Nonlinear Behavioral Modeling of FETs: Toward the Implementation of Deep Neural Networks through Large Signal Data and EDA Tools

Nonlinear behavioral Field-Effect Transistor (FET) models often rely on large look-up tables extracted from extensive load-pull characterization. Besides the numerical burden, these models have limited extrapolation capabilities and can hardly be made dependent on the device technology. In this paper, we demonstrate that a Long Short-Term Memory (LSTM)-based Deep Neural Network (DNN) is an effective alternative modeling approach. The DNN is trained with the load-pull data within a simulation platform where data exchange between an Electronic Design Automation (EDA) tool (such as PathWave ADS) and a programming platform (such as MATLAB) is exploited. As a test case, the DNN model has been extracted for an S-band MACOM 10W GaN power device, for which the Enhanced Poly-Harmonic Distortion (EPHD) behavioral model is also available in the ADS. The accuracy of DNN model is verified against the EPHD model in terms of output power, gain, efficiency, and dynamic load lines. Compared to other behavioral models, the DNN approach is expected to provide superior extrapolation capability and to be easily reconfigurable to add/combine heterogeneous device data e.g. from advanced characterization, including memory, and physical (TCAD, EM) simulations.