Machine Learning-Enabled Photonic Metamaterials

In this talk, I will discuss advances in photonic engineering in which machine learning approaches to device implementation unlock new functional capabilities. First, I will discuss new concepts in nanostructure geometric parameterization, inspired from the computer graphics community, which enable freeform layouts to be specified in a manner that can be differentiated and maintain hard constraints. Second, I will show how deep generative networks can be used to perform population-based global optimization, producing best in class devices. Third, I will show how physics-augmented deep networks can serve as accurate surrogate electromagnetic solvers and how innovations in network architecture can enable these solvers to generalize to arbitrary sized domains and grayscale dielectric media. We anticipate that the ability for deep learning models to dramatically accelerate and even automate the simulation and design of photonic systems will push the innovation cycle in all domains of photonics research.