Polarization Converter Metasurface Using Machine Learning Methods for Bandwidth Enhancement

This manuscript presents the design of a polarization converter metasurface (MTS) where Machine Learning (ML) and Deep Learning (DL) methods were used, to enhance its bandwidth. The MTS is implemented with a periodic array of multi-layered cells with chiral geometry that supports asymmetric transmission. Any impingent x-polarized wave is transmitted as a y-polarized wave, and any incident y-polarized wave is effectively reflected. In order to increase the frequency range where the incident
$\vert \mathbf{Ex} \vert$
field over the transmitted
$\vert \mathbf{Ey} \vert$
field (Tyx) is higher than 90% and the associated Rxx which is the ratio of the incident
$\vert \mathbf{Ex} \vert$
over the transmitted
$\vert \mathbf{Ex} \vert$
is less than -10 dB, the Rxx plots vs frequency were used to train an artificial neural network (ANN). The trained ANN was used to optimize the cell's geometric dimensions to create a MTS response with more than 90% conversion rate (Tyx>90%), in the frequency range from 6.5 to 12.5 GHz, resulting in more than 60% fractional bandwidth.