Microstructure reconstruction using diffusion-based generative models

Microstructure reconstruction has been an essential part of computational material engineering to reveal the relationship between microstructures and material properties. However, finding a general solution for microstructure characterization and reconstruction (MCR) tasks is still challenging, although there have been many attempts such as the descriptor-based MCR methods. To address this generality problem, the denoising diffusion models are first employed for the microstructure reconstruction task in this study. The applicability of the diffusion-based models is validated with several types of microstructures (e.g., polycrystalline alloy, carbonate, ceramics, copolymer, fiber composite, etc.) that have different morphological characteristics. The quality of the generated images is assessed with the quantitative evaluation metrics (FID score, precision, and recall) and the conventional statistical microstructure descriptors. Furthermore, the formulation of implicit probabilistic models (which yields non-Markovian diffusion processes) is adopted to accelerate the sampling process, thereby controlling the computational cost considering the practicability and reliability. The results show that the denoising diffusion models are well applicable to the reconstruction of various types of microstructures with different spatial distributions and morphological features. The diffusion-based approach provides a stable training process with simple implementation for generating visually similar and statistically equivalent microstructures. In these regards, the diffusion model has great potential to be used as a universal microstructure reconstruction method for handling complex microstructures for materials science.