Multi-Physical Field Modeling And Optimization Of Microwave Components Based On Neural Networks

With the development of electromagnetic(EM)technology,the structures of microwave components are becoming increasingly complex.In simulation and design,it is necessary to consider the impact of multi-physical parameters on component performance rather than electromagnetic characteristics.There are issues such as complex modeling processes,long modeling time and low accuracy,which can no longer meet the current requirements for device modeling with traditional modeling methods for multiphysical fields of microwave components.In recent years,with the rapid development of machine learning,the artificial neural network(ANN)technique has been widely applied to electromagnetic modeling of microwave components.ANN could be a good alternative to EM simulations to construct computer aided design(CAD)models which significantly speed up the EM-based modeling and design process.However,there are several limitations that need to be addressed.Firstly,a large number of labeled samples are required to train neural networks for the optimal design of microwave components,especially involving multi-physical field coupling,which significantly increases sampling cost.Secondly,the traditional numerical calculation methods often require iterative calculations,resulting in low simulation efficiency and further increasing the difficulty of sampling.Lastly,in multi-physical field simulations,the coupling relationships between various models are highly nonlinear,and it difficult for simple neural network models to comprehensively describe them.To further improve the modeling accuracy and efficiency,this dissertation proposes modeling,design and simulation methods based on machine learning for microwave components,and the main contents are as follows.For the large sample demand in multi-physical field simulations,this dissertation proposes an efficient knowledge-based neural network(KBNN)for parametric modeling of multi-physical fields.The input of KBNN includes geometric parameters and multiphysical parameters of microwave devices,while the output is the electromagnetic response of the components.The requirement for multi-physical samples is reduced due to a prior knowledge model,which maps the nonlinear relationships between multiphysical field parameters and geometric parameters.An input update algorithm is proposed to provide labeled samples for the training of the prior knowledge model.KBNN can handle multiple non-geometric input parameters and it also has advantages for shape optimization.Based on this model,a multi-objective model is proposed that can target different performance metrics of components,as well as a multi-level model for array modeling.The testing error of the proposed model is less than 5% compared with the multi-physics simulation software.The proposed model has great advantages in the optimization of microwave components,which can save about 40% of the time compared with the direct optimization of electromagnetic fields.Due to the low efficiency in traditional electromagnetic simulation methods,a finitedifference time-domain(FDTD)algorithm based on neural networks is proposed.The partial differential equations related to space in Maxwell’s equations at each time step are calculated by feedforward neural networks due to its approximation characteristics.The spatial difference calculation in the traditional FDTD method is converted into an optimization problem solved with ANN.ANN is trained individually at each timemarching step,and thus the results of the previous step do not affect those of the next step.It is not constrained by the Courant-Friedrichs-Levy(CFL)condition.Multi-scale problems with fine structures can be solved with ANN-FDTD.The error of ANN is determined by the gradient of its output with respect to the input vector.This error value is used to update ANN parameters with the back-propagation algorithm,so the training does not have to involve labeled samples,resulting in unsupervised learning.Timedomain excitation sources and boundary conditions that act on field data in the timemarching process are introduced as a constrain condition of ANN optimization.The computational time of the proposed algorithm is only about 4% of that of traditional FDTD,and about 35% of that of Alternate Direction-Implicit(ADI)FDTD when solving multi-scale problems.Moreover,all time steps can be divided into blocks to execute parallel computation.For the modeling of multi-physical field coupling,a coupling model based on ANNFDTD is proposed.The interaction between ANN-FDTD for electromagnetic calculations and control equations of other physical fields is studied from two aspects:electromagnetic-thermal coupling and electromagnetic-thermal-stress coupling.Leveraging the unconditional stability of ANN-FDTD,the time stepping and spatial stepping are scaled to unify the scales between the electromagnetic model and other physical models,significantly reducing unnecessary calculations and avoiding repeated mesh generation.Furthermore,a physics-informed neural network based heat conduction model is proposed to replace the FDTD-based thermal model,which improves computing efficiency by about 30% in the case of repeated calculations.