Analysis and validation of neural network approach for extraction of small-signal model parameters of microwave transistors

Extraction of a small-signal model is the first step in modeling transistors for advanced microwave applications. There are different extraction techniques, mostly based on optimizations or on direct analytical procedures. An alternative to the standard extraction methods are procedures based on application of artificial neural networks. Namely, an artificial neural network is trained to predict equivalent circuit elements directly from the measured scattering parameters without the need for any additional tuning of the elements. In this study we present results of a comprehensive analysis of the neural based extraction procedures. We examined stability of the extracted values with the choice of the input set of scattering parameters as well as accuracy of the final small-signal model. Moreover, we investigated the influence of the number of measured data necessary for development of reliable neural models. The extraction procedure was examined for a HEMT transistor working under varying temperature conditions.