74% efficiency in creating new metamaterials. ETH and others use algebraic language models to reduce dimensionality and design 3D structures

ETH Zurich and Delft University of Technology published the DiffuMeta framework in "Nature - Machine Intelligence", using generative AI to achieve 3D metamaterial reverse design. Traditional topology optimization requires a large amount of simulations and is difficult to handle nonlinear thin shell structures, while voxel AI generation costs are high and the surfaces are rough. DiffuMeta converts complex 3D geometries into algebraic language sequences, allowing the diffusion Transformer to generate structures like text. Just input the target stress-strain curve, and the corresponding unit cell can be automatically output, controlling multiple mechanical properties at the same time. In the test, 100% of the generated structures were new topologies, 74% met the requirements (only 3.2% of random combinations), and the nonlinear target error was less than 4.6%. 3D printing examples have verified its ability to accelerate the development of customized functional materials, bringing a more accurate, faster and more cost-effective new path to mechanical metamaterial design.