Evaluation of machine learning techniques for conditional generative adversarial networks in inverse design

Recently, machine learning has been introduced in the inverse design of physical devices, i.e., the automatic generation of device geometries for a desired physical response. In particular, generative adversarial networks have been proposed as a promising approach for topological optimization, since such neural network models can perform free-form design and simultaneously take into account constraints imposed by the device's fabrication process. In this context, a plethora of techniques has been developed in the machine learning community. Here, we study to what extent new network architectures, such as dense residual networks, and other techniques like data augmentation, and the use of noise in the input channels of the discriminator can improve or speed up neural networks for inverse design of optical metasurfaces. We also investigate strategies for improving the convergence of the training of generative adversarial networks for inverse design, e.g., temporarily freezing the discriminator weights when the model outperforms the generator and training data blurring during the early epochs. Our results show that only some of these techniques improve inverse design models in terms of accuracy and stability, but also that a combination of them can provide more efficient and robust metasurface designs.