Research On The Inverse Design Method Of Silicon-Based Photonic Devices Based On Artificial Neural Network

Silicon-based electronic and photonic devices produced via CMOS process platform have the characteristics of small footprint,high integration and are highly suitable for mass production.Thus,such devices can be utilized to realize electronic and photonic integrated systems with different functions,which hold a wide range of application prospects and industrial value in fields of communication,sensing,computing,etc.Nowadays,with the increasing quantity and variety of integrated devices,electronic and photonic integrated systems are also becoming more complex and demanding.Accordingly,the degree of freedom and complexity of the inverse design for electronic and photonic devices are also increasing.For traditional inverse design methods,researchers usually combine the optimizationand electromagnetic calculation algorithms to iteratively alter the structure of the device design area and conduct simulation,forcing the electromagnetic performance index of the device to approach the target value.Such approaches usually require a lot of computing time and resources.In recent years,the rapid development of artificial neural network opens up a new path for device design.Focusing on the research topic of the inverse design for silicon-based photonic devices,this dissertation improved and proposed a variety of neural network models to tackle the problems of non-unique mapping,large device parameter quantity,structural diversity and complex data preparation process.At the same time,the proposed work solved the problems occurred in traditional inverse design methods to a certain extent,such as mass computation cost,slow solving speed,and lack of memory.The main innovation points of this study are as follows:(1)Firstly,an improved tandem neural network for the inverse design of one-dimensional-structured devices is proposed,which introduces a loss term related to the device structure parameters to the original neural network model,allowing the proposed model to purposefully learn the device structure in the training set.The proposed model can effectively solve the problem that the trained inverse design model produces unreasonable device structure during actual application.Additionally,the improved model does not require specific preprocessing for the data set during training,nor does it depend on certain activation function to restrict the generated structural parameters.In the inverse design for grating coupler and multi-layer film coatings,the proposed network with the loss term showed higher inverse design ability than the original model.The spectral response of the generated device was closer to the target value,and the inverse design error was decreased by 14.79%and 8.16%,respectively.(2)Secondly,aiming to enhance the diversity of inversely designed structures,this dissertation proposes improved Wasserstein generative adversarial network with simulation module and self-attention mechanism for designing heterogeneous-structured or large quantity parameter devices,namely WGAN-sim and SAGAN-sim.The proposed WGAN-sim model is formed by connecting the generator of Wasserstein Generative Adversarial Network(WGAN)with a simulation model,so that the generator can obtain error feedbacks between the generated device spectral response and the target value,thus improving the training effect of the model.Based on WGAN-sim,SAGAN-sim model is formed by introducing self-attention(SA)mechanism into the generator and discriminator,which enables the model higher ability to learn the relationship between the local and global structures of the device during the training process,and further enhances the inverse design ability of the model.In addition,by adjusting the hidden variables in the inverse design process,the model can produce a variety of devices with different structures but similar spectral responses.In this dissertation,WGAN,WGAN-sim and SAGAN-sim models are respectively used in the inverse design for the interference region of the multimode interference(MMI)power splitter with three ports,whose results prove that the SAGAN-sim model has better training effect and stronger inverse design ability.The average quality factor of the generated device is about 0.81,which increases by 12.5%compared with the WGAN model.Also,the transmittance of the output port of the photonic device can reach 94%and the reflectivity is lower than 0.5%.(3)Lastly,this dissertation studies the application of deep reinforcement learning in the inverse design of silicon-based electronic and photonic devices,and proposes an asynchronous double deep Q-network(A-DDQN)model for exploratory inverse design.Multiple independent environments and agents can operate simultaneously in the proposed model,where each double deep Q-network agent learns to make the next optimization decision by analyzing the current device structure.The proposed A-DDQN model does not require training set preparation in advance.Instead,the experience fragments generated during the device inverse design are used for training and optimization of the model.The proposed model also adopts asynchronous execution method to share experience segments and weight gradients to speed up the training of global models and make full use of computing resources.In this dissertation,the proposed A-DDQN model is used to etch the interference region of MMI power splitter step by step,whose results proves the feasibility of this method for device inverse design.The computing time for generating four kinds of devices with different spectral ratios usually takes less than 48 hours,with the maximum error of the power ratios less than 13%.