Bayesian Optimization for Nested Adversarial Variational Autoencoder in Tunable Nanophotonic Device Design


We propose a new device optimization framework based on Bayesian optimization for efficient latent sampling of adversarial generative neural networks to expedite a complex inverse design of tunable nanophotonic wavelength splitters. Our design, at broadband telecomm-wavelengths, is electrically switchable via liquid crystal tuning.