Variational MineGAN: A Data-efficient Knowledge Transfer Architecture for Generative AI-assisted Design of Nanophotonic Structures

Abstract

Leveraging the power of deep learning to design nanophotonic devices has been an area of active research in recent times, with Generative Adversarial Networks (GANs) being a popular choice alongside autoencoder-based methods. However, both approaches typically require large datasets and significant computational resources, which can outweigh the advantages of saving time and effort. While GANs trained on smaller datasets can experience challenges such as unstable training and mode collapse, fine-tuning pre-trained GANs on these limited datasets often introduces additional problems, such as overfitting and lack of flexibility. This limits the model's ability to generalize, reducing its overall effectiveness. In this study, we present Variational MineGAN, an enhanced version of MineGAN that is less susceptible to overfitting while leveraging pre-trained GANs. This approach ensures a more robust sampling process, improving the model's generalization ability while transferring knowledge to domains with limited data. Experimental results demonstrate a lower Frechet Inception Distance (FID) score of 52.14, along with an increased Inception Score (IS) of 3.59 compared to prior methods. This implies better quality design images, which allow for the exploration of highly efficient designs with desired spectral responses and improved learning of nonlinear relationships between the latent space and the corresponding designs.