mcGANs: Topology Optimization Using Multi-Stage CGAN

Abstract

The traditional methods of solving topology optimization depends on finite element method (FEM). However, the iterative calculation of FEM increases the time cost of topology optimization significantly. In this paper, a novel strategy based on deep learning is proposed to speed up the topology optimization process. Conditional Generative Adversarial Network (CGAN) is used as the basic network in which the constraints of optimization are taken into consideration as the conditions of generation. Additionally, the stress and strain are generated from initial physical information by networks, and used as the inputs of optimization network. The cascaded network can realize structural topology optimization without iterative calculations in the whole process. In the experiments, the traditional Solid Isotropic Material with Penalization (SIMP) method was replicated and implemented multiple times. It was found that the SIMP method takes 30 - 40 seconds to conduct topology optimization, while the proposed mcGANs only need 1.7682 seconds. Moreover, performance indicators such as Mean Absolute Error (MAE), Mean Squared Error (MSE), and Peak Signal-to-Noise Ratio (PSNR) are used to evaluate the performance of mcGANs, which demonstrate the effectiveness and superiority of this strategy.

Keywords

• Deep learning
• topology optimization
• cGAN
• U-Net
• SE-ResNet