Graph-based metamaterials: Deep learning of structure-property relations

• Proposed a systematic way of constructing shell-lattices based on an efficient graph representation. • Determined the elastic and thermal properties of more than 43 , 000 distinct shell-lattices. • Proposed Crystal Graph Convolution Neural Network to describe the structure–property relationships. • Discovered new shell-lattice topologies that provide targeted macroscopic properties. The effective properties of metamaterials are tailored through the design of their internal structures. According to their main building block, the family of porous three-dimensional metamaterials is divided into truss-, plate- and shell-lattices. The exploration of their full design-space is hampered in practice by a lack of a systematic method to represent their topologies. Here, we demonstrate for the first time that graph models provide an effective representation of shell-lattices. This new graph representation is then leveraged to obtain deep learning-based structure–property models. Using finite element simulations, the stiffness and heat conductivity tensors are established for more than 40,000 microstructural configurations. We find that a modified crystal graph convolutional neural network model provides an accurate description of the structure–property relations. We anticipate the proposed graph-based modeling framework to be applicable to any man-made periodic microstructure, thereby enabling the design and discovery of new materials exhibiting exceptional mechanical, thermal, electrical or magnetic properties.