Inverse design of three-dimensional architected elastic metamaterials using graph neural networks

Architected metamaterials exhibit novel mechanical properties shaped by the spatial arrangement of periodic structures, rather than their constituent materials. Truss lattices, a notable subtype, are recognized for their high strength-to-weight ratio; thus, they hold significant potential for applications in robotics, aerospace engineering, and other fields. Despite recent advances in deep learning (DL) have revolutionized traditional design, researchers have mainly focused on the inverse design of quasi-static properties, leaving a gap in addressing dynamic behavior or simultaneous considerations for both aspects. To overcome this gap, we develop a novel inverse design framework to generate truss metamaterials with tailored quasi-static (stress-strain curve) and dynamic (transmission curve) properties. Our data-driven framework, based on graph neural networks, integrates a forward model into an inverse model, trained using deep reinforcement learning. . To demonstrate the model performance, finite element method simulations, uniaxial compression tests, and vibration tests are conducted to verify the properties of the optimized structures. The successful realization of user-desired properties in both quasi-static and dynamic domain can potentially accelerate the inverse design of novel materials towards applications such as lightweight and high-strength vibration isolators.