Principle And Application Of Artificial Neural Network In Integrated Optical Waveguide Device

The integrated optical waveguide device(IOWD)has been widely used because of its advantages of high integration,low power consumption,large working bandwidth and strong scalability.At present,it is developing towards high integration,functional complexity and intelligence.However,it is facing the challenge of efficient design and intelligent construction.In this paper,artificial neural networks(ANNs)technology is applied to the design and application of integrated optical waveguide device.The related work completed is as follows:1.In view of the low efficiency of complex IOWD design,a new method of transmission spectrum prediction,reverse design and performance optimization for complex IOWD by using ANNs is proposed.In this paper,the theoretical analysis and simulation of the plasmonic waveguide coupled with cavities structure are studied,which can excite the surface plasmon polaritons with the advantages of near-field enhancement and breaking through the diffraction limit.The effectiveness of this method is verified by the Fano resonance and plasmon induced transparency from this structure.At the same time,genetic algorithm is used to optimize the network structure of ANNs and select the appropriate hyper-parameters for it.The results show that the accuracy of prediction of transmission spectrum by ANNs is higher than 90%,and can achieve reverse design and performance optimization effectively.The effect is not lower than that of other algorithms such as binary search.Compared with the previous work,this method greatly improves the efficiency and intelligence of complex IOWD design,and reduces the complexity.It has application value for the realization of high-performance IOWD.2.In order to solve the problem of lack of effective on-chip training in the development of intelligent IOWD,a scheme of on-chip training for intelligent IOWD based on ANNs architecture using evolutionary algorithm is proposed.In this paper,the optical neural network(ONNs)is used as an example for simulation research.Two typical evolutionary algorithms,genetic algorithm and particle swarm optimization,are used to self-study the weights and determine the hyper-parameters for the ONNs.The effective on-chip training of intelligent IOWD is realized,and it is applied to the classification task of iris data set,wine data set and modulation format recognition.The simulation results show that the accuracy of classification and recognition can reach more than 89%,and the accuracy and stability are not lower than the adjoint variable method which is used in the training of the ONNs.This strategy breaks through the limitation that the adjoint variable method can not optimize the architecture of ONNs,and is expected to overcome the low computational speed of the current on-chip training algorithm when carrying out large-scale intelligent IOWD training tasks.