Abstract
Generative machine learning models have revolutionized material discovery bycapturing complex structure-property relationships, yet extending theseapproaches to the inverse design of three-dimensional metamaterials remainslimited by computational complexity and underexplored design spaces due to thelack of expressive representations. Here, we present DiffuMeta, a generativeframework integrating diffusion transformers with a novel algebraic languagerepresentation, encoding 3D geometries as mathematical sentences. This compact,unified parameterization spans diverse topologies while enabling directapplication of transformers to structural design. DiffuMeta leverages diffusionmodels to generate novel shell structures with precisely targeted stress-strainresponses under large deformations, accounting for buckling and contact whileaddressing the inherent one-to-many mapping by producing diverse solutions.Uniquely, our approach enables simultaneous control over multiple mechanicalobjectives, including linear and nonlinear responses beyond training domains.Experimental validation of fabricated structures further confirms the efficacyof our approach for accelerated design of metamaterials and structures withtailored properties.