Modeling And Optimization Of RF Circuits Based On Deep Learning

With the rapid development of wireless communication technology,the market demand for microwave RF devices grow rapidly.Modeling and parameter optimization of RF devices is currently based on accurate full-wave electromagnetic field calculations.However,the full-wave electromagnetic simulation algorithm is computationally intensive,so it requires a lot of computational resources.Some studies use genetic optimization algorithms,but it still rely on simulation algorithms or tools.If the population of genetic is too large,it is very time consuming.With the rapid development of artificial intelligence(AI)algorithms,the deep learning framework can be easily used to design encapsulated interfaces with high performance and efficiency.Because deep neural networks have the ability to simulate highly complex nonlinear mapping,this also makes artificial intelligence-based circuit modeling and optimization design possible.In this paper,we model and optimize RF devices based on deep learning.As to the combination of deep learning fast prediction ability,global optimization of genetic algorithm and fast local convergence performance of Powell algorithm,the modeling and optimization of the circuit becomes very efficient.The innovations of this paper are as follows: We perform electromagnetic simulation on RF devices based on full-wave simulation tools,use series of results from it as training data,and establish corresponding AI models.Using the model to analyze a series of RF devices suitable for the model,the electromagnetic simulation results of the device can be directly obtained without the need for time-consuming full-wave electromagnetic field calculations.At the same time,by using the AI model and the circuit topology,the parameters of the equivalent circuit can be obtained quickly and accurately by the optimization algorithm.However,one of the main difficulties in using the optimization algorithm to extract the parameters of the equivalent circuit is to apply the optimization algorithm to effectively approach the global minimum.The traditional genetic algorithm(GA)is able to find global minima,but the convergence rate is very low.On the other hand,direct search methods can easily fall into local minima.In this paper,we propose an improved parameter extraction hybrid genetic algorithm that can effectively search for global minimums.This method combines the advantages of the genetic algorithm and the Powell algorithm to efficiently extract the circuit component parameters of the equivalent circuit model.