Lightweight and Strong Lattice Structure Designs by Generative Machine Learning and Additive Manufacturing

Designing lightweight and strong structures has been a research area of interest in the materials science field. The lattice-truss structure, inspired by the lattice of single-crystalline in nature, is thought to be more lightweight compared to conventional engineering materials. However, the inherent limitations of the lattice structure rely on the trade-off relationship between density and stiffness, which creates a barrier to being used in various applications. In order to alleviate the compromise, many previous works have suggested various lattice structures such as octet, body-centered cubic, simple cubic, or face-centered cubic structures. Despite various proposed lattice structures, the shape of the beam element is often overlooked as a design variable. In this study, we suggest the optimal shape of beam elements in various lattice structures having both high elastic stiffness and strength by employing an artificial intelligence-based neural network(NN). We create numerous initial datasets of beam elements modeled by a high-order Bézier curve and predict the elastic modulus of the truss using the finite element analysis(FEA) and the homogenization method. Two NNs are trained to learn the complicated relationship between the physical properties (density and modulus) of each lattice structure and the geometry of the beam element. Because of the high accuracy of NN predictions, we are able to predict the properties of lattice structures without FEA, and the NNs can predict the relative density and modulus in about 0.1 seconds, at a much faster computational time compared than FEA. By leveraging the fast inference of the NNs, we apply genetic optimization(GO) to generate offspring datasets having high modulus compared to initial datasets. The NNs are then updated by training with all datasets including offspring datasets to expand the reliable prediction domain. We run GO again using the new parents set which has higher outputs than the previous optimization and iterate the NN-GO feedback called “active learning” until convergence. Using the active learning-based optimization approach, we are able to gradually enhance the modulus of the lattice structures and the final optimal designs show remarkable improvements in elastic stiffness compared to the typical lattice structure consisting of cylindrical beam elements. We then fabricate the optimal lattice structures along with other benchmark designs using additive manufacturing and conduct compression tests to validate our optimal designs and computational approaches. In our experimental results, the optimal design has much higher stiffness and strength than the benchmark lattice structures, which is in good agreement with simulations, and the strengthening effects are explained by the distributed stress field predicted from our FEA. Our method can be extended to design other complicated lattice structures, and the optimal specific stiffness and strength can be utilized in the development of novel lightweight structures.