Rapid On-Demand Design of Inverted All-Dielectric Metagratings for Trace Terahertz Molecular Fingerprint Sensing by Deep Learning
Xueying Liu, Yinong Xie, Yiming Yan, Qiang Niu, Li-Guo Zhu, Zhaogang Dong, Qing Liu, Jinfeng Zhu
Abstract
Metasurface design with a multiplexing scheme holds promise for enhancing trace detection of terahertz (THz) molecular fingerprints. Conventional designs rely on matching spectral resonance positions with fingerprints of trace analytes, which require laborious metastructure optimizations by performing massive optical simulations. Recently, deep learning (DL) has indicated great potential for designing metasurfaces. However, its design application for THz fingerprint metasurface sensors has barely been reported so far. Here, we present a DL architecture of a bidirectional neural network to design an inverted all-dielectric metagrating (IAM) for trace THz fingerprint sensing. Based on a given THz fingerprint spectrum, our DL design tool can flexibly customize the critical sensing structure of the metagrating with the corresponding resonance frequency. Combining the designed IAM with angle multiplexing, one can excite a sequence of guided-mode resonances in a wide THz band, which supports elevating the THz fingerprint detection performance on a flat sensing surface. The DL design is used to guide the fabrication and measurement of IAM for trace α-lactose sensing, where the experimental results demonstrate metasensing enhancement by 9.3 times and imply the fast and powerful capability of our design method. Our research will inspire more DL applications on quick on-demand designs for many other THz metadevices and metasystems.