Inverse Design Of One-dimensional Photonic Crystals Enabled By Genetic Algorithm And Machine Learning

Machine learning is a multi-region intersection subject.With its strong function fitting ability,the combination of machine learning and various fields has become one of the most popular research contents.As machine learning moves into the field of nanophotonics,new opportunities are opening up for conventional optical computing,while some of the challenges of inverse design are expected to be solved.Due to the optical response in the field of photonics is often obtained through the finite element method,finite difference time domain method,and so on,it is usually time consuming and high computational cost.Therefore,the emergence of machine learning can solve the problem that complex optical response is difficult to obtain to a certain extent,which can not only realize fast and accurate forward prediction of optical response,but also greatly improve the efficiency of inverse design of structural parameters according to the target optical response.In this paper,we study a method to implement optical inverse design by combining the genetic algorithm and machine learning,and successfully design multiple edge states in multiple band-gaps of one-dimensional photonic crystals.The main research contents include:(1)The calculation method of multi-layer one-dimensional photonic crystal band structure is improved to shorten the time required to prepare the training data set to 1/3 when compared to previous method.The neural network is trained for several kinds of one-dimensional photonic crystals,and the mean square error loss function of the network is reduced to 10^-4 magnitude,which means the network can be used as an effective tool to obtain optical response instead of numerical calculation and simulation.(2)Through the combination of genetic algorithm and trained neural network,the complete topological vector is inverse designed successfully and verified by numerical calculation.Meanwhile,effects of different fitness function are analyzed.(3)By setting the target frequency bands,the inverse design of single to multiple edge states is realized successfully and verified by simulation.After changing the materials,the inverse design of edge states is also successfully realized,and the existence of two edge states in different band-gaps is verified by experiments.Finally,we extend layers of one-dimensional photonic crystal,successful design of which proves the effectiveness and scalability of the method.