Disrupting the photonics innovation cycle with data- and physics-driven algorithms

I will discuss the role of network architecture in the GLOnet inverse optimization platform, in which the global optimization process is reframed as the training of a generative neural network. I will show how a properly selected network architecture can smoothen the design space and how the architecture can be tailored based on the type and dimensionality of the design problem. I will also discuss new methods in which neural networks can serve as high speed surrogate Maxwell solvers capable of aiding the inverse design process. These hybrid physics- and data-driven concepts can apply to a broad range of nanophotonics systems.