A Neural Network-Based Hybrid Physical Model for GaN HEMTs.

A neural network (NN)-based hybrid physical model for gallium nitride high-electron-mobility transistors (GaN HEMTs) is proposed. In this model, the artificial NN (ANN) is used to construct the terminal charge densities, which is linked with device model’s RF and dc performance using the MIT Virtual Source (MVS) theory. A detailed parameter extraction and optimization method for this hybrid model is proposed as well. The parameter extraction starts from S-parameters with the consideration of access region resistances. The particle swarm optimization (PSO) is used to optimize extraction results. After that, a two-hidden-layer ANN is constructed to model the obtained source-end terminal charge density ($q_{\mathrm {is}}$) and drain-end terminal charge density ($q_{\mathrm {id}}$). Then, the ANN-based $q_{\mathrm {is}}$ and $q_{\mathrm {id}}$ are used to construct the hybrid model under the framework of the MVS theory. The trapping effects and self-heating effects are also considered in this hybrid model. This hybrid model is built and verified in the Advanced Design System (ADS). The proposed hybrid model has better scalability than the traditional ANN-based models. It is also more accurate than the traditional MVS-based physical model. This hybrid model presents satisfactory agreement between measured and simulated current–voltage (IV), S-parameters, load—pulls, and power sweeps. [ABSTRACT FROM AUTHOR]