Deep Neural Network For Designing Plasmonic Nanoantenna And Silicon Structural Color

Nanophotonics are indispensible in optical sensing,energy conversion and electromagnetic cloaking applications by effectively controlling the interaction between electromagnative wave and material at the subwavelength scale.The efficient design and fabrication of nanophotonic devices can realize easy modulation of photonic information,thus greatly promote the development of advanced photonic technologies.However,conventional nanophotonic design usually relies on the designer’s intuition and experience,and iterative optimization is necessary to obtain the target electromganative response.This process usually requires a lot of computing time and resources.Unfortunately,with the increasing complexity of the modern devices,it is still a challenging problem for numerical computing tools to search for the global optimal solution in a huge parameter space,which greatly hinders further development of high performance nanophotonic devices.The neural network algorithms based on machine learning have demonstrated powerful applications in nanophotonics due to its computation efficiency,accuracy and capacity,which has led an important interdisciplinary research direction.By learning from a sufficient amount of training data,the neural network can accurately capture the physical relationship between parameters such as nanophotonic geometry,materials,dielectric environment,and their electromagnetic responses.Therefore,deep learning algorithms can be used for efficient forward prediction and accurate inverse design.Through innovations in the neural network architecture and training method,this paper realizes the design of bow-tie plasmonic nanoantenna and silicaon structural color.This work focuses on the study and design of bow-tie plasmonics nanoantenna.FDTD simulation tool is utilized to investigate the highly sensitive and non-unique far-field electromagnetic response of the bow-tie plasmonic structure.By constructing both tandem network and iterative network to design the far-field spectrum of bowtie nanoantenna,this paper compares factors such as design accuracy,architecture complexity and training efficiency.In conclusion,iterative network is a better choice for this type of problem and this work serves as a good reference for selecting a suitable model in similar design problems.Furthermore,in order to realize the inverse design of near-field properties in this bow-tie plasmonic structure,this paper proposes several modifications of the ordinary neural network’s structure and training process.The main measures are summarized as two kernels: The logarithmic method is used to eliminate the magnitude difference between the field strength data,effectively limiting the data range and retaining the original features;MAE instead of MSE loss function can better deal with outliers.These two findings are helpful to deal with data which are similar to nearfield enhancement data that are spanned in a large range.The optimized network enables the collaborative design of the near-and far-field properties of the nanoantennas,which can be applied to dual-functional plasmonic and SERS sensors.Finally,in order to expand the color palette of silicon metasurface,this paper uses MATLAB to generate a variety of two-dimensional structure images including rectangles,circles,rings,and crosses as input datasets.The reflection spectrum is derived through FDTD tool to caculate xy Y values of corresponding structural colors and they are used as the output datasets.The results show a significant increase of the color gamut coverage area.Convolutional neural network is adopted to deal with the 2D structural image data and realized accurate,high-freedom prediction of structural color values.Besides geometry structure,other important factors that affect the electromagnative response such as period,height,material index and so on are also examined.In the end,this paper proposes a novel predicting neural network as a general model for designing nanophotonics with high degrees of structural freedom.Furthermore,the superiority of the predicting neural network is successfully demonstrated through examples of co-designing the geometry and period in silicon metasurfaces.This paper presents deep neural network as an effective design tool which enables high performance nanophotonic devices.Such novel methods and technologies will continue serve as an important guideline for developing efficient functional nanophotonic devices in the future.