Drain-Source Symmetric Artificial Neural Network-Based FET Model with Robust Extrapolation Beyond Training Data

A large-signal FET model based on artificial neural networks (ANNs) is extended for rigorous intrinsic drain-source symmetry and robust extrapolation beyond the range of training data. Enhanced ANN architectures and training algorithms constrain the five nonlinear model state functions to transform according to the discrete symmetry rules related to the device invariance with respect to intrinsic drain-source exchange. This extends the applicability of the previous ANN-based model to situations where the instantaneous voltage crosses V_ds= 0, such as switches and mixers. The model is compiled in Agilent ADS, together with advanced extrapolation routines extending the model beyond the range of training data for improved convergence. The model has been generated for FETs from several III-V semiconductor processes, and validated with extensive independent small and large-signal measurements.