Research On Reverse Design Of Silicon-based Photonic Devices Based On Machine Learning

The application of machine learning technology in the field of photonic device design can greatly improve the design efficiency of photonic devices.This thesis studies how to use the machine learning technology to design photonic devices and proposes three design schemes: a cascaded neural network design scheme containing binary activation functions,a design method of photonic device which combined forward convolutional neural networks and genetic algorithms,and a method of full-space light field prediction.Using the above methods,we have designed an arbitrary ratio power splitter,a meta-surface beam deflector and completed the full-space light field prediction of the optical beam splitter.The main work of this paper is as follows:(1)Proposed a cascaded neural network structure including a binary activation function layer,and designed an arbitrary ratio power splitter with this structure.The use of the binary activation function can directly design a digital device structure,avoiding the performance loss of the device caused by the traditional binary scheme.This paper uses this scheme to design the power splitter with different power branch ratios of 1:1,1:2,2:1,1:3,1:4,etc.After the neural network training is completed,each of the design only takes a few seconds,and the efficiency is increased by 4-5 orders of magnitude compared with the traditional reverse design method.(2)Proposed a design method combining forward convolutional neural network and genetic algorithm,and designed a meta-surface beam deflector by this method.This method can well avoid the "one-to-many" problem that occurs when directly training the reverse network,and can obtain more accurate results.This paper uses this method to complete the design of a meta-surface beam deflector with a given real and imaginary part of the light field and a given deflection angle.The design efficiency is increased by 3-4 orders of magnitude compared with the traditional reverse design method.(3)Used a U-net structure network to predict the full-space light field.The performance of the device can be calculated from the light field in the entire space of the device.Training a neural network to predict the light field in the entire space of the device can solve the problem of training dozens of different networks due to different design goals,and greatly improve the data utilization rate.In this paper,the residual U-net structure of the neural network scheme is used to predict the full-space light field of the optical beam splitter,and the prediction results are more than 95% similar to the simulation results,which lays a foundation for the further use of space light field prediction for photonic device design.