scottyjb/HPGDiff

HPG-Diff

The dataset for HPG-Diff: Hierarchical Physics-Guided Diffusion with Differentiable Connectivity Constraints for Topology Optimization, published in Applied Soft Computing, 2026.

HPG-Diff uses displacement, principal stress lines, and strain energy density to guide topology generation. This dataset provides the structural examples, conditioning fields, and benchmark reference values used for training and evaluation.

Paper · Code

Datasets

Configuration	Split	Examples	Description
train	train	30,000	Training topologies with conditioning physics fields
test	test_level_1	1,800	In-distribution boundary conditions
test	test_level_2	1,000	Out-of-distribution boundary conditions

The examples use a 64 × 64 grid. Training and evaluation have separate configurations because their fields differ.

Data Format

Field	Shape / type	Description
id	string	Sample identifier
image	image, training only	Target topology
load	64 × 64 × 2	Load field
pf	64 × 64 × 3	Physical-field tensor used by the model
disp	64 × 64 × 1	Displacement field
psl	RGB image	Principal stress line representation
summary	structured record	Loads, supports, coordinates, and volume fraction
bc	64 × 64 × 2, test only	Boundary-condition field
compliance	list of numbers, test only	Reference compliance values used by the evaluation code

Usage

from datasets import load_dataset

train = load_dataset("scottyjb/HPGDiff", "train", split="train")
test_1 = load_dataset("scottyjb/HPGDiff", "test", split="test_level_1")
test_2 = load_dataset("scottyjb/HPGDiff", "test", split="test_level_2")

The training and evaluation scripts are available in the code repository.

Data Source

The SIMP topology examples come from TopoDiff by François Mazé and Faez Ahmed. HPG-Diff supplements them with displacement and principal stress line features.

Citation

@article{yang2026hpgdiff,
  title   = {{HPG-Diff}: Hierarchical physics-guided diffusion with differentiable connectivity constraints for topology optimization},
  author  = {Yang, Jinbo and Yuan, Mingyue and Zhang, Boyuan and Kitamura, Yoshifumi and Jing, Shikai},
  journal = {Applied Soft Computing},
  volume  = {202},
  pages   = {115899},
  year    = {2026},
  doi     = {10.1016/j.asoc.2026.115899}
}

The compliance field in each test row contains the complete reference vector for that split. Evaluation reads it from the first row and matches values by the original row order; keep that order when computing compliance metrics.

License

This dataset is released under CC BY-NC 4.0. Source materials retain their original licenses.