Research On High-efficiency Inverse Design Method For Silicon Photonic Devices

Inverse design is a new method to realize high density and high performance integrated silicon devices based on different algorithms.In the field of on-chip optical interconnection,inverse design can effectively reduce the size of integrated silicon devices while maintaining the high level of performance,and inverse design algorithms for silicon devices have been developing in the direction of high efficiency.Recently,the combination of various algorithms to achieve efficient reverse design has become a popular research direction.By giving play to the advantages of different algorithms,the efficiency of inverse design can be maximally improved.In this paper,two new efficient inverse design algorithms are used: the first algorithm combines the adjoint method with the direct search algorithm to design a mode convertor;The second algorithm combines the adjoint method and deep neural network to design the beam deflector.The main work of this paper is as follows:(1)An efficient digital subwavelength structure photonic device inverse design based on digital adjoint method is proposed.Digital adjoint method adopts the adjoint method based on sensitivity analysis to overcome the efficiency bottleneck of conventional direct search algorithm.Based on the theory of equivalent medium,the traditional binary digital subwavelength structure is extended to multi-valued structure to reduce the performance degradation caused by the digitization of device design and obtain better device performance.The digital adjoint method is used to inverse design the silicon-based mode converter.(2)An efficient inverse design algorithm for metasurface photonic devices is proposed by combining the adjoint method with the deep neural network.Based on the gaussian distribution noise source,the device pattern with global randomness is generated in real time,which overcomes the local optimization characteristics of the adjoint method.By using the adjoint sensitivity analysis,the gradient distribution of the dielectric constant pattern of the metasurface device is calculated in real time.It is not necessary to generate a mass training set offline,and the real-time training of inverse design neural network can realize inverse design of metasurface devices with high efficiency and high performance.The one-dimensional beam deflector is designed using this method.Compared with the conventional method,the average deflection efficiency of the optimized device is improved by 3% in the wavelength range of 800-1150 nm and the deflection Angle of 40°-70°,and the number of electromagnetic field simulation required to make the training set of deep neural network is reduced to one tenth of the conventional training method.