Towards neural network integration of metasurfaces in optical design software

Metasurfaces are a new tool for optical designers that could revolutionize the field by making smaller and lighter optical systems where one metasurface can replace many optical components. However, these surfaces are not defined in optical design software, so only specialized designers can exploit this breakthrough. During the recent year, we have developed a DLL using a semi-analytical model to convert incident rays on a metasurface to an output ray. In this work, we propose to use the semi-analytical model to generate data to train a neural network to output the phase gradient from an incident ray (position and angle). Once trained, the neural network is implemented in Zemax OpticStudios as a User Defined Surface to ouptut the refracted ray direction (direction cosine). We report our main results as well as the limitations of this approach.