Guided diffusion for fast inverse design of voxel-based mechanical metamaterials

Mechanical metamaterials are synthetic materials that can possess extraordinary physical characteristics, such as abnormal elasticity, stiffness, and stability, by carefully designing their structure. The representation of metamaterials through high-resolution voxels has the potential to reveal delicate local structures with unique mechanical properties. However, this approach results in a substantial computational burden. To this end, this paper proposes a fast inverse-design framework driven by a self-conditioned diffusion model. An initial set of microstructures is first generated using a physics-based optimization algorithm and used to train the diffusion model. The trained model is then employed to synthesize new candidate microstructures, from which high-quality samples are selected and incorporated into the training set for subsequent retraining. By iterating this generate–filter–retrain cycle, we progressively construct a large-scale dataset and obtain a high-performance generative model. Our model is capable of generating a microstructure to approach the specified homogenized tensor matrix in just 0.42 s on an NVIDIA GeForce RTX 3090 GPU. Compared with the state-of-the-art gradient-based topology optimization method, we achieve an average acceleration of 100 times. Furthermore, we demonstrate that the proposed model enables efficient exploration of extreme metamaterial designs, supports multiscale design workflows, and, through a dedicated guidance mechanism, generates families of metamaterials with continuous variations in both geometry and physical properties. This flexible and adaptive generative tool is of great value in structural engineering or other mechanical systems and can stimulate more subsequent research. • This paper presents a fast deep generative inverse-design method for voxel-based mechanical metamaterials. • We generate initial 128 3 microstructures via optimization and expand the dataset using active learning to achieve broad coverage of physical properties. • This rapid inverse design tool facilitates the exploration of extreme metamaterials, metamaterials sequence, and the generation of diverse microstructures for multiscale design.