Denoising diffusion probabilistic model enabled time-continuous, inheritance-preserving microstructure evolution of virtual materials

Abstract This study employs the denoising diffusion probabilistic model (DDPM), a text-to-image generative approach, to predict microstructure evolution. The model starts with a noise image generated from textual descriptions, which is progressively denoised to reveal the microstructure. The model parameters include the training intervals, algorithm selection, image resolution, optimizer settings, encoder configuration, prompts, concept script, and generator settings. The effectiveness of DDPM is demonstrated through three cases: particle coarsening, grain growth, and dendrite solidification. These examples highlight the ability of DDPM to simulate material evolution across different scales, from individual particles to multi-grain interactions and processing parameter effects. The generated micrographs closely match the original images, with prediction errors below 4% for particle and grain size, and about 9% for secondary dendrite arm spacing. The study also addresses underfitting and overfitting and shows that the DDPM predicts interrelated microstructures across successive time steps instead of random samples. This highlights the potential of model for capturing the dynamic nature of microstructural evolution.