Deep learning-based heterogeneous strategy for customizing responses of lattice structures

Abstract

Designing lattice structures with tunable mechanical behavior for multi-functional applications is of great significance. However, the inverse design of lattice structure for the specific requirement is still challenging due to the complex nonlinearity between the lattice configuration and its mechanical behavior. Herein, a deep learning-based heterogeneous strategy is proposed to design the heterogeneous lattice structure with a customized target response. Heterogeneous lattice structures comprised of octet-truss and rhombic dodecahedron cells are designed and fabricated by stereolithography using resin. Mechanical properties of heterogeneous lattice structures are determined by quasi-static compressive experiment and finite element analysis. The nominal stress-strain curves of independent heterogeneous lattice structures are calculated employing the finite element model. Based on these data, an artificial neural network is trained, validated, and tested. Influences of octet-truss cell number along the loading direction as well as interface number on the mechanical properties of lattice specimens are numerically examined. With the aid of the well-trained artificial neural network, the heterogeneous lattice structures with various specific target performances are successfully achieved, which are also experimentally verified. The results show that the heterogeneous lattice structures are more suitable for energy absorption than monolithic octet-truss and rhombic dodecahedron lattice structures. The prediction of finite element analysis can be reproduced by an artificial neural network effectively and precisely. The present strategy broadens the design space of lattice structures and provides a novel approach for designing the lattice structure with a specific response.