Metasurfaces designed by a bidirectional deep neural network for generating quantitative field distributions

Benefiting from superior capability in manipulating wavefront of electromagnetic waves, metasurfaces have provided a flexible platform for designing ultracompact and high-performance devices with unusual functionalities. Despite various advances in this field, the unique functionalities achieved by metasurfaces have come at the cost of the structural complexity, resulting in a time-consuming parameter sweep for the conventional metasurface design.

Although artificial neural networks provide a flexible platform for significantly improving the design process, the current metasurface designs are restricted to generating qualitative field distributions. Therefore, new artificial neural networks should be proposed to inversely predict metasurface-based devices with quantitative functionalities.

In a new paper published in Light: Advanced Manufacturing, a team of scientists, led by Professor Yiming Zhu from Terahertz Technology Innovation Research Institute, and Shanghai Key Lab of Modern Optical System, University of Shanghai for Science and Technology, China, and co-workers have developed a network architecture consisting of a tandem neural network and an iterative algorithm to predict metasurfaces with high-accuracy functionalities.

As a proof-of-principle example, they design multifoci metalens (predict by the network architecture) that possess multiple focal points with identical polarization states, as well as accurate intensity ratios. In addition, metalens for generating two focal points with orthogonal polarization states and accurate intensity ratios, and vortex generators for generating position- and polarization-dependent converged vortices were predicted/designed and experimentally demonstrated.

In comparison with the traditional forward-design, the inverse designed methods in this work can automatically and intelligently optimize the ratio of field intensities between two focal points. The tandem neural network enables unprecedented capability in globally optimizing the whole (target) amplitudes and phase. Therefore, one can use the optimized amplitudes and phase to easily design the metasurface to generate quantitative functionalities.

The presented approach will promote machine learning in further designing ultracompact devices with high-accuracy and quantitative functionalities in imaging, detecting, and sensing