Optimization Design Of Infrared Polarization Insensitive Device Structure In SNSPD Based On Machine Learning

In the past two decades,Superconducting Nano wire Single Photon Detector(SNSPD)has undergone rapid development and has become increasingly widely used in fields such as quantum key distribution,space communication,and remote sensing imaging.Due to the widespread use of periodic serpentine structures in mainstream SNSPD,which are highly sensitive to the polarization direction of incident light,the design of polarization insensitive device structures has always been a focus of research.At present,there are relatively mature polarization insensitive SNSPD design schemes in the near-infrared band,but in the mid infrared band,the polarization insensitive design of SNSPD is still in the development stage,and further improvement is needed to better meet the needs of practical applications.Among the commonly used theoretical modeling methods for SNSPD optical response,the Transfer Matrix Method(TMM)does not have the modeling capability for polarization insensitive design.The finite difference Time Domain(FDTD)method needs to conduct two independent simulations for the TE polarization of the incident optical field parallel to the nanowire and the TM polarization perpendicular to the nanowire,The data-driven machine learning method can obtain both TE and TM polarization results in one operation.Therefore,this article is based on the existing 3-5μm On the basis of the design of the mid infrared broadband high light absorption SNSPD,based on machine learning methods,further forward and reverse design attempts have been made for its polarization insensitivity.The former is characterized by the machine learning technology itself being used for predicting light absorption,while the optimization design requires external optimization algorithms.The latter’s main feature is to directly provide device structural parameters based on the expected light absorption,without the need for any external algorithms.The main tasks completed are as follows:1.Based on the polarization sensitivity analysis of existing mid-infrared broadband high light absorption SNSPD structures using FDTD as the main optical response theory modeling tool,the method of coating niobium nitride(NbN)nanowires with high dielectric constant material(silicon Si,refractive index n=3.478)has been adopted,which improves the absorption rate of the original structure for TM polarized incident light in the broadband range and greatly improves its polarization sensitivity.The results show that,without changing the main structural parameters of the original design,SNSPD can generate two absorption peaks with absorption rates of 47.7%and 42.8%for TM incident light in the wavelength range of 3000-5000nm by simply coating the nanowires with high refractive index materials.At the same time,the minimum absorption rate within this range can also reach 12.1%,while the peak absorption rate of the original design is only 8.8%,and the minimum value is close to 0%.Meanwhile,it was found during this process that the thickness of the high dielectric constant coating material has a significant impact on the incident light absorption efficiency of the device TM,and there is currently no optimal theoretical calculation method for the coating thickness.2.The forward design of SNSPD coated with NbN nanowires using high refractive index materials was carried out using machine learning methods.Firstly,using the thicknesses d1 and d2 of the two coating layers on and off the nanowire as inputs,and the average absorption rates ATE and ATM of TE and TM polarization in the wavelength range of 3000-5000nm as outputs,a neural network model is established.Based on the Root Mean Square Error(RMSE)predicted on the prediction set data,the number of hidden layer layers and the number of neuron nodes in each layer are determined.Then,the neural network is used as the prediction model of different polarization absorptivity,and the particle swarm optimization(PSO)algorithm is externally connected.d1 and d2 are set as optimization variables,and the difference between ATE and ATM is set as optimization goal.Then,the minimum value of the objective function is sought through optimization.The obtained results show that the average absorption rate of TM polarized incident light has increased from 5.3%of the original design to 29.65%.At the same time,compared to the previous part of the polarization insensitive design without optimization,the two corresponding peaks and minimum values of absorption rate have been increased to 53.7%,61.6%,and 15.4%,respectively.3.A preliminary reverse design attempt was conducted on the SNSPD of NbN nanowires coated with high refractive index materials using machine learning methods.Firstly,the expected ATE and ATM are used as the two inputs of the neural network,and d1 and d2 are used as the two outputs.Based on the analysis of the RMSE of the prediction set,the hidden layer structures of 1 layer and 100 neurons are determined.Furthermore,the ATE and ATM obtained from the previous forward design were used as expectations,and the corresponding coating layer thickness was directly calculated using this network.The results showed that when the set expectation was within the existing database range,the required coating layer thickness could be accurately obtained through reverse design,with a difference of only 0.29%and 0.03%from the corresponding forward design ATE and ATM.However,when the current expected value was outside the existing data range,The accuracy of the design still needs further improvement.