Deep-learning based design and modeling for chiro-optical dielectric metasurfaces

Nanophotonics employ chiro-optical effects for a variety of applications, including advanced imaging and molecular detection and separation. Due to their outstanding qualities in light-matter interactions, planar metasurfaces comprised of subwavelength meta-atoms have attracted a lot of attention. Despite of the vast potential of metasurfaces, achievement of large chiro-optical effects compactly on-chip at the visible wavelengths is still hindered by its complex design and optimization procedure. Deep-learning (DL) based modelling techniques have been put out as an alternative to the time-consuming and computationally demanding traditional design and optimization procedure of metasurfaces during the past few years. In this work, we have employed deep-learning based forward and inverse models to design and optimize achiral nano-fins to achieve giant chiro-optical affects at the visible wavelengths. A regression based forward neural network is proposed, that takes all the structural dimensions of the achiral nano-fins as input and trained separately to predict three different types of asymmetric transmissions i.e., TLL, TLR and TRL and circular dichroism. An inverse design model is also demonstrated that simultaneously considers all the three target transmissions and optimizes the dimensions of the achiral nano-fins in such a way that they experience constructive and destructive interference, resulting in an average circular dichroism of more than 60% and 70% asymmetric transmission. With potential applications in chiral polarizers for optical displays, flat integrated polarization shifter's exhibiting high efficiency, chiral-metasurface sensors and chiral beam splitters, the suggested DL-enabled design techniques ease the realization of op-chip giant chiro-optical response through planar metasurface.