A DEEP LEARNING APPROACH FOR PREDICTING TRANSMISSION SPECTRA OF METASURFACES

Metasurfaces, ultrathin 2D structures, have drawn considerable attention in recent years due to their applications in many fields. The metasurface is a subwavelength‐engineered structure with unique electromagnetic properties and exceptional light‐matter interac‐ tions. Recent research indicates that metasurfaces are a good fit for studying the polar‐ ization conversion of electromagnetic waves. In search of polarization conversion to any arbitrary angle, we have considered taking advantage of deep learning to investigate the field thoroughly. In recent years, the contribution of deep learning in optics and photon‐ ics has been undeniable. This work presents a neural network miming a conventional numerical simulator like Ansys Lumerical FDTD. We have introduced a neural network based on ResNet‐18 to predict the transmission spectra of Gold metasurfaces within the 1200 to 1700 nm region. The model is trained with thousands of simulated data of metasurfaces with varied geometries. The model can accurately predict the trans‐ mission spectra within a couple of milliseconds. Evaluation of 150 datasets shows that the model has an average prediction accuracy of 85.27%. The model aims to present a time‐efficient method for investigating polarization conversion metasurfaces.