Research On Technologies Of Inverse Design For Silicon-photonic Devices

The research goals of silicon-based optical devices are to realize the functions of light generation,transmission,control and interconnection in silicon/silicon-based materials.Simultaneous fabrication of electronic devices and photonic devices on a silicon substrate and their integration on the same silicon chip can form photonic devices with many special optoelectronic properties that are widely used in optical communication systems,optical interconnect technologies and light Calculation and other fields.The design of the traditional silicon-based optical devices tend to acquire an initial structure according to analytical theory with following complicated work of optimization.This approach is not only inefficient,but also has great limitations on the functionality of the device.On the other hand,with the improvement of the fabrication process of nano-photonics,it is no longer difficult to make a high-precision device structure.In order to improve the precision of silicon-based optical devices by utilizing the rapid development of nano-fabrication technology,a design method called reverse design has been proposed in recent years.Inverse design of silicon-based optical devices is essentially a way of mathematical computing.Firstly,the structure of the device and the target performance are mathematically modeled,and then the mathematic model is optimized by the mathematical optimization theory to finally obtain the silicon-based optical device that meets the performance requirements.The silicon-based optics that are obtained by inverse design theoretically can have any size and shape,giving us more ways to operate the light.The reverse design approach is an important milestone for the development of silicon-based opto-devices.This dissertation focus on the research of the key technologies involved in inverse design.The main contents are as follows:Firstly,this paper introduces and analyzes the high-performance computing technologies such as GPU parallel computing technology and MPI distributed messaging technology used in numerical calculation of inverse design.We optimize the algorithm in the inverse design system,and solve the problem of design calculation interruption by using blocking transmission and non-blocking receiving mode.Secondly,we elaborate on the mathematical optimization theory involved in inverse design.First of all,the algorithm of Alternating Direction Multiplier Method,which is used to solve the problem of large-scale constrained distribution,is introduced.And its application in inverse design is deduced in detail,including the implementation of alternating optimization of two variables in the algorithm.At last we analysis the performance of Newton method used to optimize the electromagnetic field variables,and propose and try the quasi-Newton method to improve the optimization performance and efficiency.We combine the high-performance computing of GPU parallel with FDFD algorithm to solve the electromagnetic field of the device in reverse design.First,we deduce the process of solving Maxwell's equations in 3D space using FDFD algorithm.Mainly,FDFD algorithm uses the difference algorithm to solve the Maxwell's equations;however it converges too slowly due to the huge matrix dimension when solving the equations in the three dimensional space.In order to solve this problem,we combine the high performance parallel computing features of GPU with FDFD algorithm and achieve the iterative solution of FDFD on GPU using CUDA programming.Moreover,we compare the performance of the GPU-based FDFD with the traditional Finite-Difference Time-Dominance(FDTD)solution of 3D Maxwell's equations by investigating the time-consuming in solving Maxwell's equations of an inverse-designed SOI device.The result show that the FDFD algorithm based on GPU acceleration has faster electromagnetic field calculation speed under the same condition.When the amount of data is 3.2×10e6,the time spent on an electromagnetic field calculation of GPU-based FDFD is less than one-tenth of that of FDTD;Further results show that the FDFD algorithm based on GPU acceleration has even more significant improvement in time-consuming when the amount of data increases.