Machine learning for the free-form inverse design of metamaterials

Licentiate thesis, 2022

We present our work on using deep neural networks for the prediction of the optical properties of nanophotonic structures and for the inverse design of such nanostructures. First we show that neural networks can indeed be used to predict the optical properties of nanostructured materials such as metasurfaces. Subsequently, we show that it is possible to perform the inverse design of metasurfaces given a set of desired optical properties. This was achieved through the careful design of the neural networks and the creation of training data which were labelled with the respective optical properties and the degree to which it is possible to manufacture these nanophotonic structures. Furthermore, a CGAN network with 5 neural networks working together was developed to overcome problems with the non-uniqueness of designs, to prevent mode collapse, and to increase the experimental feasibility of the generated structures.