A Scalable ANN-Based Large-Signal Model for GaN HEMTs Using Transfer Learning

Abstract

Traditional linear scaling artificial neural network (ANN)-based compact models face significant challenges in achieving high accuracy for device modeling. To overcome this limitation, a transfer-learning (TL)-assisted approach is proposed to develop a scalable ANN-based model that incorporates nonlinear scaling of intrinsic parameters. Unlike the linear scaling method, the weights and biases of the output layer are selected and non-linearly scaled for devices with varying gate widths and finger numbers through transfer learning. To effectively integrate these nonlinear scaling parameters into the model, a nonlinear regression technique is employed. The validation results demonstrate that the proposed method provides accurate characterization of both the S-parameters and large-signal performance. Notably, in power sweep evaluations, the proposed method achieves an improvement of more than 8% in power-added efficiency (PAE) accuracy compared with the conventional linear scaling approach.