General And Efficient Machine Learning-based Approaches For Inverse Design Of One-dimensional Photonic Crystals Toward Targeted Visible Light Reflection Spectrum

Data-driven methods have increasingly been applied to the development of optical materials and systems as inexpensive and effective inverse design approaches.Optical properties of photonic crystals are closely associated with characteristics of their light reflection spectra.Finding optimal photonic crystal constructions(within a pre-specified parameter space)that generate reflection spectra closest to a targeted spectrum is thus a meaningful inverse design problem.In this thesis we report two generally effective machine learning-based inverse design approaches for one-dimensional photonic crystals(1DPCs)in different situations,focusing on visible light reflection spectra which are of high practical relevance.The first inverse design method is based on deep neural network(DNN)and Monte Carlo(MC)methods.For a given class of 1DPC system,a DNN in a unified structure is first trained over data from sizeable transfer matrix calculations.A novel scheme that integrates DNN backward predictions(from reflection spectrum to layer thicknesses),Monte Carlo moves,and transfer matrix forward calculations is then constructed to form an iterative optimization loop.Four representative 1DPC systems,including periodic structures with two-,three-,and four-layer repeating units and a heterostructure,are investigated,all successfully reaching optimal solutions of 1DPC construction regardless of the exact achievability of targeted reflection spectra.For example,in the inverse design toward an artificially constructed “square-shaped”green light reflection spectrum or red light reflection spectrum,the approach shows excellent robustness and can automatically find optimal layer thicknesses even when they are outside the range covered by original training data of DNN.The second inverse design method is based on the deep deterministic policy gradient(DDPG)algorithm in reinforcement learning.This method realizes synchronization of data generation and model "learning" without the need to generate a complete data set in advance.For 1DPC heterostructure systems with many structural parameters and a large parameter space,this method has good applicability.We have used this method to inversely design various structural parameters of 1DPC heterostructure systems,including refractive index,layer thickness,and period,and successfully completed inverse design for the three types of targeted reflection spectra under our study.The high generality and efficiency of the two approaches developed in this thesis make them powerful exploratory tools to guide the design and optimization of 1DPC materials in future.