Hybrid Quantum Generative Adversarial Networks To Inverse Design Metasurfaces For Incident Angle-Independent Unidirectional Transmission

Abstract: Optimization of metasurface designs for specific functionality is a challenging problem due to the intricate relation between structural features and electromagnetic responses. Recently, many researchers resolved to inverse design of metasurfaces for efficient design parameters based on methods such as parameter optimization, evolutionary optimization and machine learning. In this paper a hybrid quantum machine learning method which uses quantum encoders to enhance the performance of a classical GAN is applied to implement inverse design of a metasurface. Aiming towards angle-independent unidirectional transmission, this approach combines a Quantum Generative Adversarial Network (QGAN) with a Variational Autoencoder (VAE) to optimize metasurface designs. Incident-angle independent uni-directional transmission has potential applications in efficient solar cells, thermal cooling, non-reciprocal devices, etc. However, it is very challenging to achieve such a metasurface design via forward methods and hence very few studies exist till now in this direction. The methodology employed in this work reduces the data requirement for inverse design by 30% compared to conventional GAN-based methods. More importantly, the developed metasurface designs show high fidelity of 95% with regard to the targeted far-field radiation patterns. We also provide a material look-up table for feasible substitutes of the obtained material design with real materials and yet maintaining performance accuracy. Further, we embed the inverse-designed metasurfaces into Perovskite solar cell layers to demonstrate the improvement in its performance. We observe that the conversion efficiency of the example perovskite solar cell improves by 95% and remains independent of incident angle in the range -60$^\circ$ to 60$^\circ$ within the desired frequency range.