Chiral metasurface optimization using a genetic algorithm and a neural-network based approach

We use a genetic algorithm and a neural-network based approach to design
chiral reflectors. The structures considered consist of the laterally periodic
repetition of C4-symmetric air patterns on a dielectric layer made of gallium
phosphide. These patterns are defined by the (x,y)-coordinates of the Nc corners
in the top-right quadrant of the periodic cell (Nc=3, 4 or 5 for the structures
considered). The objective is to determine the patterns that maximize the
difference between the reflection of left-handed and right-handed polarizations
for normally incident radiations on a given frequency range. We use for this
purpose a genetic algorithm and a neural-network based approach to determine
optimal coordinates for the Nc corners of the patterns considered. The two
optimization approaches are run in parallel to compare their respective results
and efficiencies. The study reveals the influence of the pattern complexity on the
desired chiral reflection effect. We provide finally an analysis of the
computational resources required by the two approaches.