M³Sculptor: A Knowledge-Driven Multimodal-Agent Framework for Automated Metamaterial Design

Abstract Metamaterials can exhibit extraordinary properties beyond those of their constituent materials, yet traditional trial-and-error design limits the exploration of vast design spaces. Here, we present M³Sculptor, a multi-agent framework for the automated design of mechanical metamaterials based on natural-language requirements and target properties. The framework integrates large language models and multimodal models to retrieve knowledge and structural images from published literature, reconstruct 3D candidate structures, and identify structures that meet target specifications through symmetry restoration, morphological operations, and computational homogenization. Systematic evaluation demonstrates that M³Sculptor reconstructs architectures suitable for mechanical analysis from structural images in the literature. Symmetry restoration and morphological operations increase the success rate for target property matching within five rounds from 34.13% to 83.25% at an absolute tolerance of 0.01 (GPa for moduli and dimensionless for Poisson's ratio). By combining knowledge retrieval, structure reconstruction, and mechanical evaluation, M³Sculptor achieves intelligent design of metamaterials.