Probabilistic inverse design of metasurfaces using mixture density neural networks

Abstract Metasurfaces are planar sub-micron structures that can outperform traditional optical elements and miniaturize optical devices. Optimization-based inverse designs of metasurfaces often get trapped in a local minimum, and the inherent non-uniqueness property of the inverse problem plagues approaches based on conventional neural networks. Here, we use mixture density neural networks to overcome the non-uniqueness issue for the design of metasurfaces. Once trained, the mixture density network (MDN) can predict a probability distribution of different optimal structures given any desired property as the input, without resorting to an iterative local optimization. As an example, we use the MDN to design metasurfaces that project structured light patterns with varying fields of view. This approach enables an efficient and reliable inverse design of fabrication-ready metasurfaces with complex functionalities without getting trapped in local optima.