Mechanics prediction of 2D architectured cellular structures using transfer learning

Two-dimensional (2D) architectured cellular structures exhibit outstanding mechanical properties unmatched by their bulk counterparts and show promising outlooks in electronic applications. Understanding of the relationship between their mechanical properties and structure patterns has yet to be fully explored. Also, traditional design rules in 2D architectured structures requiring prior knowledge of geometric parameters impose fundamental challenges for achieving desired performance within a rapid optimization process. Here, by taking full advantage of unsupervised generative adversarial network-based transfer learning (TL) and high-performing coarse-grained molecular dynamics (CGMD), we propose an adaptive design strategy to predict the mechanical performance of 2D architectured cellular structures as well as unravel hidden design rules for maximizing specific tensile strength. Results indicate that the established TL model is accurate enough to predict the mechanical properties of graphene kirigami, in which [Formula: see text] is 0.994 and 0.985 for specific strength and yield strain, respectively. The proposed design method combining machine learning with CGMD extends the ability of physical simulation beyond performance prediction, optimizing fracture mechanical properties by screening through the entire geometric design space of the architected 2D structures. Overall, this work proves that the design method based on TL can effectively obtain the power of new physical insights for structure design and optimization of interest.