MicroGen3D Dataset: 3D Multiphase Heterogeneous Microstructures for Conditional Latent Diffusion Models

MicroGen3D is a dataset of 3D voxelized microstructures designed for training, evaluation, and benchmarking of generative models—especially Conditional Latent Diffusion Models (LDMs). It includes both synthetic (Cahn–Hilliard simulations) and experimental microstructures with multiple phases (two to three). The voxel grids range from 64×64×64 up to 128×128×64. The dataset contains three categories of microstructures: Experimental microstructures (64×64×64), voxelized from real-world samples. 2-phase Cahn–Hilliard microstructures (128×128×64), generated from thresholded simulations. 3-phase Cahn–Hilliard microstructures (128×128×64), generated from thresholded simulations. For each dataset type, pretrained model weights are provided: vae.pt – Variational Autoencoder fp.pt – Feature Predictor ddpm.pt – Denoising Diffusion Probabilistic Model In addition to the complete datasets, smaller sample subsets are included for testing and demonstration purposes.The dataset is mirrored on Hugging Face, and the accompanying source code is hosted on GitHub. Preprint can be found here.