InSituNet: Deep Image Synthesis for Parameter Space Exploration of Ensemble Simulations

Area of Science:

• Scientific Visualization
• Deep Learning
• High-Performance Computing

Background:

• In situ visualization is crucial for large-scale simulations due to I/O and storage constraints.
• Current in situ methods limit post-hoc analysis flexibility as raw data is unavailable.
• Existing image-based approaches do not support simulation parameter exploration.

Purpose of the Study:

• To introduce InSituNet, a deep learning surrogate model for parameter space exploration in in situ visualization.
• To enable flexible exploration of simulation and visualization parameters for ensemble simulations.
• To facilitate in-depth analysis of large-scale simulations by generating new visualizations on demand.

Main Methods:

• Developed InSituNet, a convolutional regression model.
• Trained the model to learn the mapping from simulation/visualization parameters to visualization results.
• Applied InSituNet to combustion, cosmology, and ocean simulations.

Main Results:

• InSituNet effectively generates new images for varied simulation parameters and visualization settings.
• Demonstrated the model's capability for flexible parameter space exploration.
• Validated the approach through quantitative and qualitative evaluations across diverse simulation domains.

Conclusions:

• InSituNet provides a powerful solution for exploring parameter spaces in in situ visualized simulations.
• The deep learning approach overcomes limitations of traditional in situ visualization methods.
• InSituNet enhances the analytical capabilities for large-scale ensemble simulations.