Numerical characterization on mechanical performances of artificial cuttlebone-like structures produced via the generative adversarial network

Abstract

Advanced manufacturing techniques eliminates constraints of time and cost associated with fabrications of complex structures, thereby significantly promotes the development of biomimetic lightweight structures. Currently, biomimetic structures are often simplified as periodic lattice structures, while the imitation of natural structures' random yet coherent arrangements is highly limited. This study utilizes a progressive algorithm called SinGAN, which is a modified version of GAN, to capture the inherent randomness presented in biostructures. In this paper, we focus on the ultralight cuttlebone structure, where SinGAN is conducted feature extractions and structural learning. The input for the process includes the 2D scanning images, which are then used to generate imitated structures through the network. The 2D images are vertically extruded to convert them into 3D structures for further analyses. The stiffness of all structures is determined through finite-element-based simulations. Computed results reveal that SinGAN successfully extracts the structural features of cuttlebone. In addition, some property of the imitated structures exhibits much more superior behavior comparing to the original structure. The presented approach for generating imitated structures is one of the first few endeavors to integrate artificial intelligence into biomimetic structure design, to the best of our knowledge.