Tackling multimodal device distributions in inverse photonic design using invertible neural networks

We show how conditional generative neural networks can be used to efficiently find nanophotonic devices with desired properties, also known as inverse photonic design. Machine learning has emerged as a promising approach to overcome limitations imposed by the dimensionality and topology of the parameter space. Importantly, traditional optimization routines assume an invertible mapping between the design parameters and response. However, different designs may have comparable or even identical performance confusing the optimization algorithm when performing inverse design. Our generative modeling approach provides the full distribution of possible solutions to the inverse design problem, including multiple solutions. We compare a commonly used conditional variational autoencoder (cVAE) and a conditional invertible neural network (cINN) on a proof-of-principle nanophotonic problem, consisting in tailoring the transmission spectrum trough a metallic film milled by subwavelength indentations. We show how cINNs have superior flexibility compared to cVAEs when dealing with multimodal device distributions.