Designing three-dimensional lattice structures with anticipated properties through a deep learning method

Abstract

Lattice structures have been a hot topic recently owing to their superior mechanical properties, which are significantly influenced by the unit cell structure. By leveraging the power of deep learning, inverse design can be conducted on the unit cell structure based on the mechanical properties of its lattice structure. Assisted by deep learning, this study introduces a novel data-driven approach to design three-dimensional (3D) unit cells for lattice structures with anticipated properties. The approach can be efficiently and accurately applied to various unit cell structures. An auto-encoder is trained to extract the geometric features from unit cell point clouds. The effective mechanical properties of the lattice structures are calculated by combining the homogenization method and the finite element method. Subsequently, a mapping relationship between mechanical properties and geometric features is established through the multi-layer perceptron neural network. The models are ultimately employed to design 3D unit cells given anticipated properties of lattice structures. The results show that the mechanical properties of the generated unit cells satisfy the anticipated values. The applications of proposed method are demonstrated in orthopedic implants, new hybrid unit cells, and functionally gradient structures. Furthermore, the method can be extended to unit cell design across diverse domains.