Synthetic PBX Microstructures Generated via Deep Generative Graphs and Conditional Latent Diffusion

Synthetic PBX Microstructures Generated via Deep Generative Graphs and Conditional Latent Diffusion

Description

This dataset contains synthetic three-dimensional polymer-bonded explosive (PBX) microstructures generated using the hierarchical generative framework described in:

Poliner, J., Sun, W., Alshibli, K. A., & Regueiro, R. A. (2026). Deep Generative Graphs for Synthesizing Microstructure of Topology-Preserved Polymer-Bonded Explosives. Engineering with Computers.

The generative pipeline decomposes the synthesis task into two coupled stages:

1. Topological generation via a modified autoregressive GraphRNN (MicrostructureRNN), producing node/edge-weighted grain networks.

2. Geometric generation via conditional latent denoising diffusion, synthesizing individual grain geometries conditioned on oriented bounding boxes (OBB), effective radius, and curvature descriptors.

Generated point clouds are reconstructed using Poisson surface reconstruction and remeshed into watertight crystal grain geometries suitable for mesoscale numerical simulation.

This dataset provides statistically validated synthetic PBX assemblies preserving both geometric and topological measures relative to the reference micro-CT IDOX dataset described in the associated manuscript.

Dataset Contents

The dataset contains 100 generated subgraph clusters. Each subgraph represents a spatially assembled PBX subdomain. The directory structure is organized as follows:

generated_subgraphs/
├── subgraph_000/
│   ├── grain_000.stl
│   ├── grain_001.stl
│   ├── ...
│   └── assembly.stl
...
├── subgraph_099/

Each subgraph_xxx directory contains:

• Individual grain geometries in STL format (one STL file per grain)

• An assembled subgraph STL representation composed of all grains

• Spatially aligned grain configurations

All geometries are watertight triangle meshes reconstructed from latent diffusion–generated point clouds.

Statistical Fidelity

The synthetic microstructures reproduce the following statistical measures reported in the manuscript:

• Grain size distribution (effective radius)

• Oriented bounding box (OBB) dimension and orientation distributions

• Surface mean curvature distribution

• Orientation tensor metrics: compactness, flakiness, and elongation

• Degree distribution

• Clustering coefficient

• Graph density

• Inter-grain spacing statistics

• Polymer-to-grain phase ratio

Minor compression of distribution tails may be observed due to latent-space regularization effects.

File Format

All grain geometries are provided in STL (stereolithography) format.

• Units are consistent across all subgraphs.

• Each grain mesh is watertight.

• Assemblies are spatially aligned and directly usable for finite element preprocessing, voxelization, mesoscale mechanical simulation, and parametric studies.

Intended Use

This dataset is intended for mesoscale mechanical simulation studies, synthetic database augmentation, sensitivity analysis of microstructural descriptors, surrogate model development, and shock and ignition modeling in PBX systems.

The dataset does not contain experimental micro-CT data. All geometries are synthetic and generated using trained generative models.

Size

Total compressed dataset size: approximately 2 GB.

Related Resources

• GitHub repository (code to be released upon final manuscript acceptance): [https://github.com/jsp2195/microstructure-rnn-latent-diffusion-pbx]

• Associated manuscript: Deep Generative Graphs for Synthesizing Microstructure of Topology-Preserved Polymer-Bonded Explosives (under revision at Engineering with Computers).

Suggested Citation

If this dataset is used, please cite:

Poliner, J., Sun, W., Alshibli, K. A., & Regueiro, R. A. (2026). Deep Generative Graphs for Synthesizing Microstructure of Topology-Preserved Polymer-Bonded Explosives. Engineering with Computers.