wzhan24/ReGDiff_MetaTruss_10k

ReGDiff MetaTruss 10k

10,000 truss-lattice metamaterial unit cells, each voxelized as a binary 48 × 48 × 48 occupancy grid. This dataset was used to train and evaluate ReGDiff (arXiv:2609.34231).

Samples	10,000
Resolution	48 × 48 × 48
Values	binary (0 = void, 1 = solid)
Solid fraction	mean 17.2%, range 2.9% – 49.4%
File size	17.8 MB (bit-packed, compressed)

File format

ReGDiff_MetaTruss_10k.npz contains two arrays:

• packed: uint8 array of shape (10000, 13824). Each row is one flattened 48³ cube packed 8 voxels per byte with np.packbits.
• shape: the cube shape, (48, 48, 48).

The array axes are ordered (x, y, z).

Loading

import numpy as np
from huggingface_hub import hf_hub_download

path = hf_hub_download("wzhan24/ReGDiff_MetaTruss_10k", "ReGDiff_MetaTruss_10k.npz", repo_type="dataset")
with np.load(path) as f:
    packed, shape = f["packed"], tuple(f["shape"])
cubes = np.unpackbits(packed, axis=1, count=int(np.prod(shape))).reshape(-1, *shape)
# cubes: uint8, shape (10000, 48, 48, 48), values in {0, 1}

How the voxels were generated

Each sample starts as a truss lattice defined by node coordinates and strut connectivity inside the unit cube. The cube is divided into a 48³ grid. A voxel is set to solid when its center is within 0.06 of any strut, where the distance is measured to the strut segment (or to the nearer end node, if the center lies beyond the segment's ends).

Citation

If you use this dataset, please cite:

@article{zhan2026regdiff,
  title   = {ReGDiff: Guided Diffusion in Regulated Latent Space for Exploring Metamaterial Voxel Geometry},
  author  = {Zhan, W. and Chen, J. and Fu, D. and Zhou, D.},
  journal = {arXiv preprint arXiv:2609.34231},
  year    = {2026}
}

License

Released under CC BY 4.0. You may share and adapt the data with attribution.