Research On Intelligent Design Algorithm Of Nanophotonic Devices Based On Few Samples

Thanks to the advancement of nanophotonic devices with specific structures,light can perform optical activities such as scattering,refraction,reflection,confinement,and filtering in a new way at the nanoscale,and exhibit extraordinary physical properties.These physical properties have played an important role in tremendous fields.The design of traditional nanophotonic devices often relies on numerical calculation methods and prior knowledge,which may cause many issues such as high threshold and low efficiency.In recent years,machine learning has achieved rapid development,which provides new ideas and novel methods for nanophotonic device design.However,the existing machine learning methods for nanophotonic device design have met problems including high model training cost and poor model versatility,which greatly hinders the application of machine learning methods in the design of nanophotonic devices.Therefore,targeting the two key problems of high model training cost and poor model versatility,this thesis proposed a model-agnostic data enhancement algorithm and a scalable multi-task learning model based on the machine learning technology under the condition of few samples.The main works are described as follow:1.Aiming at the high cost of model training,the thesis proposed a model-agnostic data enhancement algorithm based on the transfer learning method,which could be combined with different machine learning models to achieve forward prediction of nanophotonic devices.Instead of directly using data to train machine learning models,modelagnostic data enhancement algorithms can significantly reduce the consumption of model training samples and improve the generalization performance of the model.Additionally,this algorithm significantly improved the generalization ability of the models of artificial neural networks,random forest regression,and support vector regression.In terms of time complexity,although the algorithm increased the time cost of model prediction,its calculation efficiency was still much higher than RCWA(i.e.,four to five orders of magnitude advantage).2.As for the problem of poor model versatility,this thesis proposed a scalable multitask learning model based on the multi-task learning method.This model can not only perform inverse design for various nanophotonic devices in a high-precision and ultrafast manner,but also can be quickly expanded to new structural designs.It overcame the bottleneck that a single model of traditional machine learning methods is only applicable to specific physical problems,and improved the versatility of machine learning models.Interestingly,based on the data sets constructed by FEM,this model realized the inverse design of low-dimensional material heterojunctions with different structures and different parameter combinations,and showed strong inverse design capabilities in multiple data sets.Relying on the similarities between physical tasks,the above works proved that it is feasible to carry out the research of nanophotonic device design algorithm under the condition of few samples utilizing machine learning.This thesis furnished remarkablynew methods and inspiring ideas for the efficient design of nano-photonic devices in the future.