Search

• Combining ML and FE methods to accelerate the design process of honeycomb structures. • Using 2D unit cell surrogate 3D honeycomb structures. • CDAE with a pixelation strategy is proposed to manage FE data. • LSTM uses latent features and external loads to predict structural dynamic responses. In this study, a novel framework was presented for accelerating the prediction of the mechanical response of honeycomb structures under dynamic crushing, using 2D cells to surrogate 3D honeycomb structures by machine learning (ML). The sizes of different honeycomb structures were designed and the necessary training data obtained through finite element (FE) simulations, but without using any explicit design parameters of the honeycomb cells in the ML model. A pixelization method of FE model was proposed to separate the complete cell structure from the honeycomb FE grid data and convert it into a matching pixel map. FE data was structuralized while reducing the large amount of computational power consumed in identifying complex array structures. Unsupervised automatic extraction of low-dimensional features of pixel maps was performed using a convolutional denoising autoencoder (CDAE). The crushing velocity and the extracted latent features were used as the input of the long-short term memory (LSTM) network to predict the crushing deformation and stress-strain curve of the honeycomb structure under different dynamic loading. Results showed that the constructed ML model could describe the dynamic crushing response behavior. Compared with the traditional FE method, the prediction model was 4.45 × 103 and 1.05 × 103 times faster in predicting the stress-strain and structural deformation response, respectively. The mechanical response prediction model provided a method for rapidly evaluating the dynamic mechanical response behavior of similar periodic array structures using FE model data, which could be beneficial for the design and development of equipment based on bionic structures. [ABSTRACT FROM AUTHOR]