CNN-based image recognition for topology optimization

Effectiveness of several currently popular topology optimization methods is closely related to the number of design variables consisted of discretized finite elements. Since the number of design variables is proportional to the number of finite element meshes, a very fine discretization needs more computational cost to carry out a finite element analysis and evaluate a compliance based objective function with the volume constraint. This paper presents a new computational method by using convolutional neural networks (CNNs) which can be substituted for the FEM process to calculate compliances. The robustness and adaptability of the proposed method are tested on a MBB (Messerschmitt-Bölkow-Blohm) beam and two cantilever beam problems. The designed CNN is trained on a 48 × 16 pixel resolution dataset taken from coarse meshes. The trained CNN can predict the information of image-based topologies composed of fine meshes. A graphics processing unit (GPU) is then used to accelerate the bulk-processing of data. •A new surrogated model is proposed to predict compliance information for topology optimization.•The proposed method can eliminate the step of FEM and accelerate optimization processes.•The CNNs are then introduced to train neural networks by using coarse elements.•High resolution image can be predicted in the trained NNs by using resize interpolation methods.•A GPU is then used to accelerate the bulk-processing of data.