Flexibly Designable 2D Chiral Metasurfaces with Pixelated Topological Structure Based on Machine Learning

2D chiral metasurface have been widely used for planar circular dichroism (CD) devices. However, the design of 2D metasurfaces usually means a time-consuming search of effective structures, which are typically limited to regular geometries and compromised performances. Here, an efficient method for 2D chiral metasurfaces with pixelated topological structure based on machine learning (ML) is proposed and demonstrated from which simultaneous high CD and extinction ratio (ER) either over broadband or at specific wavelength can be efficiently achieved. The proposed ML method combines both advantages of high efficiency of neural network (NN) and superior goal evolution ability of microbial genetic algorithm (MGA). Unlike traditional empirical-driven methods, the hybrid framework and pixelated topologically deformable structures can fully exploit the potential of design space and push design capability to its physical limit. An average CD of 94.7% and ER of 15 dB over wavelength from 1.45-1.65 mu m and a CD of >90%/ER of >27 dB at freely-selected wavelength of 1.54 and 1.616 mu m are obtained. Experiments with fabricated topological structures validate the theoretical predictions. The proposed pixelated structures with ML provide a universal method for precise tailoring of optical properties of metasurfaces which is otherwise unattainable with conventional regular geometries.