Pushing the Limits of Functionality Multiplexing Capability in Metasurface Design Based on Statistical Machine Learning

Enabled by cleverly designed nanostructures, metasurfaces offer an unprecedented way to manipulate light on a completely planar platform. The versatility and unique optical responses of metasurfaces under different illumination conditions are beneficial to encode independent information. Typically, a top-down, two-step design process for multifunctional metasurfaces proceeds as follows: First, the design goals are decoupled to obtain the phase requirements for each illumination condition. Then, the metasurfaces are built by multiplexing single-responsive meta-atoms or assembling multiple phase-tuning techniques such as the geometric phase and resonant phase. However, with an ever increasing number of design channels, this empirical design intervention becomes infeasible.<br/><br/>In this study, we propose utilizing machine learning techniques to capture the statistical features in the high-dimensional joint distribution of meta-atoms in the unit cell and the corresponding optical responses, thereby supporting a quantitative evaluation of the severity of function crosstalk and an on-demand inverse design. We enhance the design capabilities for various meta-atom design frameworks with the assistance of such framework. To provide an automated end-to-end design loop for multifunctional metasurfaces, the developed machine learning model can be seamlessly integrated with the gradient-based and non-gradient optimization processes. We have successfully constructed metasurface focusing lenses and holograms operating in eight customizable channels in the near-infrared range utilizing a single metasurface, where each channel is a combination of different incident frequencies and polarizations. We have also researched free-form meta-atoms represented by binary images, which support more resonance modes and enhance the independence of each information channel.<br/><br/>Our results suggest that the data-driven optical design technique outperforms conventional physical-guided methods, which will expedite the development of novel optical display, communication, and computing devices and systems.