High circular dichroism response predicted by deep learning in chiral metasurfaces

Chiral resonant metasurfaces have attracted significant attention due to their wide-ranging applications in chiral sensing, on-chip optical devices, and optical communication. The designed methods of metasurfaces rely on empirical judgment and parameter scanning, which constrains the enhancement of their optical performance. In this work, we apply an inverse design method based on deep learning for optimizing structure-breaking germanium dielectric metasurfaces, and a high circular dichroism value of 0.952 can be predicted. After multiple training and testing, the coefficient of determination and root mean square error can reach 97% and 0.0015, respectively. The time taken for backward prediction is reduced by three orders of magnitude compared to the forward design via finite element methods. The method of multipole decomposition is employed to reveal the physical mechanism of the chiral response. The insights gained from this research may be helpful for the applications in efficient and intelligent design of nanophotonic devices.