Efficient ANN Based Noise Modeling of Microwave FETs Against Temperature Zlatica Marinkovi , Olivera Proni -Ran i , Vera Markovi Faculty of Electronic Engineering, University of Niš Beogradska 14, 18000 Niš, Serbia and Montenegro e-mail : [zlatica, oljap, vera]@elfak.ni.ac.yu Abstract – An improved noise modeling technique for microwave MESFETs / HEMTs versus temperature is presented. It is based on an artificial neural network (ANN) that produces noise parameters as its outputs for device temperature, S parameters and frequency at its inputs. Once trained, the proposed model can be used for efficient prediction of transistor noise parameters over a wide temperature range. Since the model is based on ANN, all noise-generating mechanisms are included and therefore it is more accurate than empirical transistor models, as it is shown on a numerical example. I. INTRODUCTION During the last decade, some attempts have been made to model devices and systems in the field of microwaves by using ANNs, [1]-[8]. ANN modeling techniques represent efficient alternative to complex time-consuming electromagnetic models because once developed neural models give responses practically instantaneously, since response providing is based on performing basic mathematical operations and calculating elementary mathematic functions (such as an exponential or hyperbolic tangent function). Neural networks are able to approximate any nonlinear function and to learn from experimental data. Therefore, it is possible to develop neural model from source-response data points without the knowledge about the physical characteristics of the problem to be solved. The most important feature of neural models is their generalization capability, i.e. the capability to provide the correct response even for the input values not presented to the neural network in the training process. In that way, the developed models can be used for a reliable prediction over a wide range of input parameters. The noise characteristics of microwave transistors, often needed for the design of low-noise receiving equipment, are usually treated in terms of four noise parameters (minimum noise figure, optimum source reflection coefficient and noise resistance) that are frequency-, temperature- and bias- dependent. Since measurements of noise parameters are complex and time-consuming procedures, device noise models are usually used for device noise prediction in the microwave circuit design. But, most of the existing noise models require recalculation of elements / parameters of the equivalent circuit for every operating temperature and bias point, [9]-[11]. The authors of this paper have proposed bias-dependent device signal and noise models based on neural networks, [8], [12] and an application of neural networks in the transistor noise wave models [13]. Recently, the authors have proposed ANNs to be used for the modeling of noise parameters of MESFETs / HEMTs under different temperature conditions, [14]-[15]. A noise modeling technique presented in [14] is based on a combination of an ANN and an empirical transistor noise model. The ANN is used for modeling of the temperature dependence of elements and parameters of transistor noise model equivalent circuit. By assigning this ANN to the empirical model equivalent circuit, a temperature dependent noise model is developed. The main limitation of the model is that its accuracy cannot be higher than accuracy of the empirical noise model. The model proposed in this paper belongs to black-box type neural models. It models dependence of transistor noise parameters on temperature and frequency. In order to improve accuracy of modeling, S parameters are introduced as ANN inputs as well. The paper is organized as follows. In Section II transistor noise modeling problem is stated. ANNs are shortly introduced in Section III. The proposed transistor noise model based on neural networks is described in Section IV. A modeling example and the main results are presented in Section V. Finally, in Section VI the main conclusions are reported. II. MESFET / HEMT NOISE MODELING More than fifteen years, an empirical MESFET / HEMT noise model proposed by Pospieszalski, [11], has been accepted as one of the most convenient and accurate. + Fig. 1. Equivalent circuit of MESFET / HEMT packages with included noise sources IEEE MELECON 2006, May 16-19, Benalmádena (Málaga), Spain 1-4244-0088-0/06/$20.00 ©2006 IEEE 153