Transfer Learning-Enhanced ANN for Scalable Small-Signal and Noise Modeling of HEMTs Based on Signal and Noise Matrix Knowledge

Scalable small-signal and noise modeling for high-electron-mobility transistor (HEMT) based on a combination of the conventional circuit model theory and artificial neural network (ANN) is presented in this article. The primary advantage of the proposed method lies in alleviating the extrapolation ability of neural models can be alleviated by incorporating circuit knowledge. A transfer learning framework is employed, consisting of a source- and target-domain neural networks. Large-size devices can be regarded as composed of elementary units, and the corresponding performance can be predicted using the scalable rules derived from the noise correlation matrix and signal matrix. The proposed technique is highly valuable for neural-based microwave computer-aided design (CAD) and for analytically unified small-signal and noise modeling. Model verification is carried out through the comparison of measured and simulated S-parameters and noise parameters. Excellent agreement is obtained between simulated results and measured results for 2\times 20~\μ m, 2\times 40~\μ m, 2\times 60~\μ m, and 2\times 100~\μ m gate width (number of gate fingers \times unit gate width) HEMT devices over a wide range of bias points.