Deep learning for topology optimization design

Generative modeling techniques are being rapidly developed in the field of deep learning, and they have been applied to topology optimization. The variational autoencoder (VAE) is a generative modeling technology that extends the autoencoder to generate new images with a limited latent space. We modified the basic VAE structure to encode optimization conditions and decode latent variables for topology optimization design. The modified VAE could efficiently predict the optimized structure after topology optimization. However, it was difficult to train the neural network to predict a very detailed structure. The generative adversarial network (GAN) is another generative modeling technique for generating images and is implemented by a system of two neural networks competing with each other to make realistic synthetic data. We applied GAN to obtain more detailed optimized results from the original design. The proposed methodology of using two generative modeling techniques for deep learning is expected to significantly increase the efficiency of topology optimization.