| Abstract Scope |
Recent advances in AI and additive manufacturing have opened new frontiers in designing architected materials with unprecedented mechanical and acoustic functionalities. However, inverse design of 3D truss metamaterials with complex nonlinear behaviors—such as buckling, frictional contact, and wave propagation— remains elusive due to challenges in data efficiency, geometric constraints, and defect tolerance. We present GraphMetaMat, a physics-informed, autoregressive graph-based framework for designing truss metamaterials with programmable stress-strain and vibration responses. Combining graph neural networks, imitation learning, reinforcement learning, and tree search, GraphMetaMat generates defect-robust, manufacturable structures with tailored nonlinear behaviors spanning multiple orders of magnitude. It directly incorporates fabrication constraints—such as printability, symmetry, and missing-strut defects—into the generative process. We demonstrate its ability to discover lightweight, high-performance materials for impact protection and vibration damping, outperforming conventional foams and phononic lattices. GraphMetaMat offers a scalable path toward automated, defect-tolerant metamaterial design for engineering applications. |