Adaptively Placed Multi-Grid Scene Representation Networks for Large-Scale Data Visualization.

Area of Science:

• Scientific data visualization and compression
• Neural rendering and scientific computing

Background:

• Scene representation networks (SRNs) are used for scientific data compression and visualization.
• Current SRNs lack adaptive parameter allocation for complex scientific data, impacting reconstruction quality.

Purpose of the Study:

• To improve the reconstruction quality of SRNs for scientific data.
• To introduce an adaptive SRN architecture and an efficient training technique.
• To provide an open-source tool for neural volume rendering.

Main Methods:

• Developed an adaptively placed multi-grid SRN (APMGSRN) architecture.
• Implemented a domain decomposition training and inference technique for multi-GPU acceleration.
• Created an open-source neural volume rendering application compatible with PyTorch-based SRNs.

Main Results:

• APMGSRN dynamically allocates network resources to areas with high error, improving reconstruction accuracy.
• The domain decomposition method enables parallel training for large-scale data on multi-GPU systems, reducing training time.
• Achieved state-of-the-art reconstruction accuracy without complex octree operations.

Conclusions:

• APMGSRN offers superior reconstruction accuracy for scientific data compared to existing SRNs.
• The proposed training technique accelerates the processing of large datasets.
• The open-source renderer facilitates real-time exploration of scientific data with various rendering parameters.