Key points are not available for this paper at this time.
A longstanding objective of machine learning-enabled inverse design is the realization of inverse neural networks that can instantaneously output a device given a desired optical function. For complex freeform devices, generative adversarial networks (GANs) can learn from images of freeform devices, but basic GAN architectures are unable to fully capture the intricate features of topologically complex structures. We show that by coupling progressive growth of the network architecture and training set with the GAN framework, generative networks can be trained to output high-performance, robust freeform metasurface devices. A combination of convolutional and self-attention layers in the network enable the accurate capture of both short- and long-range spatial patterns within topologically complex layouts. In applying this training methodology to metagratings, the best generated devices have efficiency and robustness metrics that compare with or outperform the best devices produced by gradient-based topology optimization with comparable computational cost. This study showcases the capability of generative neural networks to capture highly intricate geometric trends in physical devices, such as robustness constraints in freeform metasurfaces, and demonstrates their potential as black box inverse design tools for complex photonic technologies.
Wen et al. (Fri,) studied this question.