Typical generation models include Variational AutoEncoders (VAEs) and Generative Adversarial Networks (GANs), of which GANs have recently been reported to be a better choice than VAEs. The basic idea behind GANs is that the generator and discriminator compete with each other during training. However, it is very difficult to train a stable GAN, so several types of GANs have been developed, including Deep Convolutional GAN (DCGAN), Least Squares GAN (LSGAN), Wassstein GAN (WGAN), and Boundary Equilibrium GAN (BEGAN).
The performance of these The performance of these GANs can be improved by modifying the network (e.g., generator network and discriminator network) and loss function and applying some training techniques. If the training of a GAN is successful, the trained generator can eventually generate fake data that is indistinguishable from the real data. Due to these characteristics, GANs have been successfully applied in various engineering fields such as mechanical structure design, material design, and fluid dynamics. Combining GANs with other numerical techniques such as shape optimization and topology optimization further enhances their effectiveness.
