Deep Learning-Assisted Design for Metasurfaces with Purified Single Resonant Peak

Zhenxiang Shi,Haiou Lu,Bing Yan,Xinyu Yu,Min Zhang,Kai Ni,Qian Zhou,Xiaohao Wang
DOI: https://doi.org/10.1117/12.3006742
2023-01-01
Abstract:Dielectric metasurfaces designed based on bound states in the continuum (BICs) exhibit extremely high quality factors, high sensitivity, and low loss, making them suitable for surface-enhanced infrared absorption and refractive index sensing. It has been demonstrated in some literatures that the symmetric property of the elliptical unit cell of the BICs mode can be broken by increasing the rotation angle, thereby achieving high-quality factor resonant structures. By varying the rotation angle, major and minor axis lengths of the ellipse, and the period of the bilayer elliptical unit, unique resonant properties can be achieved even without considering the existence of BIC. In this work, we have designed a deep learning-based transmission spectrum prediction network by combining the parameters of the elliptical unit. This network can replace traditional electromagnetic simulation calculations to quickly obtain the transmission spectrum of the target structure. While simulating 3000 sets of elliptical unit using the finite-difference time-domain method on an i7- 12700H processor requires 14 hours, using our neural network yields a transmission spectrum with prediction accuracy better than 10-3 in less than 6.9 seconds, significantly improving the design efficiency. The network takes the image of metasurface unit as input data and couples the scaling factor that affects the period of the metasurface into the image data, making it possible to train and predict the spectra with different structures and ratios, effectively improving the generalization ability of the network.
What problem does this paper attempt to address?