Generating 3D images of material microstructures from a single 2D image: a denoising diffusion approach.

Area of Science:

• Materials Science
• Computational Imaging
• Artificial Intelligence

Background:

• Three-dimensional (3D) imaging is crucial for material microstructure analysis and numerical simulations.
• Existing 3D imaging methods are often expensive and limited by resolution.
• Generating 3D microstructures from 2D images is challenging due to data limitations.

Purpose of the Study:

• To introduce a novel, cost-effective method for generating large-scale 3D material microstructures from single 2D images.
• To overcome the limitations of current 3D imaging techniques in terms of cost and resolution.
• To enable advanced material characterization and analysis through high-resolution 3D image generation.

Main Methods:

• Combines a denoising diffusion probabilistic model with a generative adversarial network (GAN).
• Employs chain sampling using 3D intermediate outputs from the diffusion process reversal.
• Utilizes a 2D discriminator to guide the training with limited 3D data.

Main Results:

• Generated 3D images accurately capture geometric and statistical properties of 2D inputs.
• Achieved up to a three-fold improvement in Frechet inception distance score compared to SliceGAN.
• Demonstrated enhanced accuracy in derived material properties over the state-of-the-art method.

Conclusions:

• The novel method provides a high-resolution, statistically representative alternative for 3D microstructure imaging.
• This approach significantly advances material characterization and numerical simulation capabilities.
• It offers a practical solution for generating 3D material data where traditional methods are prohibitive.