Broadband Achromatic Metalenses Design Based On Deep Learning

Metasurfaces,composed of artificially designed sub-wavelength two-dimensional array of metamaterials,have exceptional competence in manipulating optical amplitude,phase and polarization and other optical responses.Metasurface based lenses,also called metalenses,are a kind of diffractive lenses with subwavelength meta-units.They can accurately control phase profiles by manipulating the geometric of meta-units,so as to realize focusing and imaging.However,the chromatic aberration occurs in the broadband,which will cause the focal length to change with the wavelength and affect the focusing and imaging quality.Therefore,achromatic metalenses have become the focus of research in recent years.As for broadband achromatic metalenses,one key challenge is that meta-unit simultaneously needs to meet the phase requirements of different positions and different wavelengths.The traditional design method is to simulate the electromagnetic(EM)response of a large number of meta-units under a wide spectrum through numerical simulation methods,build a candidate library of meta-units,and then iteratively search for the meta-unit that meets the target requirements through optimization algorithms,which is extremely time consuming and ineffective.This thesis proposes a deep learning approach to design broadband achromatic metalenses,which establishes a deep neural network(DNN)to learn the relationship between structures and EM responses of meta-units.By feeding the target phase curves to the input of the optimized DNN,it could output the geometric parameters of meta-units to design a metalens.The main research results and innovations of this thesis include:(1)In the process of establishing the deep neural network,it is proposed to convert the abrupt phase data into x-y projection data pairs through the trigonometric function relationship,which is used for the output of the forward neural network and the input of the reverse neural network.It solves the problem of FDTD The phase of the simulated data set has a sudden change due to the wrapping,which improves the accuracy of the neural network.(2)In order to be compatible with the established design network,a target phase curve fitting method is proposed for the design of broadband achromatic metalenses.The phase curve of the design target was taken as the input of the network,and then the corresponding meta-units is output directly.Some phase curves similar to the target are artificially generate as a dataset to fine-tune and optimize the network,which improves the accuracy of the design.(3)To demonstrate the performance of the well-trained DNN,two achromatic metalenses based on elliptical nanopillars and nanofins for a spectra band in near-infrared are designed.One-dimensional simulation of the designed metalenses are carried out using FDTD,and the optical performance after the simulation are made.Their average focal length shifts are 2.6%and 1.7%,while the average relative focusing efficiencies reach 59.18% and 77.88%,respectively.