Chiral Metasurface for Near-Field Imaging and Far-Field Holography Based on Deep Learning

Chiral metasurfaces have great influence on the development of holography. Nonetheless, it is still challenging to design chiral metasurface structures on demand. As a machine learning method, deep learning has been applied to design metasurface in recent years. This work uses a deep neural network with a mean absolute error (MAE) of 0.03 to inverse design chiral metasurface. With the help of this approach, a chiral metasurface with circular dichroism (CD) values higher than 0.4 is designed. The static chirality of the metasurface and the hologram with an image distance of 3000 mu m are characterized. The imaging results are clearly visible and demonstrate the feasibility of our inverse design approach.