Volterra black-box model of electron devices nonlinear behavior based on Neural Network parameters GEORGINA STEGMAYER MARCO PIROLA GIOVANNI GHIONE Electronics Department Politecnico di Torino Cso. Duca degli Abruzzi 24 10129 Torino ITALY GIANCARLO ORENGO PAOLO COLANTONIO Electronics Engineer Department Università di Roma II via Politecnico 1 00133 Roma ITALY Abstract . With this paper we want to present a black-box model, that can be applied to a vast number of RF electron devices (e.g. FET). We will show that an analytical Volterra series approximation of the nonlinear behavior time-dependent model of an electron device can be built using a neural network and its parameters, once the proper training data are given. Key-Words: black-box model, nonlinearity, Volterra model, Volterra Kernels, neural network parameters. 1 Introduction The classical representation of nonlinearities inside an electronic device/element are usually modeled thorough equivalent circuit, assigning the constitutive relations among the current or charge-voltage relationships and the controlling voltages. This procedure is based on the known physical behavior of the modeled device that dictates the model topology. For this reason not only the model must be tailored to the device, but also the extraction of its parameters strongly depends on it. Other approaches claim to represent a general device and are referred to as black-box model, but that it is generally only partially true, since some assumption on the device is always done. The Volterra expansion approach has been proposed since many years, at least in principle, for non-linear modeling, through an input-output relationship. Due to the difficulties in identifying the higher order kernels through experimental data, this approach is complex. For this reason it is not effectively used within commercially available circuit simulator, despite the fact that it can originate intrinsically a black-box model and, therefore, device independent model. On the other side the neural network approach to device modeling has received increasing attention in recent years since model tailoring to the device under study only needs a training procedure based on experimental data. In this paper we propose a black-box time dependent neural network model able to build a Volterra series approximation suitable for electronic devices, allowing its kernels evaluation as well. In this paper we will show, as a case of study, the results obtained using a Curtice MESFET model [1] as the reference drain current. The Volterra series approach is explained in Section 2. Our proposed dynamic neural network model appears in Section 3. Simulation results are presented in Section 4 and the conclusions of this work can be found on Section 5. 2 Volterra series model of nonlinear behavior A non-linear dynamical system can be represented exactly by a converging infinite series of the form (1), that reports the dynamic expansion of a single -input single-output system. This equation is known as the