Deep learning-assisted elastic isotropy identification for architected materials

Abstract

Architected materials consisting of periodic unit cells are desirable for many engineering applications. Characterizing the elastic isotropy is of great significance for the mechanical design of architected materials. However, prevailing experimental and numerical approaches are normally too costly and time-consuming to screen out isotropic architected materials in the large design space. Here, a deep learning-based approach is developed as a highly efficient and portable tool to identify the elastic isotropy of architected materials directly from images of their unit cells with arbitrary component distributions. The measure of elastic isotropy for heterogeneous architected materials is derived firstly in this paper to construct a database with associated images of unit cells. Then a convolutional neural network is fully trained with the database, performing well on the isotropy identification with about 90% accuracy and milliseconds processing time per sample. Meanwhile, it exhibits enough robustness to maintain its performance under the fluctuating material properties in test sets. Moreover, the transfer learning of the convolutional neural network is successfully implemented among architected materials with different numbers of material components, which further promotes the efficiency of the deep learning-based approach without scarifying its identification performance. This study gives new inspirations on the rapid mechanical characterization of architectured materials, which holds promising applications in the big-data driven topological design and nondestructive testing of architected materials.