Selecting an optimal structure of artificial neural networks for characterizing RF semiconductor devices

At present, there are many various microwave structures for which their nonlinear models for CAD are necessary. However, in the recent PSpice family programs, only a class of five types of MESFET model is available. In the paper, a method is suggested for modeling miscellaneous RF semiconductor devices by exclusive neural networks or by corrective neural networks working attached to a modified analytic model. An accuracy of the proposed modification of the analytic model is assessed by extracting model parameters of the AlGaAs/InGaAs/GaAs pHEMT. An accuracy of procedures with neural networks is generally assessed by extracting their parameters in static and dynamic domains. An approximation of the AlGaAs/InGaAs/GaAs pHEMT output characteristics is carried out by means of both exclusive and corrective artificial neural networks. A systematic sequence of analyses is also performed for examining an optimal structure of the artificial neural network from the point of view its structure and complexity. The tests have been performed on both five- and four-layer artificial neural networks that serve for modeling a P-channel JFET and for the AlGaAs/InGaAs/GaAs pHEMT.