Applications Of Artificial Neural Network In The Analysis And Synthesis Of Integrated Capacitors

MIM (Metal-Insulator-Metal) capacitor and Interdigital capacitor are two important types of lumped components in the microwave circuits. They are used for RF bypassing, dc-blocking, reactive termination applications and marching networks. In additionally, they are also widely used to realize filters, dividers/combiners, couplers, baluns and transformers. From the previous works we can see that no matter we use the approximated formulas, which based on the neglect of some weak field effects, or the full-wave analysis, which based on some numerical algorithms, or the measurement-based model, it is difficult to meet not only the requirement of speed, but also the requirement of precision. Hence how to build a fast and efficient model for these integrated capacitors becomes an essential direction in current researches.Theoretically, Artificial Neural Network (ANN) could approximate any map relations of any non-linear scale. So more and more people began to be attracted by the powerful non-linear input-output mapping capability of ANN. When it comes to some problems which have large scale, complex structures or can not solved with close form formulas,. ANN is no doubly a good choice to turn to, whereas parameters extraction of the integrated capacitors is just a kind of problem with no close form solution.This article combined the ANN with full-wave analysis tools, equivalent circuit model and measurement-base components library to obtain normal Neural Networks and inverse Neural Networks to solve the problems mentioned above. The electromagnetic simulators and components library based on measurements are employed to calculate the frequency-domain responses under certain structure parameters, then the capacitors and other related parameters in the equivalent model proposed in this article can be worked out. Finally, we used these data to train a ANN so that the result net can quickly and precisely give designers the correct responses or the corresponding values of the model. Such ANNs are called normal ANN in this article.After obtaining the normal ANNs, it is time to train the inverse ANNs, which are aimed to calculate the structure parameters of the integrated capacitors given their values of capacitors. Unlike the traditional search method, this article tried to interchange the input data and output data directly and then start training. Once the inverse ANNs get ready designers can easily know what the structure parameters of the capacitors they should use if they tell the inverse ANN the frequency Domain responses of the capacitor values they want. As the result of the two kind of ANNs, the time consumed in the capacitors computation and the structure decision can both be great shorten.There are four sorts of capacitors with different structure being studied in this article, including two kinds of Interdigital capacitors, which differ in the fed ways, and two kinds of MIM capacitors, which differ in the insulate layer.