Artificial Neural Networks for GaN HEMT Model Extraction in D-Band Using Sparse Data

We describe the application of Artificial Neural Networks (ANNs) for Gallium Nitride (GaN) High-Electron Mobility Transistor (HEMT) model parameter extraction to improve the model accuracy between 110 and 170 GHz. Fully-connected ANNs trained by backpropagation relate the physics-based ASM-HEMT model parameters to RF transistor measurements. The effects of ANN activation function, number of layers, number of nodes and number of training set data points on training accuracy are studied. For the 12 model parameters that dominate the 40-nm GaN HEMT RF characterization, we obtained a combined root-mean-squared (RMS) error of 2.5% between the ANN prediction and the training set, which is acceptable for most design tasks.