Purpose of the Study:

• To introduce VDL-Surrogate, a view-dependent neural-network-latent-based surrogate model.
• To enable high-resolution visualizations and user-specified visual mappings for parameter space exploration.
• To improve the efficiency of computational resource usage in simulation previews.

Main Methods:

• VDL-Surrogate employs ray casting from multiple viewpoints to collect samples and generate compact latent representations.
• Latent encoding reduces surrogate model training costs while preserving output quality.
• Models are trained for selected viewpoints covering the entire viewing sphere; inference involves predicting and decoding latent representations, followed by interpolation for visualization.

Main Results:

• Demonstrated effectiveness and efficiency of VDL-Surrogate in cosmological and ocean simulations.
• Achieved high-resolution visualizations and user-specified visual mappings.
• Validated through quantitative and qualitative evaluations.

Conclusions:

• VDL-Surrogate significantly enhances parameter space exploration by providing high-resolution, view-dependent previews.
• The model offers an efficient solution for visualizing complex simulation data, overcoming limitations of existing methods.
• Publicly available source code facilitates adoption and further research.