Reproduce “Generating 3D structures from a 2D slice with GAN-based dimensionality expansion” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Steve Kench, Samuel J. Cooper
Researchers at Imperial College London developed SliceGAN, a Generative Adversarial Network that synthesizes high-fidelity 3D material microstructures from a single 2D image, achieving 10^5 times faster generation of 10^8 voxel volumes compared to prior methods while maintaining statistical accuracy for emergent 3D properties.
Key insights
- Training a 3D generator against a 2D discriminator by extracting slices from the generated 3D volume is an effective method for 2D-to-3D dimensionality expansion.
- Careful design of transpose convolution parameters within the generator is crucial for generating consistently high-quality 3D volumes without spatial artifacts.
- Statistically accurate 3D microstructures, including emergent 3D properties, can be learned and reconstructed solely from representative 2D cross-sections.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.