SurroFlow: A Flow-Based Surrogate Model for Parameter Space Exploration and Uncertainty Quantification.

Area of Science:

• Scientific Computing
• Machine Learning
• Computational Science

Background:

• Deep learning surrogate models offer efficient data generation but lack uncertainty quantification and parameter exploration capabilities.
• Existing models struggle with reverse prediction and efficient exploration of simulation parameter spaces.

Purpose of the Study:

• Introduce SurroFlow, a normalizing flow-based surrogate model for invertible simulation parameter-output learning.
• Enable accurate predictions, uncertainty quantification, and efficient parameter recommendation.
• Develop a user-guided framework for ensemble simulation exploration and visualization.

Main Methods:

• Developed SurroFlow, a novel normalizing flow-based surrogate model.
• Learned invertible transformations between simulation parameters and outputs.
• Integrated SurroFlow with a genetic algorithm and a visual interface for user-guided exploration.

Main Results:

• SurroFlow provides accurate predictions and quantifies uncertainty in data generation.
• The model enables efficient simulation parameter recommendation and exploration.
• The integrated framework significantly reduces computational costs for scientific surrogate modeling.

Conclusions:

• SurroFlow enhances the reliability and exploration capabilities of scientific surrogate models.
• The framework offers a powerful tool for user-guided ensemble simulation exploration.
• This approach advances deep learning applications in scientific modeling by addressing key limitations.