3D Gaussian Splatting for Scientific Particle Data Compression and Rendering

ParticleGS applies 3D Gaussian Splatting to massive particle simulations, aiming at the pain point most teams hit first: storage, transfer, and interactive visualization. Instead of compressing only in data space like traditional lossy schemes, it learns a compact representation optimized for what users actually see on screen.

The pipeline combines multi-stage, multi-orbit training with a VizMapper network that adapts one trained model to different visualization settings at inference time. It also uses spatial block training with KD-tree decomposition plus a global fine-tune pass, which is a practical way to keep training manageable on huge datasets while preserving visual quality.

The headline result comes from a 281-million-particle HACC cosmological simulation: an 8-block model reached 30.03 dB PSNR at 65x compression, beating SZ3 by roughly 5-8 dB at similar ratios. The same setup also generalized to additional HACC regions and a dark-matter-only FIRE-2 simulation without extra tuning.

For engine and tools developers, the interesting part is the rendering side: the system reportedly draws the full particle data at 662 FPS on a single GPU, more than 2,300x faster than ParaView in that scenario. That makes it relevant anywhere teams need fast inspection, remote visualization, or interactive analysis of very large particle sets.