Precise Modeling of Emerging Electronic Structures by Artificial Neural Networks

Nowadays, there are many emerging electronic structures for which their nonlinear models for computeraided design are necessary, especially for the ones from the areas of nanoelectronics and microwave techniques. However, for such structures, sufficiently accurate analytic models are mostly unavailable. This is partially caused by the fact that the physical principles of the element operation are sometimes not fully clear (especially for quantum devices), and also by bizarre characteristics of some of the elements (typically with irregularities and a hysteresis in parts of characteristics, or by negative differential conductances that are typical for the microwave transistors). In such cases, models based on artificial neural networks are necessary and useful for these elements. Majority of the elements can be characterized with a single artificial neural network. However, for certain kinds of elements, a cooperation of more artificial neural networks is necessary. This case is described in the paper first, where the Pt−TiO2−x−Pt memristor characteristic with an extraordinary (but typical) hysteresis is approximated by a set of cooperative artificial neural networks, as a single network is unable to characterize this unconventional element. Second, an ability of the artificial neural networks for modeling the negative differential conductance is demonstrated by characterizing the 110GHz pseudomorphic high electron mobility transistor (pHEMT). Moreover, a semiautomatic selection of an optimal structure of the networks (both numbers of hidden layers and the numbers of the elements in the layers) is also suggested.