Inverse machine learning framework for optimizing gradient honeycomb structure under impact loading

Abstract

In this study, an inverse design framework was constructed to explore gradient honeycomb structures (HCS) with high impact resistance. By establishing the relationship between the height of the cell and the cell-wall angle, HCS with different gradient modes were designed. The developed machine learning (ML) framework consisted of a forward regression model based on neural network (NN) and an inverse optimization model based on generative adversarial network (GAN). The regression model surrogated finite element analysis (FEA) to rapidly provide corresponding mechanical properties for the generated data of GAN. To ensure that the generated data met the desired design targets, additional physical constraints were incorporated into the generator of the GAN. This ensured that the network output not only consisted of valid data but also exhibited superior impact resistance. The results demonstrated that the trained ML model optimized both specific energy absorption (SEA) and initial peak stress of HCS. It even discovered high-performance gradient designs that outperformed the original data, with a maximum energy absorption 49.5% higher than that of uniform HCS. Moreover, the gradient modes and structural parameters of superior designs were identified, which is of great significance for the gradient design of HCS under impact loading. As a general design approach, this ML framework holds significant potential for the optimization design of metamaterials in other structures or with different mechanical properties.