Source-linked AI summary
3D Gaussian Splatting for Scientific Particle Data Compression and Rendering
Bo Jiang, Youyuan Liu, Taolue Yang, Sheng Di, Sian Jin
TL;DR
Large particle simulations are costly to store, transfer, and explore, while conventional compressors target data-space error rather than rendered-image fidelity. ParticleGS learns a compact 3D Gaussian representation that adapts to visualization settings; its 8-block HACC model reaches 30.03 dB PSNR at 65× compression and renders at 662 FPS.
Problem
Large particle datasets strain storage, transfer, and interactive visualization, while data-space compression does not ensure visual fidelity.
Source: §I · pp. 1–2Method
ParticleGS optimizes a compact 3D Gaussian representation for rendered images, using multi-orbit training, VizMapper adaptation, and spatial blocks with global fine-tuning.
Source: §I · p. 2, §III · p. 4, B. VizMapper: Visualization Parameter… · p. 4Results
30.03 dB PSNR at 65× compression for the 8-block HACC model, outperforming SZ3 by 5–8 dB at matched ratios; rendering reaches 662 FPS.
Source: §I · p. 2Takeaways & Limitations
The representation generalizes without tuning to additional HACC regions and a dark-matter-only FIRE-2 simulation.
Source: §I · p. 2Fidelity is highest within the trained camera-orbit and radius/opacity ranges; quality degrades outside that viewing envelope, so arbitrary interior fly-throughs are not fully covered.
Source: §V · p. 11
Abstract
from arXiv · showhide
Large-scale particle simulations produce hundreds of millions of particles, straining storage, transfer, and interactive visualization. Existing lossy compressors such as SZ3 operate in data space and provide no guarantees on downstream visualization fidelity. We propose ParticleGS, a visualization-aware framework based on 3D Gaussian Splatting (3DGS) that learns a compact representation directly optimized for rendered image quality, combining (1) a multi-stage, multi-orbit training pipeline, (2) VizMapper, a lightweight network that adapts a single trained model to user-specified visualization parameters at inference time, and (3) spatial block training with KD-tree decomposition and global fine-tuning. On a 281-million-particle HACC cosmological simulation, our 8-block model reaches 30.03 dB PSNR at 65x compression, outperforming SZ3 by 5-8 dB at comparable ratios, and generalizes without tuning to additional HACC regions and a dark-matter-only FIRE-2 simulation. It renders at 662 FPS on a single GPU, over 2,300x faster than ParaView on the full particle data.
I. INTRODUCTION
Particle simulations are difficult to store, transfer, and explore interactively at scale, while data-space compression does not ensure faithful visualizations. ParticleGS adapts 3D Gaussian Splatting to particle data and reports higher visual fidelity than SZ3 at comparable compression ratios.
I. INTRODUCTION
- ParticleGS targets rendered-image fidelity, whereas data-space compressors control coordinate error without ensuring visual quality.SZ3 and ZFP can introduce axis-aligned streaking, while LCP's spatially coherent quantization artifacts also reduce visual fidelity at comparable size.Source: §I · pp. 1–2
- The approach addresses the need to vary particle radius and opacity and to represent fine detail across multiple viewing scales and large datasets.A single model otherwise lacks capacity for fine-grained structures throughout the volume, and single-distance training does not resolve features at other scales.Source: §I · p. 2
- ParticleGS combines multi-orbit, progressive-resolution training with VizMapper, which adapts a trained model to visualization settings without retraining.Spatial block training uses KD-tree decomposition, parallel per-block models, and global fine-tuning; it improves quality by up to 1.28 dB over the single-block baseline.Source: §I · p. 2
- 30.03 dB PSNR at 65× compression for the 8-block HACC model, outperforming SZ3 by 5–8 dB at matched ratios.The single-block model reaches 28.80 dB PSNR at 290× compression; the pipeline also generalizes to additional HACC regions and a FIRE-2 simulation.Source: §I · p. 2
II. BACKGROUND
Particle datasets are unstructured point clouds whose visualization depends on user-controlled radius and opacity. Conventional lossy compressors primarily target data-space error, which does not directly ensure rendered-image fidelity.
A. Particle Scientific Data
- Particle positions form an unstructured point cloud, so techniques designed for structured grids do not directly apply.Particle visualization parameters such as radius and opacity reveal different structural scales, making support for varying settings important.Source: A. Particle Scientific Data · pp. 2–3
B. Lossy Compression for Particle Data
- Conventional compressors bound or reduce data-space error, but those metrics do not directly optimize rendered-image quality.SZ3, ZFP, and related compressors often process particle coordinates as independent one-dimensional x, y, and z arrays; particle-specific LCP reorders particles to exploit spatial locality but can produce structured artifacts through coherent quantization errors.Source: B. Lossy Compression for Particle Data · p. 3
C. 3D Gaussian Splatting
3D Gaussian Splatting represents scenes with differentiable, renderable Gaussians, but standard models assume fixed appearance. ParticleGS formulates particle rendering as image matching and adds VizMapper to adapt Gaussian scale and opacity to visualization settings.
C. 3D Gaussian Splatting
- 3DGS represents a scene as anisotropic Gaussians that are rasterized and alpha-composited, with differentiable image losses optimizing their parameters.Its explicit representation supports real-time rendering, but existing methods target fixed-appearance natural scenes.Source: C. 3D Gaussian Splatting · p. 3
A. Problem Formulation
- ParticleGS seeks a compact Gaussian representation whose rendered images match ParaView across viewpoints and visualization settings.The representation contains far fewer Gaussians than particles, and its rendering objective is defined over camera pose, radius, and opacity.Source: §III · p. 4, A. Problem Formulation · p. 4
B. VizMapper: Visualization Parameter Adaptation
- VizMapper modulates each Gaussian’s scale and opacity from user-specified visualization parameters, allowing adaptation without retraining.It is a 3-layer MLP with 4,610 parameters; its inputs include relative radius and opacity changes and per-Gaussian scale and opacity attributes.Source: B. VizMapper: Visualization Parameter… · p. 4
- Scale and opacity adjustments combine each Gaussian’s base value, a global visualization factor, and a learned per-Gaussian correction.Opacity correction is applied to the stored logit before the rendering sigmoid, with a lower bound on the multiplicative opacity factor.Source: B. VizMapper: Visualization Parameter… · p. 4
C. Training Pipeline
The training pipeline normalizes the particle volume and trains a 3DGS model across varied visualization settings, camera orbits, and progressively higher resolutions. It uses a content-masked L1 objective and omits components that increased model size without improving masked PSNR.
C. Training Pipeline
- Training uses full-particle renderings as supervision; the approximately 200,000-particle subsample initializes optimization, while densification expands the model to 171k or 769k Gaussians.The subsample is not used as the sole target for reproducing the visualization.Source: C. Training Pipeline · p. 5
- Particle radius and opacity are sampled independently from Beta distributions centered on defaults, concentrating training near common settings while covering extremes.The sampled ranges are r ∈ [0.0025, 0.0175] and α ∈ [0.0125, 0.0875].Source: C. Training Pipeline · p. 5
- Three camera-orbit families at far, mid, and near distances support both overviews and close-ups during training.Their radii are 1.0×, 0.7×, and 0.5× the base radius.Source: C. Training Pipeline · p. 5, Figure 4 · p. 5
- Training progresses from 1920×1080 for 12,000 iterations to 5760×3240 for 27,000 iterations, with densification in both stages.The second stage adds 20% internal cameras during its final 12,000 iterations to improve robustness for internal close-up views without materially affecting external-orbit metrics.Source: C. Training Pipeline · pp. 5–6
- The objective is pure L1 loss over a content mask that excludes the black background; adding DSSIM increased model size by 68% without improving masked PSNR.Degree-0 spherical harmonics also save 59% versus degree 2 because particle rendering is view-independent.Source: C. Training Pipeline · p. 6
D. Spatial Block Training with Global Fine-Tuning
For volumes too large for one model, ParticleGS trains spatial blocks independently, merges their Gaussians in a shared coordinate frame, and fine-tunes globally to address seams and recalibrate VizMapper.
D. Spatial Block Training with Global Fine-Tuning
- KD-tree decomposition recursively bisects the axis of greatest extent into spatially compact blocks with roughly equal particle counts, balancing per-block training cost.Blocks use camera orbits centered on their local bounding boxes and can be trained in parallel across multiple GPUs.Source: D. Spatial Block Training with Global… · p. 6
- Each block’s Gaussians are transformed from its local frame into the global frame and merged into one PLY model.Positions and log-scales are adjusted according to the normalization-scale ratio and coordinate-frame centers.Source: D. Spatial Block Training with Global… · p. 6
- The merged model is fine-tuned end-to-end on the original multi-orbit training data to repair block-boundary seams and recalibrate VizMapper for the more heterogeneous Gaussian population.Fine-tuning uses reduced resolution because its goal is seam repair and recalibration, not recovering additional high-frequency detail.Source: D. Spatial Block Training with Global… · p. 6
- Global fine-tuning runs for 60,000 iterations, with mild densification enabled during the first 30,000.The correction ranges are widened to δmax_s = 0.3 and ϵ = 0.1.Source: D. Spatial Block Training with Global… · p. 6
E. Particle Recovery via GMM Sampling
ParticleGS reinterprets its optimized Gaussians as a Gaussian mixture model and samples from it to recover particle positions. The recovered particles preserve large-scale density statistics but not local structure.
E. Particle Recovery via GMM Sampling
- The learned Gaussians also support particle-position recovery as a byproduct of optimizing the representation for rendering fidelity.Source: E. Particle Recovery via GMM Sampling · p. 6
- The optimized Gaussians can be reinterpreted as a Gaussian Mixture Model, with each component’s mixing weight proportional to its opacity.Particle positions are recovered by sampling component indices and then sampling positions from the selected Gaussian.Source: E. Particle Recovery via GMM Sampling · p. 6
- Recovered positions are mapped from the normalized [−1, 1]3 frame back to the original coordinates.Source: E. Particle Recovery via GMM Sampling · p. 6
- Recovering 281M particles takes approximately 74.5 seconds on one GPU without additional training, preserving large-scale statistics but not local structure.The preserved statistics include the power spectrum and density PDF; recovered particles are a statistical proxy.Source: E. Particle Recovery via GMM Sampling · p. 6
IV. EXPERIMENTS
The experiments evaluate design choices, scalability, generalization, compression quality, rendering performance, and preservation of spatial statistics.
IV. EXPERIMENTS
- Ablation studies isolate the effects of individual design choices.Source: §IV · p. 6
- Block-training analysis and generalization experiments assess scalability and transfer to other data.Source: §IV · p. 6
- Rate-distortion comparisons with SZ3 and LCP, pipeline profiling, and spatial-statistics evaluation test compression, rendering performance, and distribution preservation.Source: §IV · p. 6