Explorable INR: An Implicit Neural Representation for Ensemble Simulation Enabling Efficient Spatial and Parameter...

Area of Science:

• Computational science
• Scientific simulation

Background:

• High-resolution ensemble simulations in fields like cosmology and oceanology face significant storage and computational challenges.
• Existing surrogate models require full field reconstruction for parameter exploration, limiting efficiency and flexibility.
• Current methods struggle with capturing physical attribute distributions and identifying optimal parameters.

Purpose of the Study:

• To introduce Explorable INR, a novel implicit neural representation-based surrogate model.
• To enable efficient point-based spatial queries and parameter space exploration without full field reconstruction.
• To address computational bottlenecks in spatial exploration and facilitate ensemble analysis.

Main Methods:

• Developed Explorable INR, an implicit neural representation-based surrogate model.
• Utilized probabilistic affine forms (PAFs) for uncertainty propagation within Explorable INR.
• Reformulated parameter exploration as optimization tasks using gradient descent and KL divergence minimization.

Main Results:

• Explorable INR facilitates point-based spatial queries, bypassing the need for full field computation.
• Probabilistic affine forms enable efficient statistical summaries for ensemble analysis and visualization.
• The proposed optimization approach ensures scalability for parameter exploration.

Conclusions:

• Explorable INR significantly reduces computation and memory costs for spatial and parameter exploration.
• The model provides effective ensemble analysis capabilities, overcoming limitations of existing methods.
• This approach enhances the efficiency of exploring parameter spaces in large-scale scientific simulations.