Deep Learning-Based Programmable Design of Plasmonic Born-Kuhn Metasurface for Sensing Application

The intensive local light-matter interaction in the vicinity of chiral plasmonic metasurfaces facilitates their robust application for chiral medium sensing.¹ However, the target-specific design of the ultrasensitive sensor is hard to achieve due to the severe complexities in inverse engineering of the plasmonic metasurfaces. Even several researches experimentally demonstrated the significance of correlation between the sensing capability and the features of plasmonic metasurfaces, architecting the on-demand nanostructure for specific purposes is still challenging due to the absence of well-organized but simple methodologies.² In this regard, artificial intelligence may offer the programmable basis of chiral plasmonic metasurfaces to overcome the limitations, based on its extraordinary capacity to deal with intricate and non-intuitive problems.³ On this, we successfully retrieved the chiroptical properties of the representative optically active plasmonic model (Born-Kuhn) and their optical responses to the adjacent medium based on a deep learning-based study. Herein, the plasmonic Born-Kuhn model, vertically displaced coupled nanocuboids, was found to be suitable for the systematic study thanks to its availability of analytic interpretation and capability of straightforward modulation in dimension.⁴ At first, a bidirectional neural network predicting the chiroptical response of a single plasmonic Born-Kuhn model, a fundamental descriptor of chirality, was constructed by employing multi-task and tandem deep neural networks. Then, by combining the multi-task deep neural networks with a genetic algorithm, we explored the spectral change that occurred by the intervention of various refractive indices and chirality parameters and figured out specific configuration of the on-demand ultrasensitive sensor for the target medium. Once the dataset for training was prepared using the finite-difference time-domain (FDTD) method under batch simulations, the neural network was found to be exponentially resource-saving than repeated numerical computations. Our study provides insight into deep learning-based cost-efficient strategy architecting plasmonic nanostructure for a specific purpose and serves a general methodology in designing nanophotonic metasurfaces.<br/><b>References</b><br/>[1] M. Kadodwala, et al., Nature nanotechnology, 2010, 5.11: 783-787.<br/>[2] B. Kanté, et al., Nature Physics, 2020, 16.4: 462-468.<br/>[3] H. Suchowski, et al., Light: Science & Applications, 2018, 7.1: 1-8.<br/>[4] H. Giessen, et al., Nano letters, 2013, 13.12, 6238-6243.