As particle-based simulations grow in scale, efficient compression of scientific data becomes critical. Neural point cloud compression (PCC) methods excel on computer vision (CV) benchmarks, but do they transfer to scientific particle data? We highlight that scientific particles differ from CV point clouds in fractal dimension: CV point clouds are locally planar, but particles are not. This key difference limits the effectiveness of the octree data structure that underpins most PCC methods. Neural codecs achieve only marginal bitrate reductions over the traditional G-PCC baseline on scientific data, revealing a significant performance gap. Supported by an evaluation spanning domains of cosmology, molecular dynamics, and plasma physics, we outline where the structure of scientific data limits current neural approaches and where learning can help in the future.
Figures & tables
Figure 1: Fractal dimension reveals a geometric difference between CV and scientific point clouds. (a) Human body scan. The dimension remains near 2 at all levels, consistent with locally planar surface geometry. (b) Vehicular LiDAR. Fractal dimension increases from near 0 at fine levels to approximately 2 at coarse levels. This reflects the physical nature [ 16 ] of spinning LiDAR sensors: scan rings appear locally linear at finer levels and appear denser at coarser levels. (c) Cosmology N -body simulations. Fractal dimension rises quickly from 0 to 3, indicating a sharp transition from sparse to volume-filling geometry. (d) Molecular dynamics simulations show the same sharp sparse-to-dense transition as cosmology simulations. This contrast illustrates the fundamental geometric difference between CV point clouds and scientific particle data discussed in Sec. 2.1 .
Dataset (# test samples)
Post- Quant.
Raw Octree
G-PCC
RENO
Oct- Attention
SZ3
LCP
Scientific
Quijote halos (200)
54
87.40 (0.6)
35.92 (1.5)
35.55 (1.5)
35.78 (1.5)
60.18 (0.9)
36.22 (1.5)
Quijote particles (200)
54
88.37 (0.6)
36.03 (1.5)
35.77 (1.5)
35.92 (1.5)
59.46 (0.9)
36.25 (1.5)
Quijote boxes (1600)
45
53.89 (0.8)
23.52 (1.9)
22.68 (2.0)
22.82 (2.0)
40.36 (1.1)
24.52 (1.8)
YiiP (900)
45
59.70 (0.8)
25.35 (1.8)
22.89 (2.0)
23.91 (1.9)
44.70 (1.0)
27.08 (1.7)
WarpX (30)
54
55.84 (1.0)
23.74 (2.3)
23.35 (2.3)
23.33 (2.3)
41.25 (1.3)
24.70 (2.2)
CV
MVUB, Phil (245)
30
2.55 (11.8)
1.07 (28.0)
–
0.79 (38.0)
–
–
Table 1: Neural methods achieve far smaller gains on scientific particle data than on CV data. Bit-per-point (bpp) by method, with compression ratio in parentheses, computed over post-quantization bitrate. General-purpose scientific compressor SZ3 suffers on particle data, while particle compressor LCP performs on par with neural CV point cloud codecs on scientific data. “–” marks entries where the method was not evaluated on the dataset originally.
Figure 2: CV point cloud compression benefits across levels from conditioning on a single parent node, while particle datasets only benefit from it at coarser levels. (a) On MVUB (human body scans), points concentrate on surfaces, so a child’s occupancy can be strongly predicted by its parent’s: conditional entropy drops significantly across levels. (b) Quijote halo data benefits from the same model only on coarser levels. Gains in these levels are less significant, because the number of voxels to code here is negligible compared to the entire octree. Gray area indicates levels that contribute less than 1% of the total bitstream – gains in these levels have little effect on overall bpp.
Lossy compression is widely used to reduce storage and I/O costs for large-scale particle datasets in scientific applications such as cosmology, molecular dynamics, and fluid dynamics, where clustering structures (e.g., single-linkage or Friends-of-Friends) are critical for downstream analysis; however, existing compressors typically provide only pointwise error bounds on particle positions and offer no guarantees on preserving clustering outcomes, and even small perturbations can alter cluster connectivity and compromise scientific validity. We propose a correction-based technique to preserve single-linkage clustering under lossy compression, operating on decompressed data from off-the-shelf compressors such as SZ3 and Draco. Our key contributions are threefold: (1) a clustering-aware correction algorithm that identifies vulnerable particle pairs via spatial partitioning and local neighborhood search; (2) an optimization-based formulation that enforces clustering consistency using projected gradient descent with a loss that encodes pairwise distance violations; and (3) a scalable GPU-accelerated and distributed implementation for large-scale datasets. Experiments on cosmology and molecular dynamics datasets show that our method effectively preserves clustering results while maintaining competitive compression performance compared with SZ3, ZFP, Draco, LCP, and space-filling-curve-based schemes.
Congrong Ren, Sheng Di, Katrin Heitmann +2
The Ohio State University · Argonne National Laboratory
Plenoptic point clouds (PPC) are novel data structures that represent the light from different viewing directions in order to provide a higher degree of realism to regular point clouds. This is achieved by associating each point to multiple colors instead of a single one. Here, we present a method to efficiently compress the attributes of a PPC, consisting of a Karhunen-Loève transform over the color attributes followed by multiple attribute coders with intra prediction capability. This compression scheme can be incorporated within the MPEG's geometry-based PCC (G-PCC) standard, using any of G-PCC's existing solutions for attribute coding. Compression performance assessment using PPCs of different spatial resolutions reveals competitive results in comparison to existing methods, such as RAHT-based or video-based PCC solutions. We believe our coder to be the new state of the art.
Davi R. Freitas, Gustavo L. Sandri, Ricardo L. de Queiroz
Inria - Rennes Bretagne-Atlantique Rennes, France · Instituto Federal de Bras´ılia Bras´ılia, Brazil · Universidade de Bras´ılia Bras´ılia, Brazil
We introduce Neural Particle Automata (NPA), a Lagrangian generalization of Neural Cellular Automata (NCA) from static lattices to dynamic particle systems. Unlike classical Eulerian NCA where cells are pinned to pixels or voxels, NPA model each cell as a particle with a continuous position and internal state, both updated by a shared, learnable neural rule. This particle-based formulation yields clear individuation of cells, allows heterogeneous dynamics, and concentrates computation only on regions where activity is present. At the same time, particle systems pose challenges: neighborhoods are dynamic, and a naive implementation of local interactions scale quadratically with the number of particles. We address these challenges by replacing grid-based neighborhood perception with differentiable Smoothed Particle Hydrodynamics (SPH) operators backed by memory-efficient, CUDA-accelerated kernels, enabling scalable end-to-end training. Across tasks including morphogenesis, point-cloud classification, and particle-based texture synthesis, we show that NPA retain key NCA behaviors such as robustness and self-regeneration, while enabling new behaviors specific to particle systems. Together, these results position NPA as a compact neural model for learning self-organizing particle dynamics.
Hyunsoo Kim, Ehsan Pajouheshgar, Sabine Süsstrunk +2