Deep super-resolution neural network for structural topology optimization

Abstract

ABSTRACT A deep-learning approach is proposed to predict an optimized high-resolution structure with multi-boundary conditions. An enhanced deep super-resolution (SR) neural network and a convolutional neural network are constructed and trained to establish the mapping relationship between low- and high-resolution structures for the topology optimization problem. The data set for training and testing is generated using the solid isotropic material with penalization method, in which the training and test sets have different geometric boundary conditions. Each sample contains both low- and high-resolution structures. The deep neural network is trained with limited training samples (4000), and numerical experiments demonstrate that the proposed method can obtain an accurate high-resolution structure in negligible computational time. Moreover, the proposed method has the generalization ability necessary to predict high-resolution structures with multi-geometric boundary conditions. The effective incorporation of a deep SR neural network and topology optimization has enormous potential for future practical applications in large-scale structural design.