Global Optimization and Deep Learning Techniques For Freeform Nanophotonic Metadevices

Engineered freeform components are advancing the state of the art in the field of nanophotonics. To this end, optical metasurfaces are increasingly desirable due to their ability to control the phase, magnitude, and polarization of transmitted (and/or reflected) light and improvements in nanofabrication techniques. While metasurfaces have demonstrated unprecedented control over optical performances, it is not well understood how to determine the geometries necessary to achieve a targeted optical performance. That being so, optimization and inverse-design techniques have proven invaluable to engineers seeking to realize metadevices with specific targeted functionalities. In this presentation, we discuss the advantages of our metadevice design process which utilizes a combination of global optimization and deep learning to realize an accelerated mutliobjective framework for generating highly performant designs that are robust to fabrication uncertainties.

I. Introduction

The field of nanophotonics has grown significantly in recent years due to developments in both nanofabrication and device design and optimization. Perhaps the most exciting nanophotonic devices are those based on optical metasurfaces [1]. These so-called “metadevices” are often realized through the patterning of unintuitive freeform structures which have been engineered to achieve a desired functionality over a specified frequency band and angular field of view of interest. Metasurfaces work by introducing engineered phase discontinuities on a surface which lets designers exploit the generalized form of Snell's law to achieve tailored reflection and transmission behaviors not possible with conventional optical components [2] including non-mechanical reconfigurability [3]. These phase discontinuities are often realized through the patterning of constituent unit cells (i.e., “meta-atoms”) which have been designed to produce a desired dispersion-engineered phase profile. However, the performance potential of the metasurface is ultimately determined by the behaviors of its individual meta-atoms.