Optimizing Metamaterial Inverse Design with 3D Conditional Diffusion Model and Data Augmentation

Abstract The inverse design of metamaterials is critical for advancing their practical applications. Although deep learning has transformed this process, challenges remain, particularly with insufficient data and less realistic, diverse generation for 3D metamaterials represented as voxels. To address these limitations, a data augmentation technique is developed based on topological perturbation and introduced a 3D conditional diffusion model (3D‐CDM) to optimize 3D metamaterial generation. This original dataset, comprising 200 voxel representations of lattices and triply periodic minimal surfaces, is labeled with effective physical properties computed using homogenization methods. This dataset is expanded to 5000 entries using the proposed data augmentation technique. Training the 3D‐CDM with the augmented dataset significantly improved the quality and accuracy of generated designs. The model successfully produces realistic 3D metamaterials with targeted properties, including volume fraction, Young's modulus, and thermal conductivity, outperforming existing voxel‐based generative models in terms of fidelity and diversity. The 3D‐CDM can be further optimized and extended for the inverse design of a broader range of material microstructures.