Denoising diffusion algorithm for inverse design of microstructures with fine-tuned nonlinear material properties
Vlassis, Sun, CMAME, 2023.
We propose a denoising diffusion algorithm for discovering microstructures with fine-tuned nonlinear properties. These generative models utilize diffusion dynamics for denoising images and generating synthetic samples. By learning the reverse of a Markov diffusion process, the algorithm manipulates microstructures' topology, generating numerous prototypes with constitutive responses closely aligned to defined nonlinear responses. A convolutional neural network surrogate is trained to identify the subset of microstructures with precise properties, replacing high-fidelity finite element simulations. Our study confirms the algorithm's ability to create microstructures with specific nonlinear material properties within the latent space of the training data. Furthermore, the algorithm can be extended to incorporate additional topological and geometric modifications. Validated against the open-source mechanical MNIST data set, the algorithm performs successful inverse design of nonlinear effective media and quantitatively maps the nonlinear structure-property relationship.
Read more →
Read more →